community-help

[faint pawn]

No more thumbs Up or Down feedback?

[green summit]

It should still be there, however, it doesn't always show for every single message/response. (98% does ofc)

[faint pawn]

It’s been gone for like 6 hours. Everything is updated

[green summit]

Could be an IOS issue. (Wouldn't be the first time)

[green summit]

Looks dope!

[queen birch]

Looks amazing fr

Anyone in Ai automation?

[faint pawn]

It took me a minute to read your name 😭

[mild yarrow]

Lol 😆 😆

[tidal depot]

Is anyone else getting error in message stream constantly right now?

[valid herald]

Yes I do have.

[hoary epoch]

Same.

[tidal depot]

Think it’s good now, sort of, still get it every 2-3 chats, also thinking is a bit slower too

[true mauve]

Hey I need to talk with a mod or sum, y'all just basically stole $22 from my girls card and didn't even give me the sub

After it took the money out it just showed this.

If anyone knows how I can go about a refund or sum please reach out cus this is BS

[charred parcel]

nobody here can help with this. You need to contact support at https://help.openai.com

[north tusk]

Idk if this is the right channel to put this, but why is it that my Android app isn't opening the chat? But my iPhone, desktop, and desktop app can all open this one particular chat? It's just the Android app that's acting up. The chat isn’t even maxed out. Nothing's wrong with it...

[karmic folio]

How do I fix guys? Help please??

[trail cairn]

poor bait to post 2 years old stuff

[slender vapor]

in codex the browser use plugin is just gone im on windows i had access to it the whole time but now its just gone i cant even install it in the config.toml the bundled browser is activated

[charred parcel]

I don't know how to fix this on Windows but Codex is much more painless on a Mac

[frail summit]

what's up with gpt 5.5 pro via the API?

Seems like it's a completely different model. For instance it states its knowledge cutoff is June 2024 when the docs say it should be December 2025

and it's suspiciosuly fast

The above doesn't seem like a hallucination either. As here it has no knowledge of the louvre heist that happened in October 2025

[trail cairn]

first time using the api? cutoff date is not part of the model training data and cutoff date means some of the data can be up to given date, not full knowledge

[tepid leaf]

anyone here could help me or knows what to do when a password reset gives this error?

Oops, an error occurred!
Username not found.
error_code: username_not_foundrequest_id: 4912bc91-dfc3-40bd-9667-3a5e0435da47

[trail cairn]

do you actually login with username or just use google, apple etc login?

[green summit]

Log out/back in. This might help

Tho, it might already gotten fixed.

[tepid leaf]

i used to login with just my email, but logging in isn't working through the email-passcode (error_code: invalid_state) and then when i try reset password i get the username_not_found. I tried everything from removing cookies/data + incognito, different devices. But nothing changes the fact i get these errors. So looks like an account problem.

[agile gale]

Yeah but it always let me read it at least In the past now not even that? That's dumb.

Yeah yeah but it has still let yiu at least read it in the past now it won't and I'm upset about it

[frail summit]

oh, so this is the model? it's just it seems very fast by default 😂

[vast ferry]

hi! can anyone help me to contact someone about getting enterprise access for my business? i havent had luck via email or website forms

[onyx kraken]

I'm sorry that happened to you. Any too-long chat can bug out weirdly, that's a risk we all take if we get really long chats.

[onyx kraken]

We're nearly all of us community members here, like yourself! Even the mods are volunteers.

The offered ways to request enterprise are, as far as I know, the only ways.

[charred parcel]

ask in #ai-discussions as we can only discuss OpenAI products and models here

[upper junco]

Ahhhh ok thank you

[patent jay]

Can I get help with codex here?

I want to use the chrome plugin but I don’t believe it’s working.

[rustic marlin]

chrome plugin?

are you talking about chrome debug?

[patent jay]

I’ll admit I’m not an expert in coding but maybe it’s that. My basis is a video recently uploaded of the plugin where codex can engage in many tasks opening tabs in an ongoing chrome session.

[rustic marlin]

it is not chrome debug

[patent jay]

This thing

[rustic marlin]

are you on mac?

[patent jay]

Windows

[rustic marlin]

it is for mac

[patent jay]

The video said it was for both Mac OS and windows

I can’t copy link it’s this one

[patent jay]

But I can’t get it to do something like this.

Maybe it’s my prompt and I’m overreaching? I’ll try a short one.

Or is it exclusive to the pro plan?

It seems I’m not the only one with this issue after checking Reddit

I can’t wait to get computer use plugin

[rustic marlin]

it was only on mac AFAIK

but if it is on windows, even better

[patent jay]

I can’t get it to work and it makes me profoundly sad

Maybe it’s only for pro (I have plus)

[faint jolt]

Hi all, quick question. Is the web searching capability disabled or altered when in the same session, MCP server tools are used? Before it used to work, but since this week the web searching is not doing as expected and limits itself to just searching arxiv and wikipedia. Any hints?

[rustic marlin]

it is working for me normally

[faint jolt]

did you try web search + custom app (mcp server) in the same session?

[faint jolt]

this is what the open ai help chatbot replied:
the behavior lines up with a product constraint: when a custom MCP app is enabled for the run, ChatGPT may not be able to use the full live web Search stack in the same turn, so the “search query” step can degrade (including returning “no results” or drawing from a very limited set like Wikipedia/arXiv).

[rustic marlin]

I often do.
i didn't used today, just woke up, but i use codex daily with some mcp tools. and often doing web search

[faint jolt]

maybe it's different in codex

[haughty igloo]

how to fix chatgpt unusual activity from your device?😑😑😑

[onyx tree]

Can anybody do a sora 2 video for me

[green summit]

Anything specific?

[onyx tree]

I rly wanna create a disney trailer about my friend ( i have consent for this btw)

[green summit]

Shoot me a dm with some information and I'll try to help you out.

[distant tulip]

Hi everyone, i am trying to use Codex + GPT 5.5 to work on a proprietary coding language. I basically created a plugin and named piped codex and the proprietary IDE and it managed to mod the IDE find its way to trigger reconciliation and created a feedback loop. It created also created bunch of skills recognizing those file headers etc. So my question, is this a good way forward by building on skills or there's better way to do it? Are there limitations with this approach?

[dull garden]

Is here a support chat somewhere?

My ChatGPT account shows PRO in the top-right corner and I paid for Pro $100, but Codex usage still shows an upgrade banner saying “Upgrade for 20x higher usage than Plus” and my 5-hour limit appears to be treated like Plus. Please refresh/re-sync my Codex entitlement and usage tier for this account.

Can any admin help here?

[charred parcel]

for support you need to go to https://help.openai.com

[tepid leaf]

how long does support usually takes? im locked out of my account which i paid $200 for in subscription. already 5 days waiting..

[charred parcel]

I had same day support a few days ago, I would poke them again

[tepid leaf]

okay, will do!

[livid cape]

hi

can anyone help?

when i use custom gpt with 4.5 (pro subscription)

the thing slows down :/

chatgpt just hangs, doesn't reply

this is on iOS

any help?

[green summit]

One of many IOS issues

[livid cape]

but its also bad on the desktop web app

[green summit]

Seems to be a very common thing lately. I hope it gets fixed soon.

[queen scarab]

i cant sign in with codex it asks for phone number but i dont get the code

any ideas?

[near void]

When I uploading files it shows error

[zinc lodge]

Hello everyone. I have been having issues logging into ChatGPT on mobile and on the web. I have been repetedly getting the below error:

error_code: get_chatgpt_account_error
request_id: 88887c82-d4f5-48ba-a34b-520841cd1634

I was in emails with someone at the support team but after doing what they asked I never heard back. Can someone help me?

[charred parcel]

you're going to have to sort that out with Support, nobody here can help with account level items

[zinc lodge]

is there a better way to reach support than generic email?

[charred parcel]

you can use the chatbot at https://help.openai.com and ask to be escalated to a human

[zinc lodge]

thank youi

[dry hatch]

i built some custom GPTs but it seems they cant do things like add things to google sheets or google docs is there a way i can change this or is that just how it is?

[tepid leaf]

ive got a similar problem, locked out of my account and still waiting for their helpdesk to help. I will def cancel my subscription after this failure.

[brave prairie]

Hey @Jesse! You can definitely connect GPTs to Sheets. It's the best way to make them useful for business. You just need to set up an Action using an automation tool.
I specialize in building these workflows and I've done this many times. It involves creating a webhook that the GPT can talk to. Happy to explain the steps here if you're still stuck!

[gusty moon]

Hey all, I updated to the latest version of Codex on Windows 11 this morning and am now getting this error when trying to open the Codex app. It appears to be the latest update as Windows Store shows it up-to-date.

Do I need to uninstall/reinstall, or is there an easier fix?

Edit: I'm running on WSL

[heavy oriole]

Anyone having this problem too? I cleared my cache like I was told to and now no matter what browser I use I get this. It won't let me make any requests, won't let me log out, won't load any history, I click on new chat, search chats, codex, more and it does nothing.

[brave prairie]

Glad the suggestion helped, @Jesse! I actually have a high-level workflow diagram that shows exactly how the GPT-to-Sheet data flow looks (including the error-handling steps). Would it be helpful if I dropped a screenshot of it here for you to use as a reference?

[brave prairie]

Hey @Jesse! I figured I'd just drop a teaser of that workflow diagram here to save you some time.
This is how I structure the 'Logic Gate' between a GPT and a Spreadsheet to make sure no data gets dropped if the API glitches. It’s a standard part of the AI Ops framework I use for business systems. Hope this helps you get those GPTs connected!

[zinc lodge]

I’m having a similar issue for my business account. Support ghosted me so I’m trying to get a new case open

[heavy oriole]

I think I figured it out. Would like me to dm you since the support's ghosting us? -_-

[zinc lodge]

Please, I’m willing to try anything at this point to recover my data. Thanks

[heavy oriole]

Np!

[tawdry dawn]

bro why is chatgpt such a bum, i told it to make an exportable file for the whole chatlog and it said it did what i asked, i looked at the file, had no more than 15kb of data. i asked it, said it was not the whole chat. and now, i cant use that chat anymore because appranetly i cant use chats with analyse used in them.

atleast gemini tells me it cant do it rather then faking it and ruining my chat

is there any way to like delete the analyse file so i can still chat even when i run out of my analyse and file perms because i need chat to help with SQLs in my backend

also when i use that session with chat for more then 10 messages a time, chat's ui becomes unresponsive until its done with its message which due to the lag and unresponsiveness, takes around up to 2 minutes for a simple prompt rephrasing

[floral forge]

Could someone give me a definitive answer on, if I have a ChatGPT Plus subscription that I usually only use for Codex Desktop, if I use the chat feature on web for planning/research, will that eat into my Codex ratelimits too?

[gentle vigil]

Keep getting this message

[brazen quarry]

Same here.

[distant shoal]

Hello guys. Wanted to try out codex.. getting this error. chat gpt cant help me 🙁

can someone help?

[brave prairie]

@Preston Thanks for the thumbs up! Are you seeing many founders struggling with the GPT-to-Sheets connection lately, or are most moving toward full API builds now?

[spare fulcrum]

I am experiencing an issue where ChatGPT completely stops responding. Every time I send a message, the AI begins to process it, but instead of showing the response, it immediately displays the error message: "Error in message stream."

Here are some details about my current setup:
Device/Platform: Handphone, Oppo Reno 14F 5G
Browser/App: Firefox & Built in Android App

Network: Home Wifi & Cellular Data.

What I have tried so far:
Refreshing the page / restarting the app.
Checking my internet connection (other websites load normally).
Has anyone encountered this specific stream error recently? Any advice on how to fix this would be greatly appreciated. Thank you!

[ashen mountain]

Problem with my paid Plus account after the global outage.

Symptoms:

ChatGPT won't send messages in any mode

Console errors: 404, 429, 500, ERR_HTTP2_PROTOCOL_ERROR

What I've tried:

Incognito mode, a different browser, my phone (cleared cache and data)

Key clue:

When I log out (guest mode) it works perfectly. So the issue is definitely with my profile on your server.

Request:
Please ask the engineers to perform a hard reset of my profile. I haven't been able to use my subscription for 3 days.

[onyx kraken]

Hi, we're all community members here, even the mods are volunteers. #server-staff . You'll want to reach out to OpenAI tech support, https://help.openai.com/en/articles/6614161-how-can-i-contact-support?

[ashen mountain]

I guess this isn't just my problem, and customer support hasn't responded. Has anyone else managed to solve this issue yet?

[onyx kraken]

It can take hours to weeks for support to respond, OpenAI does have hundreds of millions of weekly members using the services, so it can take time to help everyone who might need help.

If you're using a VPN or have browser extensions on both your phone and any other devices you tried on, I'd try with VPN off, and consider deleting browser extensions as some can affect problematically even when turned off.

Otherwise, you describe a decent range of potential fixes. I'd consider trying from a different device with a different internet provider, maybe. But chances are you might need OpenAI's direct help.

[trail cairn]

those errors combined can happen with shared connections such as vpn and would also explain why non logged in works as it has less restrictions

[bronze vapor]

Hello, I hope you're doing well.
I'm writing to you because I've tried everything on my own and with helper to recover my Sora data, but nothing has worked. Do you know if it's still possible to recover my data, and if so, how?
I want to clarify that since my Sora and ChatGPT accounts were linked, I immediately thought of downloading the entire contents of OpenIA via the privacy gateway, but I don’t think I see Sora in the files within the ZIP archive.

[ashen mountain]

I'm working without a VPN on my phone and computer. Support has stopped responding.

[trail cairn]

chatgpt and sora were different services but just shared same login system. if this does not work then there is not much you can other than contact support https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation#will-i-be-able-to-export-my-data-after-the-sora-app-is-discontinued

[bronze vapor]

as I said in my message I had already contacted the chatgpt helpers and they failed to help me 🥲

[trail cairn]

vpn was just one example of a shared network. people have had similar issues with proxies, in school environments, isp that shares public ip, etc.

[bronze vapor]

That's why I came here, all my attempts have been unsuccessful both alone and with the team

[brave prairie]

I see a lot of people struggling with how to handle "edge cases" in their automation. Here is Part 2 of my series on SOP Architecture, focusing on the Logic Flow that handles the chaos.
Most automations break because they are "Linear" (If A, then B). But real business operations are "Dynamic." If you want to build systems that don't break, you need to architect for the "Human-in-the-Loop" and "AI-Verification" cycles.
The "Intelligent" SOP Logic Flow:
Why this Architecture wins:
The AI Brain (Node B): Instead of just passing data, we use AI to identify the intent. Is this a refund request? A high-priority lead? A technical bug? The system decides the path before it takes a step.
The "Standard" vs. "Edge Case" Split (Node C): If it's a standard SOP, the machine handles it. If it's an "Edge Case," it doesn't just "fail"—it routes to a human for review.
The Feedback Loop (Node I): This is the "Elite" part. When a human handles an edge case, the AI learns from that action. Over time, your "Standard SOP" bucket grows, and your "Human Review" bucket shrinks.
Verification (Node F): We use a second AI call to verify that the steps were executed correctly before notifying stakeholders. This eliminates "Silent Failures."
The Goal: Move from "Manual Chaos" to "Automated Clarity."
I’d love to hear how you guys are handling "Human-in-the-Loop" transitions in your current builds. Let’s discuss!
#AI #Automation #SOP #SystemsArchitecture #NoCode #Make #Zapier

[trail cairn]

yeah but the default helper is ai bot. you can ask to escalate to a human if you didnt try that. no one here can do anything but suggest things but you said nothing worked without details such as error messages, etc.

there was a sora channel with export tricks but it might be gone by now or only visible with the archive role

[bronze vapor]

Is it an AI that answers you on laposte. net ?

[trail cairn]

no? support is at https://help.openai.com bottom right

[bronze vapor]

so it wasn't an AI that answered me, it was human beings ?

[trail cairn]

where did they answer? support bot and humans are only at that site or via openai.com mail

[bronze vapor]

because it was through my email that I was responded to

[bronze vapor]

Well they offered me various solutions like checking that I'm not using a VPN (I don't have one at all so they quickly responded) and other things like that but no concrete solution

😅

[trail cairn]

how did you contact them and are you sure you used the official sora and not some third party?

[bronze vapor]

I don't know which site I relied on anymore, it was a site on chatgpt in the official resources and at the beginning it was an AI who asked me for some information then told me that I would be contacted by email by a human person

This was done but unfortunately the help they gave me did not help me

[trail cairn]

login with your openai account and use the above site chat. if you are from france you can ask it again and say to escalate to human and say you want your data and solution is needed (gdpr)

[bronze vapor]

I was just connected before requesting help I had to connect

Gdpr you want say rgpd ??

[ashen mountain]

Here's the error: POST https://chatgpt.com/backend-api/f/conversation
Status: 503 Service Unavailable
Response: “upstream connect error or disconnect/reset before headers. reset reason: connection termination”
It has nothing to do with the network; the mobile network isn't working either

[bronze vapor]

I should do this on which chatgpt help site ?

[trail cairn]

no i mean gdpr and english message. if you do it in another language the person helping might use translator or be regional support that lacks the ability to resolve it

[trail cairn]

https://help.openai.com

[bronze vapor]

Uhm what is gdpr ?

[onyx kraken]

General Data Protection Regulation, the European Union data privacy law

I'm not very familiar with it, as I live in the US and am a community member like nearly everyone here, so it's not part of my job or life. But Radium recommends you mention it as you request help from https://help.openai.com/en/articles/6614161-how-can-i-contact-support?

[bronze vapor]

the site was able to display for a few seconds before giving me an error message

[bronze vapor]

Ok it's rgpd in france

Thank you for explaining to me

[bronze vapor]

this site is displayed for a short time then displays the same message that said I now know that I used this one

this site is displayed for a short time then displays the same message that said I now know that I used this one

[onyx kraken]

Happy to try and help.

Are you able to reconnect to the help website?

You can also publicly report bugs here, #1070006915414900886

Note you must click on the first post, read it, and follow its instructions to even make a bug report, it's not like just typing a message into this channel, but we can make bug reports there too.

I would use that if I can't get the website to work at all, I would still use the website help system for my data export request assistance if at all possible because that is personal, not public. Bug reports on this Discord can be read by everyone, your https://help.openai.com/en/articles/6614161-how-can-i-contact-support? chat is private to you and OpenAI tech support.

[onyx kraken]

What happens when you click 'Go back home'?

[bronze vapor]

It brings back to https://help.openai.com/en and then gives me the error message and continually takes me back to this page when I click "Go back home"

[onyx kraken]

I would restart my device and try again. That sounds like a cache issue or similar problem. I'd consider turning a VPN off if I was using one, I would consider deleting browser extensions if I have any on my device (they can affect and cause issues with interacting with OpenAI services even turned off, so delete and reinstall after is a good idea if that is the issue).

[bronze vapor]

But then I used this site ( https://help.openai.com/en )
to contact the humans on the team and as a result I didn't get much help ( To download Sora's data )

I've been asked this and it's not the first time I've been asked this even though I've re-explained it several times.
That I don't have any VPN, no ad blocker, no extension

I'm trying on the computer

[onyx kraken]

If you mean 'The OpenAI Discord' as 'this site', we are nearly all community members, like yourself. Even the mods are volunteers, only the incredibly few gold name accounts actually work for OpenAI, #server-staff . OpenAI doesn't run tech support through this Discord, and the people answering you here, like myself, have no idea what you said before unless for some reason someone is so curious and interested they use the Discord search posts from you feature, I sure hadn't done so (still haven't), I'm just trying to help you here in this chat.

Great that you have no VPN, extensions (not sure that ad blockers matter, never saw anyone discuss that). I'd restart the computer and see if I can connect to the website properly, as when I run into problems like that a comp restart often fixes the issue for me.

If it doesn't fix, I'd #1070006915414900886 the 'I can't connect to the help website' issue, with screenshots and a description of the problem in detail, timing and such, that you're connecting from France but not your IP or other personal identifying details, and try again in a few hours/tomorrow on the help website.

The help website is how OpenAI handles tech support, and the company surely attempts to track issues with access to it and get them fixed immediately.

[bronze vapor]

I put the site I used I wasn't talking about discord

[bronze vapor]

Actually it was on my phone that it doesn't work so now I'm going to test on the computer

It does the same to me on the computer unless I don't click on cookies

[onyx kraken]

Okay. This laposte site you mentioned, I don't know anything about it. As far as I know, it's not connected to OpenAI, so I wouldn't expect help from it. Researching, it's the email service ran by France’s national postal/mail system. I do not know of OpenAI using that system at all. The help file about official OpenAI communication, which I understand maybe you can't open right now, is https://help.openai.com/en/articles/11725090-verifying-communications-from-openai

And it shows this info, among other less specific to your situation (I think) info:

As a customer of OpenAI, you may receive emails from the following domains:

@c-openai.com

@mail.openai.com

@openai.com

@sales.openai.com

@tm.openai.com (used for workspace and GPT invites from [e-mail])```

It does not list laposte at all.

[onyx kraken]

Are you able to click on https://sora.chatgpt.com/sunset and get that link to work for you?

If so, click on the 'export' button, does that work for you?

[onyx kraken]

I've offered all my ideas for your 'how do I contact OpenAI support, it's bugging out' issue,

But you also have a 'how do I get my Sora data?' issue.

This is how we are instructed to request an export of our Sora data:

https://sora.chatgpt.com/sunset and click export

Does that work for you?

[bronze vapor]

To explain what I told them I explained to them that when I made the request to export my file never and I checked everywhere even in spam even in junk everywhere I checked everywhere I never received an email telling me that my file had been exported and when I returned to the site to make a request again and it showed me that and still to this day it clearly shows me the same thing

[onyx kraken]

I do not understand why you are being shown that 'the Sora application will shut down ... ' message.

This is what I see at https://sora.chatgpt.com/sunset

[onyx kraken]

If you are seeing something different, that is a very appropriate Discord #1070006915414900886 to make, because that does not seem correct based on what I see when I click https://sora.chatgpt.com/sunset

[bronze vapor]

yes I saw that also the screenshot shows what happens once I click on export

Besides, they offered me to export again

I click on the button

I hope this time I will have the file

[onyx kraken]

I would try the export again, or re-export button.

Sometimes when I request my regular ChatGPT data export, it does not appear. If it's been more than 24 hours, I request a new one. Sometimes it takes a few days to successfully give me my data export, this has happened across years. It either works, or does not work. If it does not work, I try again. It eventually works.

I strongly believe trying the re-export button once every 24 hours is the way to get your Sora data export, possibly after a few tries.

[bronze vapor]

before there wasn't this button so I couldn't sit down again, I'm waiting to see if I receive it tomorrow

Thanks youuu

[onyx kraken]

I really hope you get it tomorrow!

With my ChatGPT data export, sometimes it's taken quite a few tries, I think across a week or more once. But I did get it after that. If I was in your situation, I would try 1x per 24 hour period until success.

I don't know about Sora data exports, but I would expect they are similar to ChatGPT data exports, and likely succeed or fail in similar patterns.

[bronze vapor]

I hope too

Thank you

[strong eagle]

I’m trying to understand when the “Open in App” selector appears in the top-right corner and when it doesn’t.
In the later case, there is code generated in the underlying folder but the selector doesn't show up in the Codex UI. When I start a new chat, I can see the drop down. Does anyone know if this is intended or a bug?

[tight shoal]

what's the proper repo for making requests for the vscode codex extension?

[merry oar]

Try Plus free for 1 month not working

[junior sedge]

Using Plus Free for one month will result in a hidden ban via phone verification.

Consequences of the crackdown on bots creating large numbers of GPT Plus accounts being sold on Telegram.

[hybrid coral]

does codex have only gpt models ?

[onyx dew]

Hi everyone,

I’m currently unable to access ChatGPT due to a persistent login error:

“There is a problem with your request. (401,)”

My API access still works normally, and I never received any deactivation email from OpenAI.

I’ve already contacted support and submitted an appeal, but I wanted to check whether anyone else has recently experienced the same issue or knows if this could be related to an authentication/risk-control problem.

Any advice would be appreciated. Thanks!

[trail cairn]

try to login in incognito mode?

[onyx dew]

I tried.

Show this error

I've been using ChatGPT with this account for more than a year, and suddenly it just show the error and looks like lost the whole world. And no human support can reached to help......And I also have Max subscription.

[trail cairn]

that looks one of their apps and not browser?

[onyx dew]

Yes, Desktop app.

[trail cairn]

try in safari private mode

[onyx dew]

Browser can login but no error and type no reponse.

[trail cairn]

what does the browser error console or network tab show when you refresh the page

[trail cairn]

have you installed some network, privacy or safety apps that may alter network traffic? since both browser and app are affected

[onyx dew]

No, I'm in AU, and my laptop is developer environment, no privacy and fasty apps. And no VPN as well. Pure research usage.

[trail cairn]

does your account work properly on some other computer?

[onyx dew]

No matter how this issue was resolved on OpenAI’s side, this incident became a very real reminder for me as an AI researcher/engineer:

No matter how powerful cloud-based AI becomes, once the connection, authentication, or account access breaks, a huge portion of your workflow, memory, and accumulated context can instantly disappear.

This experience honestly made me appreciate the importance of local inference engines and self-hosted AI systems even more.

Cloud AI is incredibly powerful.
But true reliability may ultimately require hybrid systems:

• cloud intelligence
• local memory
• local fallback inference
• user-controlled persistence

This was a surprisingly strong real-world case study for why local AI infrastructure still matters.

Looks like mobile and desktop all not working properly now

The good news is support human get involved and start to helping me resolve this. Hopefully it not taking couple days.

[merry oar]

the offical one behind chatgpt? im using it with no issue

[rustic marlin]

did you used ai to write this message?

[pine flame]

how can I stop really long chats from freezing my browser? is there some sort of functionality that allows me to delete the for some reason still rendered or loaded parts of the chat that are just outdated ?

[sour pollen]

Hello, I have a question about usage limits - if I have the Pro 5x subscription, what exactly are they? Especially for the Pro model

I dug around the OpenAI website and it says that Pro users have unlimited access, so long as they don’t abuse it, but at the same time the Pro 5x subscription says that, well, you get 5x the Plus usage!

That’s a contradiction

[pine flame]

what falls under abuse??

[sour pollen]

Yesterday I also received a pop-up saying I’m out of Pro usage until (or after?) May 24, but it’s the 23rd and I was able to access the Pro model today

So that’s two contradictions at once

[sour pollen]

🤷‍♂️

[pine flame]

what did you do to lose accsess ??

did you try parsing wikipeda or smthn? 😭

[sour pollen]

The Pro model

[sour pollen]

No lol

I asked about Roman grain surpluses

[pine flame]

what?

[sour pollen]

I asked it how Rome could afford grain doles

A very basic historical question

I just wanted a thorough answer

[trail cairn]

the 5x 20x is more related to codex and does not affect much in chatgpt other than some things

[sour pollen]

Does “some things” include the Pro model? The pop-up specifically said I’m out of usage for the Pro model

[trail cairn]

dont think so as the lower plans dont even have it

[sour pollen]

Maybe I forgot to take my meds

[sour pollen]

Ok, the limit is back

Is there any documentation on what those limits actually are?

[trail cairn]

doubt it

[sour pollen]

At least Claude shows you your usage 🫠

[sick niche]

Does anyone know what the ChatGPT deep research capabilities actually are? I've been trying to get it to help me work on a physics engine, and I was previously using Claude, and Claude was able to view github repos and view code directly, but ChatGPT keeps making notes in its documents that it wasnt able to access some repositories code so its mostly speculative???

[scenic portal]

Error running remote compact task: { "error": { "message": "Invalid 'input': array too long. Expected an array with maximum length 16384, but got an array with length 18006 instead.", "type": "invalid_request_error", "param": "input", "code": "array_above_max_length" } }

how to fix this in codex?

[gritty sage]

hi, guys, my business account invites one, shows that : Team has a delinquent subscription, how should i do ?

[hollow jetty]

does go plan have more codex usage than free?

[hollow jetty]

ask codex how to fix bro

[scenic portal]

codex isnt working itself

[hollow jetty]

then ask chatgpt...

[mellow helm]

How to get free credit in seedence 2.0?

And are there any better video generators, something like sora for example?

[candid kiln]

Sora is gone

And for the other questions other channels may be better than codex

[green summit]

Well they said "Something like Sora", so they know.

[candid kiln]

„And are there better generators (compared to seedance) - like sora for example?“ is how I understand it.
Anyway not really the point - it’s the wrong channel, I think that #ai-discussions is probably more informed on this type of thing than codex

Duh sorry

🤣

Not sure how I ended up in the community help channel lol
@mellow helm sorry you obviously at the right place - I somehow totally messed up channels lol, thought this is #codex-discussions 🫤

[green summit]

Nah you're good. Tbf, you're also right. Since they're also asking for/ talking about non OpenAI products.

[urban plover]

How does one get assistance that is not an ai agent from this company. I had some email thread created through their own portal, ask me to log in and start another chat to validate my identity despite when I log in the entire email thread is pumped out into their chat window....

I just get canned responses claiming I need to do some nonsense repeatedly and apparently escalations do not exist because ai is flat in reality.

[dapper basin]

Ask it to elevate you to a human responder.

[urban plover]

did like 5 times, "human manager"

eventualyl it just stops responding

[dapper basin]

It sounds like a conspiracy to delay and obfuscate responding to you in order to keep banking your subscription.

[urban plover]

guess they will fall in alignment with every other service provider offering inhumane levels of support

[candid wedge]

yo

anyone could help me what is going on?

[dim ermine]

hello there!!

its not working, tried 2 cards. working on other totally normal. (international transcations)

but not on chat gpt? 1st one had error of payment not approved while there was not any payment to approve? 2nd one is being declined.

@onyx kraken or @analog spear if you guys could help!! would appreciate

or any other staff!!

[analog spear]

Hi! We, or any other Discord Staff member cannot help with this matter, please contact https://help.openai.com , as they are the team able to assist you the best ^^

[onyx kraken]

We're all volunteer community members here, like yourself. Nobody on this Discord can help with a billing or other account issue except with helpfiles.

https://help.openai.com/en/articles/6614161-how-can-i-contact-support?

Reach out to OpenAI tech support there.

https://help.openai.com/en/articles/7232916-why-was-my-credit-card-declined may be informative too.

[onyx kraken]

This helpfile might also help: https://help.openai.com/en/articles/7242622-why-did-my-chatgpt-plus-or-chatgpt-pro-renewal-transaction-fail

[dim ermine]

Alright, thank you @analog spear and @onyx kraken. :))

[modern aurora]

hey everyone! i’m gonna get the feedback ball rolling and say i would LOVE the option to embed the text of the prompt i used into the file name! that way i would never lose my place

i’ve got so many ideas, i will have to just bullet point a few here bc i’m so excited

a scrollable viewport for inpainting is the dream, but i understand this would take some time

down the line, i’d love the option to save my favorite “primers” (which is what i call the giant pile of style adjectives) as little “favorite styles” buttons

[pallid sedge]

I’d love to hear more what you mean by “never lose my place”! Can you tell me a bit about your workflow and how filenames would fit in?

(We started exploring designs around saving favorites, so I’m wondering to what extent save-related features would solve this problem)

[modern aurora]

sometimes it can be hard to remember what prompt i used for a specific pic long after i’ve saved it to my camera roll? so having the file name contain the prompt would enable me to simply check each pic to remember what wording got me that image 🙂

also, regarding saving: currently, on ios (starting at the “results” screen for a given batch) it takes 5 user actions to save any given photo:
(tap desired photo,
tap download icon,
tap “view”,
tap-and-hold image,
tap “save to camera roll”)

i am wondering if there could be a way to reduce this so that you can simply tap-and-hold any given photo from the results screen and get the “save to camera roll” option, or maybe just making the download icon trigger a download to the camera roll instead of safari downloads (where it currently goes), as that menu can be hard to manage on ios. i may be explaining this poorly, and if so please let me know and i’ll try to rephrase.

[pallid sedge]

is tapping “view” a safari thing?

this is what i see (tried attaching a video but it was too big of a file)

[modern aurora]

ah! yes it must be safari!

[pallid sedge]

(I def agree that anything more than 2 taps is cumbersome! Was just wondering where the extra step came from)

[modern aurora]

my dream would essentially be to try and see if the “save image” could be accessible via tap-and-hold on any give image from the grid screen

[pallid sedge]

same — this is a technical constraint (webp file instead of png) but we’ll keep investigating!

[fiery drum]

A smaller diameter option for edits would be 🙌

[jade dagger]

How much of the prompt is used in variations? Anything at all?

[tight warren]

https://youtu.be/Sm1HV9nJYBY I've gotten nearly 500 prompt requests and I want to answer all of them, wondering if anything can be done here 🤔

It's time for DALL-E bulk beta, even if it starts with sharing a link to a CSV file hosted on AWS

[pallid sedge]

none — variations are image as input and output!

[pallid sedge]

we’re working on getting a lot more people access — which should hopefully reduce the load here! Eventually we’d like to give access to anyone who wants it (once we figure out safety & gpu constraints)

Question for you all: what were some prompts that got blocked by the system that you found confusing or especially annoying?

[tight warren]

This is the end solution yes, thanks for sharing. I think there's still need for parallel requests/bulk at an individual user level though. I'll see what I can do in the meantime.

[pallid sedge]

one other thing you could do for now is to direct them to the instagram account @openaidalle (managed by @lusty creek ) — that account gets a lot of DM requests and we try to respond to as many as we can

[tight warren]

Thx

[jade dagger]

Adding "ecstatic smile" to a prompt got my results blocked. I see where it's going, but was surprised nonetheless. It had no problem with "delighted smile"--but the part that has me more concerned is that I might get automatically booted, especially when I don't know how many strikes I get

[fiery drum]

I’ve had prompts blocked when describing emotional states for various statue series- regardless of positive or negative emotion. Like Michael, I am also very hesitant to re-prompt because I don’t want to lose access

[pallid sedge]

Super helpful — thank you! We started w erring on the side of the model being overly aggressive, but are working to improve it (which is where your feedback would come in). FWIW if you do get blocked but we don’t see any abuse we’d happily reinstate it (although it would still be annoying for you to write in / wait etc)

[fiery drum]

Thank you. I will definitely start noting any future incidents

[south fox]

In my case some names and terms were blocked, like "in style of David Lynch" (because of "lynch") or "Bullet Time from Matrix" because of "bullet".

I found a workaround though, sorry for hacking: by writing the terms together, like BulletTime the system understands the prompt.

[regal solstice]

In just 4 weeks of using DALL-E 2, I've saved 4gb of my favorite images to my computer. Is this similar for others? A quick 'bookmarking' tool would be nice, so I don't have to save every image I'm into. (It's not always bc I love the final image, I'm often saving them as a mental placeholder, then putting the corresponding link to the image into a spreadsheet, and copy & paste the prompt, which tedious and time-consuming). I'd want my 'bookmark list' to be private, unlike the share button, which is more for 'final pieces'. Curious what your thoughts are on saving / managing images.

[native rapids]

curious to know how long the images cache for -- I've noticed I can open up the (non-public) URL days later and the images are still there

[regal solstice]

I'm afraid to ever clear my browser history again, since I don't want to lose access to my all-time prompt list 😬

[tight warren]

Yes my camera roll on my phone is crazy full, outnumbering even photos I took IRL at this point

There's something about DALL-E it turns me into a hoarder, interesting psychological phenomena to think about

[regal solstice]

I relate to that—it's like you're collecting specimens that otherwise you may never see again! It's also made me jot down more notes about styles / scenes I'm encountering IRL. So I'm like bookmarking things in the real world to play with on dall-e later.

[pale shell]

curious to know how long the images cache for
images are stored indefinitely at the moment, so you can also bookmark URLs instead of downloading.

improving the image saving process is definitely on our radar!

[hazy quartz]

one bit of feedback is that the model seems to have some trouble keeping track of too many elements. a friend has a family band called A Pocketful of Hedgehogs, so I was trying to make a logo for them. it's pretty easy to get hedgehogs (cute ones!) in a pocket, somewhat harder to get exactly four of them, and really tricky to get four hedgehogs wearing hats peeking out of a jacket pocket

[tight warren]

I just dropped this new article on what I'm calling, "Recombinant Art" which I think is a good description of the kinds of images we're seeing users generate with DALL-E 2. In this article, I attempt to come up with a term and definition, describe this phenomenon and where it comes from, and also try to explain how it fits into the existing digital media history/landscape we've seen up until today. I also briefly answer what could be the next wave for future multimodal AI systems. Please let me know what you all think, I'm eager to hear everyone else's thoughts as fellow beta users https://bakztfuture.substack.com/p/dall-e-2-recombinant-art-and-design

[steel torrent]

i've noticed that dalle loses track of the "by [artist]" or "in the style of [xxxxx]" as elements are added to the prompt. it seems to revert, in all cases regardless of the "by [artist]" to the same homogenized "generic oil painter" look. for instance, i could start out with beatrix potter as the artist, and as I add more elements, beatrix's style slowly is replaced by a generic oil paint look. has anyone else noticed this?

[steel torrent]

could this be tricky on really long prompts? win vs mac char limits? i suppose it could just truncate BUT i would rather see a library IN dali where i can save images to by clicking something on the image. from there, my own curated private collection of work / work in progress, i could click "download" and get the image but also see the prompt on the screen, maybe with a "copy" javascript shortcut.

one other idea that MIGHT work is to embed the prompt in the metadata so that it's easily discoverable by looking at image properties, photoshop, etc.

that said, i'd prefer to have a private gallery inside a dalle UI, with some folder options for organizing after they have been saved from the generation part of the UI, where i could save this work.

but in the short term, a simple save button might be nice to include the prompt in metadata. my understanding is that images shared on social media have all their metatdata stripped out for security reasons (like long lat, names, etc). if you shared images directly on your own hosted CMS like wordpress, that would be up to you to cleanse, if it were to be embedded "on save" to your local pc.

my workflow for the most part is to create a local folder on my machine, and save images there with the folder name and a number in the file name, and then create and update a text file to keep track of the prompts. it's not a one-to-one relation of prompt+image name because it doesn't need to be. there are many prompt tweaks, and then inpainting which can change the prompt. the main thing is to keep track of the core phrases, or "seeds" that get the gist of things so you can pick back up where you left off. locally, it look like this:

i just toss a text file in each folder. i was doing it in apple notes but that get's messy and easy to lose. i could use excel but i really dislike the formality 🙂 a free form note seems to track all i need to stay organized and productive.

[steel torrent]

webp has been more of a hassle than benefit in my professional web work... 😦

[steel torrent]

i got knife and gun blocked, but both were in the context of using dalle to inpaint on existing comedic cartoon / pen and ink art i was uploading. the "you will be blocked" message that pops up is definitely a little stressful!

[steel torrent]

the messaging is definitely a four-alarm stresser. I run afoul of the language algo and pivot to wondering if I'll lose my access 😦

[steel torrent]

i'm on 3 strikes and am nervous...it actually affects how i'm writing prompts. i'm trying to say without saying kind of thing 🙂

[steel torrent]

i have done similar things 😛

[steel torrent]

we could have a whole big chat on this. i'm a software / web UI UX designer and developer and spend lots of time in figma. would be glad to lend a hand with wireframing some concepts if we had a good text list of desired features rounded out. @pallid sedge

[steel torrent]

i find myself more looking to breaking down image construction to steps, big to small, and try to avoid inpainting but also build the prompt so that i'm ready to try inpainting at any step where prompting begins to falter. this is the step where i usually get to 5-8 sets of 10 and realize the prompt is not going to work no matter how many times i click "again" and then cut over to prompting.

[jade dagger]

Always love your observations, @tight warren! One of the new experiences that I'm getting from using DALL-E is the feeling not of making art exactly but of exploring it, like it's a terrain I'm visiting for the first time, and I have to look around and get to know it before settling in on a trophy image to take home. A lot of this has to do with being able to make variations on variations. I'd pour dev into expanding the UI of this.

Another thing that's hit me is that I feel like my art execution skills are being used less, but my art discernment skills are being used more than ever. I like this feeling. Imagine the world where millions of artists have incredible artistic direction chops from having weighed in on so many ai results. It's gonna raise the bar

[jade dagger]

I have a feature request, is this the right place to put it?

On the loading screen, I'd like the option to keep up the original image (instead of the gallery slideshow).

Often I'm in a mindset of contemplating a particular image and I find the 20-30 seconds it takes for generation to be not long enough to go do something else, but if I stay the slideshow takes me out of concentrating on the image and the world of experience it's invoking within me.

If you could add this as a toggle it would be a material improvement to my current exploration process. Thanks for considering it!

[tight warren]

Thanks Michael! 100% I feel like all great art already exists, I just need to enter the right terms to discover it

[tight warren]

Creative acceleration!!!! Artistic acceleration x 1 million

[regal solstice]

I relate to this concept of 'visiting terrains' — I keep relating it to the feeling of being inspired when I'm at a museum, except in the case of dall-e I can visit really niche and surprising museums (and freely hop between them)! It's also expanded my exploration / paying attention to specific details of styles out in the physical world (I'm doing way more visual notetaking than usual).

[modern aurora]

on this note! for anyone who haven’t tried my game, it’s basically this exact concept as a game!

the game is called treasure-map-telephone:

you write out a long prompt and you send that prompt alone. no pics. that’s the treasure map!

the other person’s job is to enter that prompt and see where it goes! and then, the best part: explore that area of latent space! make ones own changes word by word, then send the prompt back to your friend nice it has been sufficiently “telephone-d”

my DMs are always open if anyone wants to play.

[modern aurora]

i am expanding the claim of my “digital art” tip to include photographs. anything that is NOT a normal photograph should have “, digital art” to boost coherence. especially anything remotely imaginative at all.

this now stacks with “photograph of” and actually makes the resulting images MORE photorealistic compared to an imaginative prompt with JUST “photograph” as the style.

“a photograph of a dinosaur mech, digital art”

seems to perform better, even as far as photorealism, than the same prompt without “digital art”

again i will attach examples.

here’s “photograph of a dinosaur mech”

as we can see, all miniatures. it is trying to consider what a “real” photo of a dinosaur mech would be. almost certainly a toy by dalle’s logic.

here’s the same prompt, ending with the classic “, digital art”

much more menacing!

but still looks as photorealistic as i would expect considering the subject matter.

enjoy!

[errant rock]

Hi! Why don’t we have access or ability to batch download or make a training generation ? I notice that Aditya is posting mostly generation frame animation videos and they look stunning I want to be able to make them as well. https://twitter.com/model_mechanic/status/1513588042145021953?s=21&t=BNSQYsStnVzekfedIobx2w

Aditya Ramesh (@model_mechanic)

Modernizing a victorian house with #dalle text diffs

Likes

36745

Retweets

4401

▶ Play video

Twitter · Apr 11, 2022, 6:41 PM

[modern aurora]

this is actually not a training video! this is interpolation from text-diffs. a feature that will be added soon!

[modern aurora]

very exciting tho

i am also foaming at the mouth for text diffs

[errant rock]

Can you talk more about this feature or link to more reading on it?

[paper cedar]

yes, i feel the exact same way. it's the deep-seeded (seated?) fear of losing artwork/information to the sands of time...

[paper cedar]

This is exactly what I was thinking when I went through the file-name idea in my head. Embedding the prompt in the metadata of the png feels like the perfect solution, and reasonably straight-forward to implement (?)

[tight warren]

Makes sense, but this triggered the content filter

[split orchid]

@lusty creek Good day! I wanted to ask what should someone do when he/she has an access to DALL-E 2 from a country where this AI is not available?

[lusty creek]

Hi! Are you having login issues? If so, LMK and we'll work on it with the team.

[split orchid]

Yes, it shows that dall-e is not available in our country :(. I waited for so long with an excitement and whe i finally had an access... that pops up. Would be great if you help to solve this problem.

[lusty creek]

Could you email me about this? We'll connect with the right folks. [e-mail]

[split orchid]

Okay, thank you!

[split orchid]

Sent ✅

[steel torrent]

any other solution is going to require 2 files on download, which means a zip, which means more steps. ugggh. if i had to choose today, i'd pick prompt in metadata until i had better idea...

[pallid sedge]

small update: #announcements message

[errant rock]

noticed that the descriptor of "woman" outperforms the descriptor of "man" in many different use cases. Wondering if this is an embedded bias but found it interesting and notable

[regal solstice]

I noticed in the recent UX change to prompt history that if you open up an old link, it won't show up on the 'recent' tab. Instead it will show the history from your previous sessions, which are unrelated to that prompt search. This is a little jarring, since I've moved on from that session.

I liked that in the previous UX, I could have multiple DALL-E tabs open, each with their own journey

If I start with a link to an old prompt to continue my exploration in that zone—that starting link will not be saved when I enter the next prompt. (I think in the previous UX, whatever link I opened would automatically show up in the 'recent' list).

[steel current]

wondering if the feedback for bias or offensive could be split up into two different buttons? i am seeing a lot i want to flag for bias but not much for offensive (comparatively). might make backend reviewing easier?

[willow meteor]

I'm finding myself getting warnings for things that don't seem to come close to violating policy, any ideas about that? i.e this search "a precocious, silly, cantankerous, loud, easily annoyed, clown"

[errant rock]

I mean I can see how that would fail cus it’s too many modifiers on one subject but interesting that it’s a violation. The sum of the sentiment is pretty negative so maybe you just tricked the negative sentiment trigger on the NLP lol. Or maybe violations happen if you create too many processing errors? Lol.

[deft silo]

Same here, for instance "Two dogs picknicking in front of Mont Saint Michel, digital art"

I have no idea why exactly

Would be nice of DALL-E could highlight what it thinks are sensitive words in the same style as what GPT-3 does

[deft silo]

Also noticed that when some words can have a double meaning (but are perfectly fine in the context you write them) they are still not accepted

[south fox]

Here is one interesting thought from my timeline. Any ideas?
https://twitter.com/vovahimself/status/1527972337366204416

Vova “words are a motherfucker” Zakharov (@vovahimself)

What do you think of a really minimalist website where people can put (and vote for) #dalle generation requests, and those who have acces would upload generations (for free perhaps)?

As a key holder, would something like this be fun to you? @bakztfuture @Merzmensch?

Twitter · May 21, 2022, 11:19 AM

[junior crypt]

Interesting idea for sure, I'm sure alot of people would be interested in that sort of stuff 😁

[deft silo]

Maybe it would be nice if it was possible to include the prompt in the EXIF description field of the generated images (maybe as an opt-in option), this way you can be sure you won't forget what prompt you had used

[worn folio]

Hi team, access is not available in Ukraine. Please can you help with this problem?

[iron mirage]

Is a side profile of a face like this ok to share? Working on a surreal photo series

[boreal kelp]

i think this is to realistic, but im not sure

[primal lichen]

Does somebody had problem with singing in Open AI. It says that it's not available in your country. I tried using VPN, but it didn't help. Can somebody help me?

[iron delta]

I'm not getting the verification code. Might be my country?

Literally just got the code

But it's no longer valid

False Alarm

[lusty creek]

Hey Arsen! I got your email. We hope to get back to you on this mid-next week

[lusty creek]

Hi! Yes, this is too realistic to share. Thanks for checking!

[primal lichen]

Ok thanks

[worthy gate]

Hi, I have a quick question inspired by this bullet point on the recent OpenAI blog post:
"Less than 0.05% of downloaded or publicly shared images were flagged as potentially violating our content policy. About 30% of those flagged images were confirmed by human reviewers to be policy violations, leading to an account deactivation."
Does this mean that simply downloading an image that might possibly violate the content policy (such as the one shared by Luke above) could get our accounts deactivated, even if we never share the image?
For example, I uploaded a picture of my face, generated variations, and then downloaded some of them. Obviously I won't share them publicly, but I'm concerned that human reviewers might see that I downloaded them and flag my account

[pallid sedge]

oh that sounds totally ok!

We only ask that you not share photorealistic faces, so generating / downloading in the context you specified is totally fine.

To get a bit graphic (please stop reading if you’re in a wholesome Saturday mood which is what I wish for all of you), the content policy violations that led to account deactivations were things like lots of hentai & furry porn-adjacent content (when there was a clear user intent in prompts to generate that kind of content), repeated (albeit blocked) attempts to get dall-e to generate child porn or political disinformation, etc.

Although we had to err on the side of using “scary” language / aggressive filters as we scale to lots of people, we promise we’re reasonable when we review violations & freezes!

[boreal kelp]

well there goes my wholesome Saturday mood 😃

[pallid sedge]

told you to stop reading 😾

[worthy gate]

Thanks for the info – that's reassuring! I appreciate how hard you guys are working to make it safe for everyone

[young lagoon]

This is actually very reassuring. I generated a greek statue yesterday and had it try to create variations but got a warning for it and I was kinda concerned. I didn't try to make more but I'm glad to know that the people deactivated were clearly breaking the rules.

[deft silo]

Hey! A few questions regarding what's authorized:

• The rules say that no drug use is allowed, but are smoking cigarettes/cigars and drinking alchocol considered drug use? Say I ask for a drawing of a koala drinking a beer would it be ok or not?
• Is it ok if there is potentially copyrighted content in my prompt as long as it stays personal? For instance if I cite a TV Show movie or game name / character name is it fine as long as : I either don't share the output online or I can clearly see the output is far away enough from the original that there wouldn't be any issue?

[boreal kelp]

Is it maybe an idea to create a new discord channel for questions and general chat?

[pallid sedge]

• re: beer, only if the koala’s over 21 years old in the US or 16 in Germany 🙂 (all jokes aside yes alcohol / cigarettes sound ok)
• yes that sounds fine; the rule of thumb here is not violating other people’s rights (e.g. profiting off of someone else’s creative work, hurting their reputation, etc)

[pallid sedge]

ooh good idea! Let me make a general chat channel right now

[deft silo]

Thanks !! 🙂

[deft silo]

Is the cutoff date of the training data public like for GPT-3 ? Does DALL-E 2 know of recent events ?

[pallid sedge]

I don’t think it is, sorry! Not sure how recent you might be thinking but let me know what you end up trying!

[deft silo]

Was specificallt trying The Queen's Gambit based digital art

And it won't output anything related, just chess pieces ahah

[deft silo]

Something that would be nice with inpainting — but I have no idea of the technical feasibility — is if it could use your selection to make the main changes you require but if it could still modify the rest of the image to make it better match what you inpainted

So if I add a candle next to a subject, it gets lit by the candle etc...

[modern aurora]

i use a couple methods to mimic this. one of my favorite things to do is replace light sources with inpainting, but keep the light bouncing off of the environment. this works especially well with colorful LED lights. ill attach some examples

[modern aurora]

all of these are made by inpaint-replacing a light source but not its immediate surroundings. typically i zoom in very close to inpaint, then i composite/merge the image together in photoshop

[tall kernel]

This one took awhile to generate the issue I had is that it would do either the "coffee with milk designed as a rocket" or "doughnut shaped like the moon" but it wouldn't do them together as one prompt by saying one "next to" the other. To get around this I had it generate 2 coffee cups then uploaded the picture again and had it replace the coffee cup with the moon doughnut. Super impressed it was able to do that!! Seems it's not great at multitasking two unique things described next to each other in one prompt or maybe I'm doing it wrong. So impressed by the outcome though this is awesome

[regal solstice]

Wondering if it would be possible to save 'Original' images? I sometimes will do a lot of inpainting, then I can no longer find where the original came from, because of the overwriting of results in the new Recent layout.

[charred imp]

I don't know if something like that counts as a bias in the model but I asked for a "portrait of" + wikipedia definition of dalle (portrait of an artificial intelligence program creating /that creates images from textual descriptions ) and it only resulted in portraits of white people

[restive lake]

Nothing crazy to say just wanted to say that I'm having so much fun making images. The AI gets lost here and there but a few variations down the line it's usually okay

[south fox]

A question regarding our threads:
should we create a thread where we will show entire projects, we used DALL-E for? To separate it from this thread where we share our DALL-E creations?

[deft silo]

Thanks for the tip, really nice technique indeed!

[deft silo]

Found a small bug : if you start changing the prompt on a page with results, then click a result to zoom on it, then go back to the page of results, it resets your changes to the field (even if there is no page reload). Seems like a weird interaction flow but its more common than it seems as I already encountered the issue multiple times

[odd crater]

I like the "generate variations" function and most of the time its working very great. But at some point I wished we could have more control over it. For example I uploaded this picture of a penguin robot 3D model I've made several years ago. I generate variations of it 5 times and got many cool penguins in the same ambiance, but non of them looked that robotic. I understand it's because the original picture isn’t that robotic as well. There are just a few details like the legs and the shiny material. It's a really specific case and no necessary feature but I had the idea, that it would be nice to have more control. Be able to use the text description as well while generate variations would be great (in my example the words plastic, metallic or shiny could help). Or having a slider to control how close the variations are to the original would be a great feature too.

[glacial canyon]

i agree adding an optional description for variations could be a nice feature

[modern aurora]

this is coming 🙂 its called text diffs

[clear sorrel]

Anyone know why sometimes it spits out noise?

2nd image looking sus, ngl

i guess it doesnt recognize illuminati, i replaced it with cult and it works

[sleek edge]

I just got warned for the prompt "A computer evaporating into millions of little stars in space", and I can't tell what's wrong with it?

[deft silo]

Getting a content policy error for: "Stutenkerl aliens invade Paris, 1960's newspaper cover"

[clear sorrel]

I got a content warning for anything nuclear

the weird thing is, it’s not that hard to make photorealistic faces

[rugged wing]

I always seem to get them on accident

with things unrelated to the prompt

[regal solstice]

💭 Meta-feedback. Would anyone prefer switching the DALL-E community chat to Slack? I've talked to a few folks about this who would prefer Slack.

Slack gives us the ability to bookmark posts (super useful for saving tips), and best of all, the ability to reply to individual posts as a thread! Threads are so good for responding to people's posts.

Discord is way more ephemeral and hard to ever respond to posts that were made more than a few hours ago. So any topic you bring up might get lost forever.

I think Slack would be more helpful for the DALL-E team to parse what we're talking about too. (By looking at threads on individual posts). It's also easier to make DM groups in Slack if any of us want to collaborate.

Now is the chance! Before the community grows. I know Slack is less cool-sounding, but it's way more useful!

[rugged wing]

I wouldn't mind a switch

[worthy gate]

Switching to Slack?

[lusty creek]

This is on my list of things to get started on tomorrow. We'll build a Slack so those who are more comfortable with that interface can communicate with us there.

[hot goblet]

I don't know if this was already pointed out, but it would be great to have a functionality that allows to download all pictures generated in a session.

and! that creates a folder with the textual input for those images 🥰

[wicked drum]

Hi guys, I've worked a lot with discord within the past years and I do think all the functionalities we are discussing here are possible with good discord configuration and backup of discord bots to make it easier. I can think of something to set up in order to make all of this discord flow better. Just hmu if you want to explore this idea.

[pallid sedge]

@lusty creek and I would love your help with this!

[wicked drum]

I already offered my help to @lusty creek in dms, would be great to schedule a call in order to discuss about this

[boreal kelp]

A battle has started, I’m available as well to improve the discord 🙃

[wicked drum]

Lets join forces then 😉

[boreal kelp]

https://tenor.com/view/yoda-the-force-star-wars-gif-10763720

Count me in!

[south fox]

In case you need a moderator - glad to help here as well!

[worthy gate]

Exciting stuff! I don't know much about discord configurations, but it would be super helpful to use the roles feature so we can easily tell OpenAI team members from general artists, and maybe we can have moderators too like Merzmensch suggested

[half flicker]

"peace pigeon, digital art" not sure what this pigeon is up to 🧐

[rugged wing]

i’m wondering, is the data set that trained the model updated regularly or contemporary so to speak? if you put in a query like iphone for example, it serves up iphone models that used to have a button, mostly similar to the iphone 5 which was launched 10 years ago

and adding “futuristic” to different queries serves up pretty present day designs

[regal solstice]

Curious if we'll ever regain the ability to see what prompt (and other 9 results) that an individual image came from? Right now if I open an old saved link to an individual favorite result, I can't see what that prompt was, or the other 9 images. (we used to be able to! but that was removed in the most recent UX change with the recent tab) 🙏@cinder slate @wanton forum

[glad trout]

Anyone know why the country Pakistan can’t be included in a prompt? Every other country seemingly works fine

[sharp furnace]

ive been getting this as well

[deft silo]

Looks like "Mont Saint Michel" always triggers the content policy filter

[clear sorrel]

Anyone else want to make a separate channel or group of channels for mini games using dall-e?

I’ve thought about a telephone game where one person makes a prompt, uses dall-e to make images, and the next person tries to guess what the prompt was and uses that guess as a new prompt

i think it’ll be a good way to see how intuitive the AI makes the image and it’ll be interesting to see what people think the images are and how replicable they are

[rugged wing]

i'd be down for that 👍

[midnight siren]

So I just got a warning for the following "a negroni with a laptop out of focus in the background" but when I tried it with the word "drink" instead of "negroni" it was fine, and negroni on its own is fine too. strange.

[pallid sedge]

thanks for flagging — will fix! (this is most likely due to the first five letters of negroni)

[pallid sedge]

oh interesting, this feels like a degradation — will ask the team what they think tomorrow!

[boreal ember]

Is DALL-E capable of re-styling an image? Like uploading a selfie and asking it to turn it into a oil or watercolor painting

[pallid sedge]

(sorry I haven’t been following convos in discord for the past couple days — it’s been a bit crazy! so if you had any questions above and I didn’t answer them — it’s because I didn’t read them, not because I think it’s a dumb q etc)

[pallid sedge]

not now, but on our roadmap! This is something we’d like to prioritize higher than other capabilities.

[boreal ember]

Cool 👍

[rugged wing]

this is the output i got for "photo of god". i'm guessing that no christian/orthodox representation came up because there are no statues/paintings depicting god in those specific religions 🤔

[lusty creek]

Yes! Let's connect. I'll move to DM to schedule!

natalie9578 and I would love your help

[rugged wing]

"the mona lisa, by henri matisse" came back with a content policy violation

[pallid sedge]

Weird, will look into it thank you!!

[rugged wing]

i think something like << think this shouldn't have been flagged? report it here >> would be a nice addition to fine tuning content policy violations

the mona lisa, by Jean-Michel Basquiat also came back with a violation 😦

other painters were ok though, picasso, van gogh, bansky

going down this route, im pressing that generate button with fear now 😅

[deft silo]

I have the same issue too, I always hope I don't get auto banned because of this x)

Would be a nice addition indeed

[deft silo]

If generating 10 images is actually 10 times slower than generating a single one (which I'm not even 100% sure of?) maybe it would be nice to enable people to set the generation to less images, for instance I'm pretty sure that even just 3 images would enable me to know if a prompt seems to be worth it or not

Also if it doesn't deteriorate the creatvitity of the AI, and if it also has a proportional impact, maybe enabling generating smaller images would be nice?

So for instance, dividing width and height by two results is 4x less pixels to generate

Combine it with generating only 3 images this would lead to 12x faster generation times if that's how it works, so 2.5 seconds?!

[south fox]

This is great idea.

1. For generation of images: smaller sizes
2. For picking up, and/or publishing/sharing: upscaled image.

Of course, only if original images (1024px) are already created and not just upscaled, it could be other story...

[open badge]

Might be good to add a field to add some feedback to the bias report to make it easier to identify the bias in question.

[signal arrow]

That's true! I reported a pic for 'government', a prompt systematically showing flags for certain countries. But I'm not sure whether they actually understand what I was reporting 😂

[static rapids]

A lot of hindus in the world and they love taking pictures

[iron delta]

Not sure what this result is, pretty off from the prompt
"a potrait of a barn owl, Sigma 24mm f/8 1/1000 sec shutter"
Pretty impressive that it handles misspelling really well

[steel torrent]

I know the mods here must be slammed, but 2 quick questions maybe anyone can answer:

I have an acquaintance on Twitter who is on the waitlist and said that they just got a Google form asking them if they wanted to attend a one hour onboarding session for the “artist access program“, and was wondering if that was legit or some type of phishing scam. Does anyone have any insight into that?

Also, is the “artist access program” really the same as the access that many of us currently have? Or is that some other special program?

[lusty creek]

Hey @steel torrent! That's me! Basically, if folks on the waitlist who indicated that they are artists are open to joining the onboarding call that so many of you joined, we can get them on sooner because we have a higher degree of confidence that they will use DALL-E responsibly.

The actual access is the exact same. It's just the onboarding session required for the earliest levels of access that not everyone on the waitlist will get.

So totally same access. Just trying to get more education out there to those willing to participate. Artists who opt out will still get access with the rest of the folks on the waitlist just as they would have before.

[steel torrent]

Thanks Natalie! I had secret hopes that the "artist access program" was some super secret special high powered art-focused preview of Dalle-3 with a full blown UI and 8K resolution with 8 second 10x render grids 🙂 But I will let my friend know! Thanks! Loving Dalle-2 and the access and all that is going on. So crazy.

[lusty creek]

Haha — nope! Honestly it's just my attempt to make sure artists get the information they need to understand DALL-E (some of their questions are different from what we hear from developers and researchers) and also that we can get them a safety briefing.

[steel torrent]

Artists def don't think like developers and vice versa 🙂

[south fox]

Exactly! And @lusty creek if you are developing some artistic courses, I would like to share my experiences there as well. DALL-E fascinates me not because it can depict "cat with blue hat", but it can re-interpret human cultural heritage and combine semantic concepts in a stunning way.

Prompt design is A&O and we can unleash so many creativity out of DALL-E. Would love to take part on experience sharing with growing AI artist community (I'm creating AI art with different approaches since 2016)

[steel torrent]

I was just commenting on this in a private thread: use cases for DALL-E artists are not the same, really at all, as a developer or passerby. there could be a unique art community opportunity to talk about workflow, art related prompts, and UI and UX (my day job), and a lot more.

The big question on my mind is, after this tech preview, is there a DALL-E website with a unified UI and set of experiences for all, or will DALL-E be an API for some other company (Adobe) to wrap their UI UX around?

I'm having fun making insane images like anyone else (how can I not) but I'm also ready to move on past the party trick images and cultivate what I see as a revolution in art, exactly like @south fox says above. We are on the same wavelength there, and I know a few other DALL-E users on twitter that feel the same way. I think we are talking about the raw power to shape culture, not cats playing guitars on the beach (which i was making today on twitter) 🙂 So we want to continue to have fun, test the limits of what can be done in this tech preview, but many of us know this is just the appetizer. Engaging culture seriously in a novel way...that is the main course <gurrggle_rumble_noises>

[willow meteor]

Sharing this feedback with full understanding that the pictures pull from data sets that naturally contain bias: I notice that "photograph of a boss at work" returns only light skinned people, could be good to show more diverse representation in authority figure searches. I generated a couple times and similar results. (I know not to share these publically because of faces anyway)

[lusty creek]

Thanks for this, Max. We'll pass this along to the team as another example of this bias. We are seeing it too. (Thanks also for your diligence about the guidelines/social media sharing!) 🙏

[oak vessel]

Something I really would love to hear a discussion on is a way to embed a "qr code" into the very fabric of each generated image. Obviously I am not well versed on programing and computer language and don't know how to actually create this system. The rational behind this idea is a means to ensure art made by an AI is identifiable by others as a means to prevent deception and scams. Perhaps having a pattern that is not detectable by humans but when fed through a tool it would be flagged, could be a method for this to work.

[south fox]

Interestingly I have a good example for diversity. A simple prompt "A Judge" creates series of people of different genders and ethnics (I just want to share it with you here, of course, I won't share it outside because of semi-photorealistic faces). Fortunately, not only "white old men"

[radiant dirge]

I tried to add a rosette and a kettle to this beautiful 70s style carpincho but I've got a warning 😢

I'm afraid of writing something that isn't allowed and I get banned. Other day I tried to add a snake to an illustration with inpainting and got a warning too. It only happens when I use inpainting.

[radiant dirge]

Would it be a bad idea if we could make a chat to list the forbidden words and avoid making mistakes?

[young lagoon]

When you inpaint something I think it technically re-uploads it and scans if it could be offensive content. If Dall-E generated something like somebody holding a sword, if you try and inpaint on that you can get a warning for example. Despite it being generated by Dall-E. Because I think its seen as a potentially violent uploaded image. So might not have necessarily been the word that caused the warning. At least thats what I've noticed because I've been hit for a warning of Darth Vader where he appeared to be holding a sword in a banksy and another of a statue which I'm guessing is because his muscles were defined but that wasn't part of the prompt or the intention it was just something Dall-E generated when I was creating statues.

[worthy gate]

I had the same fear at first, but Joanne reassured us that people are only banned for repeated attempts to get dall-e to generate extreme & obviously-inappropriate content. Here's a link to her message for more details:
#community-help message

[static rapids]

I used the word snake without issue

[clear sorrel]

I realize there has to be a private API behind this, but I’m wondering how public this would be. This is orders of magnitudes more powerful than GPT-3 in the hands of almost any person, so I wouldn’t be surprised if there’s no public API ever. It would just be cool to peer into that code

[deft silo]

So swords are an issue? maybe that explains why my "Mont Saint Michel" prompt fails, Saint Michel being often represented with a sword

Yet the Mont Saint Michel itself is just a city on a small mountain island 🤣

[young lagoon]

From my experience anyway. It can generate something with a sword but if you inpaint or upload an image that has a sword it seems to give you a warning. I think I had another issue with this. But I definitely had one and the only thing that I could tell that was offensive was the fact vader had something that resembled a sword

[deft silo]

I see, hmmm.... did you still find a way to inpaint it?

If you didn't I think maybe just re-drawing on the sword would work? Before uploading

[young lagoon]

Generally if you get an odd warning that doesn't relate to text its probably from inpainting because of this.

I didn't try, I didn't wanna push the warning system

[deft silo]

I see

[deft silo]

If this works (12x faster generation times) then it could maybe be combined with "generation matrices", for instance you could input something like:

A {dog,cat,mouse} plays {piano,guitar}, {oil painting,digital art}, high quality

And it would input the possible combinations
So using this technique, it means such prompt could still output 2 or 3 images for each of the combinations in the same time as it currently does.
Or it could also maybe just generate one per combination at the usual size and then you could just hit generate again if you want more results, but that would be a nicer or at least different experience than just generating 10 versions of the same image, then 10 versions of another etc...

[clear sorrel]

Depending on Dall-E’s limit, making generational matrices should be possible even now

You can just make some POST requests. I’ve tried it before

Though their closed API might have a couple seconds to limit the amount of requests one user may do

[deft silo]

Yes would be possible today indeed, but then you'd have to either tradeoff on generation time making people wait longer or just generate one image per combination (if the number of possible combination are around 10)

[clear sorrel]

I’m talking about parallel requests, not one after the other

[deft silo]

Oh I see! Yeah true you can do up to two parallel requests indeed

And at three you get a funny error message about how OpenAI loves your enthusiasm 🤣

[deft silo]

Looks like Pepe the frog also triggers content filter 🐸

Maybe because it was at some point associated with racism? It never was never a racist symbol in the first place and still hasn't become one tho 🤷‍♂️

Maybe I was also unlucky and some other part of my prompt triggered it but I don't really dare trying it again

[rugged wing]

Since the generated images are always different for the same text, it would be great to have a longer history, e.g. where you can select the day or month first.

[open badge]

How about a feature that allows you to generate fewer but higher resolution images? don't know if it is possible, maybe if the resulution is a multiple of the original? I feel like some things aren't really possible to create due to the limitation of image size, like making big scenes forces you to lose detail to the point of not being readable.

[sacred valley]

Have you tried using inpainting to stitch a bunch of images together? Can get some really massive detailed results that way

[open badge]

How can you stitch an image?

[sacred valley]

Let’s say this is the image you’re starting with

You can then bring it into a new canvas via photoshop etc and slide the image halfway to the left

Or do a double sized canvas like this

And then crop within Dall-E like this

Erase the white area + a little part of the middle area with the inpainting feature, and then fill it in with whatever prompt you’d like to use

I typically use the same prompt or a slight variation of the prompt that the original image used

So this is the result of inpainting that other half of the image

And then you’d just need to line up the left half of the new image with the right half of the old image in photoshop

@open badge Lmk if that makes sense, there’s probably a less convoluted way to describe it haha

[open badge]

Big brain move 🤯🤯🤯

[sacred valley]

I think someone also posted a video tutorial somewhere in this discord that could be more efficient

Here’s a stitch I made of 9 different images

[static rapids]

Ive been curious for numbers behind dalle ? How many images have been made so far and it what time span, how many prompts , how many users all that ,

[iron delta]

I was looking at the EXIF data for one of the images generated and it has a counter of sorts. The last image I generated was around 4 million

So right now should be about 4.5 million. Educated guess

[plucky bison]

Oh hmm.. I managed to hit a "bad prompt" on my third attempt. I originally asked for A yellow umbrella in the style of Jean Michel Basquiat but found that the backgrounds were a little boring. Changing the prompt to add A yellow umbrella in the style of Jean Michel Basquiat painted on wood. resulted in the following error message:

I can't really think of what could have caused that?

[plucky bison]

Ran another with A collaboration between KAWS and Jean-Michel Basquiat and got hit with the same error. Is there an issue with asking for stuff in another artist's style/with basquiat specifically? 🤔

I suppose some of Basquiat's real works can be a bit rough but the style as a whole goes far beyond that

[errant rock]

It would be interesting if certain artists were moderated. I kinda doubt it but I hope it’s true.

[naive moth]

Ran into an error when running searches in Chrome in iOS. Phone went into sleep mode mid-search. When I started back up I closed the old window as progress bar was stuck, and now getting an error “You can only run 2 generations at once”. I tried closing Chrome windows and rebooting phone, but still not working. Any ideas?

[south vessel]

I was about to ask the same question, put on sleep mode my laptop, when to pick something at the door and when I came back (and woke the laptop) the progress bar was stuck loading and now I cant put any prompt and there the same error "You can only run 2 generations at once". I closed Google, deleted the cookies and nothing 😦

[naive moth]

Did the same. Tried logging out of the site, but no luck

[real furnace]

Same here, tried logging out too. Nothing

[pallid sedge]

#bugs-and-issues message

[real furnace]

That's amazing. What was the prompt?

[sinful quarry]

Hi, looks like I'm still unable to access DALL-E after indicating that I aim to be a content creator with it. Is there anything I'm doing wrong there?

[lone vessel]

Hey! Is there an API for DALL-E that is publicly accessible, or would it be kosher to automate the creation of some images with a bot (while still respecting the normal query limits)?

I feel like the answer is probably no… just asking!

[clear sorrel]

I asked this earlier, I guess there is a very limited psuedo-api

you can make POST requests and GET requests

but no more than 3 simultaneously

hmmmm

  "error": {
    "message": "You didn't provide an API key. You need to provide your API key in an Authorization header using Bearer auth (i.e. Authorization: Bearer YOUR_KEY), or as the password field (with blank username) if you're accesing the API from your browser and are prompted for a username and password. You can obtain an API key from https://beta.openai.com. Feel free to email [e-mail] if you have any questions.",
    "type": "invalid_request_error",
    "param": null,
    "code": null
  }
}```

it might use regular OpenAI api keys

i tried using my API key as the password

it did not work

[lone vessel]

Hmmmmm it’s not in the rules I was sent upon joining DALL-E, just wanna know how the team feels about this since it would essentially have to be a web scraper

[clear sorrel]

I assume if you don't do multiple at the same time and make proper attributions, you should be fine

also, i think this is interesting

https://openailabsprodscus.blob.core.windows.net/private/

You should be able to see all of your private images

would be cool to access that

  "object": "task",
  "id": "task-RfkEWs5QwAKZch5sqs9x7Bt0",
  "created": 1653617097,
  "task_type": "text2im",
  "status": "pending",
  "status_information": {},
  "prompt_id": "prompt-LCPPPTfLEuIWIMPhsh5fHapx",
  "prompt": {
    "id": "prompt-LCPPPTfLEuIWIMPhsh5fHapx",
    "object": "prompt",
    "created": 1653617097,
    "prompt_type": "CaptionPrompt",
    "prompt": {
      "caption": "A torus shaped planet with a moon in the center, digital art"
    },
    "parent_generation_id": null
  }
}

This is basically what we post to the Dall-E API

interesting

the API Key works fine for GPT-3

I think Dall-E has a different API key

makes sense

but they 100% have a pretty organized api

[pallid sedge]

Hi! We don’t allow programmatic access at this time 😿

[lone vessel]

Thanks so much for answering, Joanne! I figured that would be the case, but I just wanted to make sure so I wouldn’t risk losing access to this amazing project! Thanks for clearing this up 🙂

[pallid sedge]

ok I just fixed this so you should be able to run these in 12ish hours if not earlier

[pallid sedge]

https://openai.com/blog/dall-e-2-update/

this post answers some of your questions!

[pallid sedge]

just fixed this so you should be able to run these in 12 hours if not sooner!

(we block certain politician names, and overtriggered on “Michel” bc the president of the european council is Charles Michel) cc @plucky bison

[deft silo]

Thanks a lot for the fix! 🙂

[plucky bison]

Thanks for the explanation !

[split orchid]

Good day! Any country related problems solved yet?

[eager stone]

So Dall-E already detects other languages?! That's amazing

[split orchid]

I mean, Dall-e is not available in some countries

[plucky bison]

Is there a branding kit somewhere we can use to readd the logo in areas where we might have to erase it to put our own work?
@pallid sedge

[lucid ravine]

Hey! Would these be okay to share? This is from a novel prompt that I use in other text to image tools that seems to work exceptionally well in Dall-E.

[neat bridge]

Thanks for checking -- I think the first one is fine but the second one might be too photorealistic (it's admittedly quite arbitrary as we calibrate)

[lucid ravine]

totally fair. the interesting thing is that Dall E knew this artist only did portraits/art shots of faces

so even though I didn't even include any words about people or faces

just the artist's name in the prompt sent it to that. and it does that in every text to image tool I've used.

[plucky bison]

Another q regarding content: I'm getting interesting results for drawings of eyeballs using the term 'LSD'. Technically no drugs are being shown but it makes the veins in the eyes a bit more apparent and gives a more 'wobbly' style to drawings. Is that OK?

[eager stone]

Well it somehow understands italian pretty well, not sure why but it does

[deft silo]

Is it ok if your post title doesn't specify "AI generated" but your first comment under the post does? On some websites such as Reddit, if you want to make a fun/meme post, it makes it look almost like an ad to speicify it from the title

On the same subject I have been wondering if there will be any evolution of the rules for other cases, say you want to create a sticker pack with DALL-E creations, it seems pretty hard to correctly respect the rules for now :/
Makes me realize it would be nice if there was some "AI generated" checkbox you could tick on these sites so that they add a small badge or something that make people know about the origin of the image without you needing to modify your title.

[clear sorrel]

Can we make photorealistic images of ourselves? Like editing an image of me to have a mohawk?

Make and share

[glacial canyon]

nop you can t. Even if you add yourself with photoshop on a background made by dalle, even if dalle hasn't generated your face you can't share it

[junior crypt]

Got this twice today. Is anyone else getting this?

[latent marten]

yes i get that too on mobile, on pc it works however

[clear sorrel]

I got it, but after I waited around a minute it worked for me

[neat bridge]

yes that's fine!

[deft silo]

Thanks

[neat bridge]

if it's your face that's ok!

[neat bridge]

I think we're ok with this now (we're loosening restrictions as we learn more) 😺 (thank you for remembering though <3)

[glacial canyon]

Ohhhhhh since it was a question I asked during the Google meet of course I remember 😂😂

[lucid ravine]

accidentally hitting the word filter. was trying to use the name "Ted Gore", a photographer.

Not purposefully trying to push any boundaries. I guess we need to be very careful and not just use old prompts from other tools?

[gentle ether]

yeah i was thinking about that when I was trying some of my old MJ stuff

gotta double check some of those prompts lol

[lucid ravine]

So Mj prompts should be okay-ish.

[somber quartz]

Who owns the copyright to images that DALL•E creates for me?

[plucky bison]

OpenAI but it's an evolving situation according to their introductory call the other day

[somber quartz]

is there a recording of this call, or notes anywhere?

[plucky bison]

I'm not sure... but @pallid sedge will know

[somber quartz]

I'm curious about what the developments around copyright hinges on—yet to be passed legislation, legal work by OpenAI's team, or is it a business model decision?

[plucky bison]

My understanding is that they have no desire to own the copyright to artists' work, it just so happens to be the simplest legal framework for now. It was a very late night session for me so details are blurry lol.

[somber quartz]

I see. Thanks for replying!

@neat bridge what is the most useful kind of feedback the DALL•E team is looking for right now? and what's the best way to deliver it?

[junior crypt]

That's a great question, was wondering that myself 😁👌

[deft silo]

I wonder if there is any difference between how images are generated now and what would happen if you generated some digital art of something and then did some style transfer to a photographic style

Maybe this would enable getting high quality photorealistic results of surreal scenes more easily than today?

Right now it seems harder to diverge from reality when using a photographic style vs digital art style

[naive moth]

Is there a proper way of using punctuation to flesh out longer concepts? Or is it best to use large run-on sentences? I’m wondering because in many examples the comma seems to be viewed as the end of the description, only followed by a stylistic choice

[plucky bison]

I kinda just give it a lot of keywords

[naive moth]

But when going for more specifics, I’m finding it difficult sometimes since portions of the prompt seem to get ignored or misconstrued.

One that gave me a possible content policy violation warning today was the following, but in hoping to avoid this I can’t figure out what portion could be causing this:
“A skateboarder skating through a large puddle of molten glass as he leaves tracks in the glass and has sparks shooting out from behind the skateboard’s wheels”

[neat bridge]

“shooting” probs 😿 I think it makes sense for the system to tell users what caused the block in the near future whenever it can fwiw — considering this!

[naive moth]

That makes sense, and hadn’t occurred to me. I was focusing on the the liquid aspect or even the term “tracks” as possible issues. I’ll try it without “shooting” and see what happens

[naive moth]

That was it. Changed “shooting” to “flying” and got this: https://labs.openai.com/s/6eFnTDAeteeABShHR3mxin1A

[neat bridge]

honestly, any and all feedback would be helpful at this point since it’s so early — but top of mind for me as a product manager are use cases & misuse potential.

We learn about a new use case every single day; e.g. you told me yesterday re: your language learning use case and that’s something I hadn’t thought about! This then lets us consider these use cases as we prioritize research problems / product-ization directions.

[neat bridge]

oh another thing we love learning about is how DALL-E fits into artist workflows & how it can expedite creative iterations / processes!

from stitching images to alpha channel needs to how @hybrid drift explained how to get around the close-cropping problem — we’ve been in awe of how creative this community is 😻 Natalie knows a lot more about this since she works super closely w the artist community (not tagging her since she needs to detach heh)

[indigo rune]

Is there a way to see all the links you shared?

[neat bridge]

I think I’ve seen a hacky way to do this somewhere, but not officially. We are exploring a save feature fwiw?

[somber quartz]

@neat bridge do you have any answers on the copyright questions at this point?

[neat bridge]

@somber quartz, @plucky bison is right and it’s “all of the above” — currently OAI owns DALL-E generations because it’s the simplest legal framework (we’re in fairly uncharted territory right now); our legal team is looking to figure this out though!

(will be less available on discord to answer questions the next few days!)

[leaden pollen]

@neat bridge Is this you guys? https://www.instagram.com/openaidalle_2/

[plucky bison]

no the official acc is @openaidalle

[twin river]

@neat bridge is it okay to download pics of realistic-ish people as long as we don't post them? I'd like to have some for reference when drawing. The way light hits faces, body posing structures, expressions, etc can all help with my own creations.

[lucid ravine]

question - in other text to image tools I have done brief experiments to see if there are interesting, "random" results that can be generated from non text prompts. examples - entering in large numbers, negative numbers, punctuation and symbols, etc. are there any rules/restrictions that prohibit exploring DALL-E in this manner?

[neat bridge]

nope, go ahead and let us know what you find 😺

[plucky bison]

fwiw emojis get parsed quite well, 🍋 got me lemon tea lol

[knotty furnace]

The prompt was "ADN inmortality structure", could this AI be used to achieve certain technical advances in any field (in the future)?

[plucky bison]

This one? no. Others? yes!
This specific article is grim but the opposite is doable too
https://www.theverge.com/2022/3/17/22983197/ai-new-possible-chemical-weapons-generative-models-vx

[somber quartz]

AlphaFold is working on that:

https://alphafold.ebi.ac.uk/

[plain plume]

Over the last days I`ve got a few "Policy Violations". As it "may lead to an automatic suspension of [the] account". If that should happen, is there a chance to get a manual review?
Because I feel it was deserved only in one case (where I accidently asked about "Weapon looking like a Balloon Animal" ... other times it was something like "Dog with table tenis bat on his head" or "pink character with fur and small arms"..

[plucky bison]

It seems to me it's not actually automatic; see:

[open badge]

https://tenor.com/view/saywhat-huh-confused-iamconfused-youlyin-gif-22766635

[deft silo]

Inside an empty all-glass cubic room, frameless, high glass pane ceiling, glass pane walls, (10 by 10 by 10 meters) with garden outside, digital art, creative, modern
I'm trying to replicate a room that would look a bit as if you was inside an Apple Store glass cube, unfortunately there are always pretty big "window frames"/beams and I don't really have any idea of how to remove them especially after specifying "frameless"

Does anyone have an idea of what else I might try?

[plucky bison]

...adding 'apple store'? 🤔

[open badge]

I think you have to add frameless in the context of the glass. the "," probably makes the image just generate without a frame added to the image itself

[deft silo]

Thanks @plucky bison and @open badge will try both suggestions

[open badge]

hmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmm

didn't work for me

[deft silo]

It does help to some extent, the frames are still present but they look a bit thinner in some cases

[open badge]

i made some frameless 🤷‍♂️

[deft silo]

I tried "Inside an empty all-glass frameless beam-less cubic room, high glass pane ceiling, glass pane walls, (10 by 10 by 10 meters) with garden outside, digital art, creative, modern"

Still has frames but also better structured maybe so that's still an improvement

[open badge]

It's a hard prompt to nail down it seems

[open badge]

adding cubic room makes it adds cubes to the floor of the room for me xD

[deft silo]

Yeah ahah, I also had it a few times, same with "on the moon", typically adds another moon in the sky even if astronauts are on the moon

[deft silo]

Not bad indeed

[open badge]

I have an idea

brb

[plucky bison]

Maybe frameless windows just looks like no windows?

So what if you ask for no walls

[open badge]

less go

Inside an empty room with massive glasspanes as walls, (10 by 10 by 10 meters) with garden outside, digital art, creative, modern

it's just an attempt, but maybe making the scale so big that you cannot see the beams make beams not exist 🤔

[deft silo]

Interresting, closer in some aspects indeed

Thanks for testing!

[open badge]

no problemo

and look, i solved it

not a window in sight xD

[deft silo]

Almost yeah, the issue is that this is a case where it's hard to specify precisely what you want, for instance if you look at Apple Stores closely there are still some places where you can see some sort of very small "frame" in the corners and I think it makes it look better / more realistic if you are inside it but it's hard to describe

[deft silo]

Gonna live underground from now on 🤣

[open badge]

Apple store from the underworld 🤔

[deft silo]

Some good prompt maybe

[neat bridge]

to clarify: most suspensions are automatic (since we don’t have time to monitor filter stats 24/7), which we then review the next workday to determine if it was actually justified; if it wasn’t, we reinstate access right away!

[plucky bison]

Nice! That seems pretty fair 😄

[plucky bison]

If an artist friend just signed up to the waitlist, is there a way to get them nudged up the list a bit?

(bonus point: she's quite enthusiastic)

[neat bridge]

let me ask natalie on Tuesday! (Not tagging her bc weekend)

[plucky bison]

Thanks !

[neat bridge]

is it just me or do you all get the welcome screen every time you open the app too?

[plucky bison]

Happens once in a while to me

[wicked drum]

Should be only when you join the server but might appear if it is changed frequently

[south vessel]

Sorry to bother but I've got a question, how different does a derivative work needs to be in order to be sold or commercialized in some way?

For example I did this

Based on a Dall-e image I made

As you can see, it's kinda similar haha but I was just curious

(I'm not trying to sell it hehe, pure curiosity)

[somber quartz]

#community-help message

[south vessel]

Oh I know about the Dall-e outputs but my question goes more to derivative works, because on the Meet call they mentioned that in order to be sold it needs to be a derivative work and we need to inform them about it 😅 or at least that's what I remember hearing, please correct me If I'm wrong

[south vessel]

Thanks btw :)

[somber quartz]

My understanding of copyright is that you would need explicit permission in writing from the copyright holder to create a derivative work. Most likely scenario right now is that we just have to wait until they have figured out a copyright policy (probably a lot of legal work that needs to be done)

[south vessel]

Oh right! Thank you so much bro and sorry to bother hehe

[somber quartz]

no worries, I'm just trying to be a helpful dalle community member!

[lucid ravine]

thanks! here is a silly one. I can post the prompt via DM if you don't want folks "playing along". the prompt was math based.

[coarse prawn]

Hi, I am using DALLE to recreate lost old paintings. I have seen that in some case the results are cropped. Does anyone have had the same experience?

[coarse prawn]

Example of cropped image for the prompt "a highly realistic oil painting of a full figure saint francis in ectasy on his knees wearing a brown habit. Painted by Rembrandt, artstation"

[open badge]

You can uncrop it though

Use the same method as when making a compisite. Take the image into an editing tool. Then move it down and tell DALL-E to generate the top part of the image. Then add them together and then you have the full image

The channel composite is dedicated to that

#community-help message

Go There for a better example than mine

[coarse prawn]

Thanks for your message. I have tried and it works very well. However, I think it is interesting to give feedback to OpenAI, as this cropping seems to happen often when asking for representations of religious paintings, specially when asking for full body portraits

[open badge]

I think that is what the suggestion channel is for :)

[deft silo]

Let's say someone (not me) loves very much an AI generated image, how could they put it it in profile picture? I feel like this is a bit tricky since OpenAI owns the the image and since it's not me I think it's not exactly equivalent to me sharing it on some social media (which is explicitly allowed by the rules)

[plucky bison]

Well, what could you do if you posted it on social media and they grabbed it to use as an avatar? I don't think this is different

[deft silo]

True, but technically this still looks like copyright infringement (but I'm not a lawyer 🙃)

So if they want to do things right, I don't know if they really have a way

[neat stirrup]

I have a question regarding the content policy warnings (and since I just read that suspensions are actually automatic of course I would like to avoid them). My prompt "a reproduction of the mona lisa which is true to the original in every detail" triggered the warning and I do not quite understand which part of it is the offending one here. I tried a few variations which did not trigger any warning, but I don't want to try my luck by raising too many alarms.

[plucky bison]

Reproduction probably.

[neat stirrup]

Oh - now that I see it spelled out like that I get it 🤭

That was it - looks like i can safely use "repro" or "replica" instead.

[oak vessel]

Are we allowed to generate art with depictions of "PG" conflicts? For instance two samurai soldiers facing each other while holding swords, or two tanks firing across each other, a city exploding etc

[odd crater]

Would you flag this picture for "doesn’t match my text description"?
The description was "a basketball by nike". I mean, it's still a kind of basketball and you could imagine that the white thing is the "swoosh". But having an animal head coming out of the ball is definitely a strange outcome for the description of a simple basketball.

[somber quartz]

I think you could probably just make a tweet when you update your profile pic attributing the image to OpenAI. e.g. the image in my profile avatar was generated by OpenAI’s #dalle

I think that beta testers are just expected to attribute all images to OpenAI when they are shared, but other people who see the image aren’t under that obligation.

[deft silo]

Thanks! It is the subject of other people I find more complicated: while I suppose they can't be under the obligation of attributing to AI generation if they did not agree to in the first place, what about the copyright on the generated images?

If an image is copyrighted then I would suppose you can't just use it as your profile picture

I think this is a bit of an edge case and it doesn't matter too much (I was just wondering it because somebody asked me about using it in their profile pic) but still interesting to discuss nonetheless

[neat bridge]

I’m not sure I fully understand the question but if it’s an image you like and you want to set it as your profile pic, I don’t see why not? I did it 😅

[deft silo]

The TL;DR is: I created an image and someone said they would love to use it as a profile pic, but from what I understand OpenAI owns the copyright of the generated images, so I'm unsure of wether they have the right to use it

[deft silo]

But yeah I agree for us DALL-E users (or OpenAI employees 🤣 ) it seems it is fine

[neat bridge]

oh I see — as long as they keep the DALL-E signature and don’t commercialize we’re happy that more people get to enjoy it!

[deft silo]

Thanks for the clarification! 🙂

[indigo rune]

frame to use if you like how it looks

[plucky bison]

I was thinking it's a bit loud, but maybe if you just kept tbe original five squares in the bottom right corner but rounded?

[indigo rune]

oh i see what you mean

[plucky bison]

Oh and perhaps a lefty version for discord, as the icon would overlap otherwise

[indigo rune]

rounded

yeah you can rotate it to be in any corner

like that

since discord supports animated gifs if you have nitro you can make an animation where it travels around the circle 😆

also if you do inpainting you can get other frames for your gif like make your avatar blink

[young lagoon]

How strict is the violence and shocking images policy? Like take the following two images I think they're pretty cool because it kinda shows half of vader where he's injured but idk if it would be considered shocking and then the second I've been messing around with comic book stuff and technically its violence but nothing too obscene just like standard star wars stuff

[deft silo]

Thanks!!

[spring wyvern]

how militant are you about 'the signature' if we add a proximate disclaimer as an alternative?

i'm mainly talking about displaying images in non-square contexts (e.g, Instagram Stories offer a 3x1 grid layout which makes all the photos landscape, or a Wordpress template where the hero image is non-square) - places where it's easy to add a text caption saying 'AI image by OpenAI's DALL-E' but not so easy to 'keep the sig'

[plain plume]

Had a few friends who wanted recreatetions / Variations of their work by Dall-E (with consent). They realy where thrilled but also a bit afraid. Is there a way to tell Dall-E (for example via upload) that it shouldnt be able to do Variations from certain Artworks?

[neat stirrup]

Q: Does capitalization in the prompt matter or does dall-e convert all prompts to lower case before tokenization like other models?

[crude birch]

After getting early access, I've gotten a lot of requests for prompts from strangers - How do I respond to this? Can I DM random people their request? Couldn't see or remember the rules about this sort of requests.

[plucky bison]

My understanding is that you're free to do it, but you bear all responsibility

[naive moth]

I had asked this before, but wondering if an admin can chime in? Is there a proper way of using punctuation to flesh out longer concepts? I've found trying to address multiple components in a scene, the AI tends to get things mixed up, but I'm not sure if using full sentences and punctuation works best, or bullet points or some other method that the AI "understands" better?

[patent sequoia]

Is making art of people kissing ok?

[naive moth]

As a fellow beta user, I'm going to make an educated guess based on the briefing we got and the agreement/list of rules we signed. I would guess it's ok as long as it's a G-rated kiss, not explicit or overly intimate, no tongues. Also not showing any faces if it's photorealistic, so maybe best in silhouette or stylized as a painting or cartoon/animation? Again, just giving you my thoughts based on how I self-moderate my own content searches

[spring wyvern]

Aha I have a hit a 'daily request limit!'

[jade dagger]

I just hit that daily request limit, too! Anyone know what the limit is? Guess I have to be a little more selective with my experiments!

[clear stone]

I hit the limit too, and I know I haven't used it that much today. I wonder if there are server troubles?

[fickle hawk]

same 😭 hoping this is a bug, i haven’t used it that much today i don’t think

[hybrid drift]

a group of artists typing prompts, sad that they have hit their daily limit on dall-e, digital art

[jade dagger]

Message now says limit is 50 in 24 hrs.

[alpine falcon]

yup! I got it too now for the first time

[south fox]

Me too now! I guess I can go sleep now 😂

[neat bridge]

capitalization does not matter!

(Posted an update in #announcements to answer the other questions re: usage limit; I’m truly sorry I didn’t proactively communicate this beforehand!)

[neat stirrup]

Since you are very concerned about ethical questions related to the use of dall-e one aspect that i am curious about is the average CO2 footprint of a single prompt. Given the time it takes to render the 10 images i can imagine it is non-negligible. And whilst of course I am also not super happy about the new usage limit, I think if you were displaying the actual estimated energy use of a prompt it might make people prompt a little bit more responsibly. (Says the hypocrite who has been prompting like there is no tomorrow)

[brisk thistle]

A friend saw a DALL-E image I posted and wants to know if he can use it as a base to make a vector art for his music album cover?

[somber quartz]

is there a special reason dalle generates square images? or was it just an arbitrary choice?

[high pulsar]

Can we use these Dalle generated images in book covers?

[proud creek]

I know we are not supposed to share human faces, but i saw some people putting some squares of colors on faces.
could this kind of " censorship" be allowed by the team if we share it like that ?

[lucid ravine]

is that.. an 80s boombox camera? 😄

[proud creek]

prompt : A still from Total recall of a young woman hodling a medium format camera"

[lucid ravine]

nice result. cheers for sharing.

[molten socket]

To clarify, is the 50 total for 24-hours A) 5 generation entries / 50 resulting images... or B) 50 generation entries (i.e. 500 resulting images)?

[paper cedar]

Definitely the latter, I've already blown past 5 entries today.

[molten socket]

Ahh... it seems you have a "50 total in 24-hours" counter... so if I've generated 50 over a 6 hour period, the additional generations will "unlock" over 6 hours.

[pale shell]

Use of DALL-E for commercial purposes is not allowed at the moment (I'm assuming they plan to sell this album)

[pale shell]

Use of DALL-E for commercial purposes is not allowed at the moment. If it's a non-commercial endeavor then it should be okay – e.g. https://archive.org/details/1111101000-robots/mode/2up

[clear vector]

can I refer a friend?

[patent sequoia]

i want to refer a friend too

[charred imp]

What could be wrong with my prompt? I got 3 warnings for that, trying different phrasing "colorful birds taking flight from tree branches by Martin Johnson Heade, oil and silver leaf on wood, extremely detailed"

[errant rock]

Can we get a clearer idea of what kind of moderation is being implemented around photo upload. I uploaded a picture of a femme presenting body in shorts and a tank top and got moderated. Lol. Example photos? A photo of masculine body in shorts was not moderated.

[red bay]

that's a good question

[sullen silo]

I also incurred in similar issues with prompts that doesn't violate the content policy but may seem to be triggering the moderation system somehow.
"A risograph poster with a black background, 135mm portrait of a man dressed with a soviet-era jacket, he is backlit by the moonlight"

Initially I thought that it could be because of the soviet theme (somewhat 'political' although USSR doesn't exist anymore for a long long time?), but after some prompt tinkering it seems that what was causing the issue was instead the "135mm" lens word. Got me thinking... is it a safe to avoid realistic face generations? (In my case I was not planning to use the face, I was going to edit/replace it with a flying saucer later anyway... but if I add it to the prompt the flying saucer appears on the background)
Or did the AI associated the '135mm' with the gauge of some kind of gun or something because there was 'soviet' in the prompt...?

I'm kind lost now and a little afraid to try anything remotely near anything historically oriented.

(Just for ref, the image I was trying to re-build as a risograph on DALLE is this one, a poster I made a long time ago as a parody of sci-fi/action 80s movie covers):

[neat bridge]

just fixed this -- this should work in the next 12ish hours!

[stone bay]

I got a violation warning for “U.F.O. abducting a hamster” 😓 should I tell someone about that?

[neat bridge]

I just checked, and it looks like it's the word "soviet" -- I agree that the prompt itself shouldn't be moderated, but I just looked at the data and we've gotten far more political type content for that prompt that I don't feel comfortable asking our product policy team to remove it (for now). Sorry for the inconvenience! We'll continue monitoring this and loosen it if we see fewer of those attempts

[neat bridge]

This is helpful, thank you! I'll add this to a list for our team to consider loosening the violence classifier!

[trim wharf]

how much of an image needs to be altered in order to avoid copyright issues? 10% ? 20%? what if I just flipped the image horizontally or converted it to grayscale? could I then use it in print media?

[sullen silo]

That's understandable. I imagine that some people are probably really 'testing' the security measures trying to find breaches right now. 😑

One thing that I thought that, maybe, can help with situations like the ones we are dealing (where you have one word that has ambiguous context, like 'soviet', 'abduction' or, I don't know, the 'gun' in 'water gun') is to proceed in generating the content (server-side only, not showing the results to the user), and use some inverse-image classification on the result to see if DALLE 'variations' identify something really out of the rules (explicit violence, firearms, politicians etc). That way, maybe, with time, the results can help the model learn in which contexts the word may/may not be violent/against the rules and not just rely on a hard limit on any usage of the word by itself.

But for now I understand that, to stay on the safe side, some situations like these will occur, c'est la vie.

[plain plume]

Isnt there something like a General law? In germany it says something like "if the original isnt directly Recognizable // if your work is a work for its own". Just mirroring or changing colors wouldnt work (at least for german laws)..

[patent sequoia]

can I pay someone for unlimited generations, like, rn?

[coarse prawn]

Out of scientific curiosity, is it known how big is the dataset that has been used for training DALLE-2?

[spring wyvern]

Yes:

When training the encoder, we sample from the CLIP [39] and DALL-E [40] datasets (approximately 650M images in total) with equal probability. When training the decoder, upsamplers, and prior, we use only the DALL-E dataset [40] (approximately 250M images).

From the DALL-E research paper: https://arxiv.org/pdf/2204.06125.pdf

[tight warren]

where can someone find the link to this discord community? Someone tweeted me, they're saying they already have access. Is the only way to get invited here through openai support?

[lucid ravine]

correct, invite generation by non-admin staff is disabled. so you're saying there is a person with access to DALLE2 who wants access to the discord?

I'm sure joanne could help out.

[sacred valley]

Curious to know how many people have DALL-E 2 access at this point 👀

[sly fog]

Can I share in social media like this?

All the results...

Would that be ok?

Is that a problem? Thanks!

[brisk thistle]

I was experimenting a bit today to test DALL-E for simple logic, which it often does not display. For example asking for 'a candle in the shape of a number that equals three plus four', 7 is never generated. Asking for 'spheres in all colours except red', shows mainly red spheres and some other colours. a 'pedestrian without a backpack' generates guys with backpacks. Any expressions that can help or is logical inversion and exclusion simply out of it's capabilities?

[lucid ravine]

Just curious - have you worked with other text to image generative tools?

I ask because I'm wondering if any other tool can do that yet. Its my understanding that these tools are examining images and doing their best to match what they see in images. subtraction/reduction of features isn't their strong suit. in fact, most other tools have some type of negative prompting to try to influence and steer it to not do those things.

[tight warren]

Yes he says he has access already to DALL-E 2

[sullen silo]

In the access email he should have received a link to the discord server from Natalie (at least I did) ... maybe he didn't notice?

[indigo rune]

I just tell people to ask the person/contact who gave them access to invite them to the discord

[sacred mirage]

Just started off on DALL-E. Is there a way to programmatically give the prompt and see the response? Like we do with GPT3? Vs. using the website?

[indigo rune]

you could look at what api the website is using and call those endpoints with your auth token. not sure if they want us doing that though

[spring wyvern]

there isn't ATM.

[livid stream]

Forgive me if this has already been asked/answered, but when does the new 24-hour reset for 50 new generations, or is it different for everyone?

[sullen silo]

its rolling time, based on the time of your generations. e.g. say you generate 20 images 10am and then 30 images 5pm, you will have met your 50 images/day limit.
Next day 10am you will be able to generate 20 more and then 5pm 30 more, etc...

[livid stream]

Not confusing at all... lol

That counter can't come soon enough...

[subtle knot]

Is there an online database of the image sources/websites the model is trained on?

[spring wyvern]

No, it's proprietary

[lilac ledge]

Hey @neat bridge, I've put together a small narrative animation using generated images and am very eager to share it. However, it's pretty strange but (I think) not objectionable. I'm wondering if it is possible to run it by OpenAI staff first?

[sweet garnet]

Hello there, I've been uploading images from my work ( and another friend's work ) and I received the message that I cannot upload these images, the realistic one I understand, but the colorful? what happens here? Just to understand ( thanks )

[dense mortar]

I think the AI thinks it's human flesh

[neat bridge]

hi! our classifier detected smoking as a "not safe for work" attribute in the left image (although I think it's ok), and thinks that the right image is... phallic 😓

[sweet garnet]

OK! I understand! Thanks! for the tip. Also: one question more: why when you do some cartoon characters in real movies some of them are anthromorphics ( 100% human beings ) and others are just charachter adaptations? I still don't get this parameters after different tries.

[autumn crow]

Are there any plans to add the text diffs shown here to the consumer-facing Dall-E product? https://twitter.com/model_mechanic/status/1517205808680869888?s=20&t=pf3nPFyKmuVZk9eM15EYhQ

Also, semi-unrelated, but is there anyone exploring potentially leveraging this approach to allow animation? Each frame of an animation is a very minor text diff 🤔

Aditya Ramesh (@model_mechanic)

Unmodernizing an iPhone with #dalle text diffs

Likes

1123

Retweets

164

▶ Play video

Twitter · Apr 21, 2022, 6:17 PM

[sullen silo]

The frame by frame part I guess can be done using the edit mode. You save the original image for the structure and then proceeds editing it for each new detail, using new prompts to try to reach the final goal. Can be a fun thing to try, although, with luck, for a two second video 24fps, there goes the entire quota for the day... 😦

[sullen silo]

hi @neat bridge , just to add on that scenario, I had a similar problem with an image depicting a smoking character as well. The funny thing is that dalle itself had generated the original image. Once I tried to edit it, it showed me the warning.

[rugged wing]

since i registered at 5pm do the 50 prompts always get reset at that time or if i spend all the prompts by 8pm do they reset at 8 or still at 5?

[serene bronze]

They are reset on a rolling 24hrs, so use 1, it gets back to you in 24hrs, use 3 an hour later, those 3 get given back 24hrs after they are used.

[rugged wing]

thats cool, thanks 👍

[serene bronze]

each sort of credit/prompt/token whatever you want to call it has its own 24hr refresh timer essentially

[rugged wing]

thats a good system tbh

gotta stop the temptation to do all the prompts in an hour lol

[sand basin]

This is almost a philosophical question. My conclusion after the last month experimenting: Do not ask of Dalle what you NOT want, ask of Dalle only what you DO want. Ask a child NOT to lick the window and what will it do next? Just don't inspire the kid that licking the window is an option in its momental reality. 😁

[charred imp]

Actually, in my experience negatives do work. I used 'no text no letters' and it worked for a particular prompt that previously generated random silly 'words'

[broken kernel]

Funnily enough I found out yesterday that you're not allowed to ask DALL-E for windowlicker

[sand basin]

aha! good to know. Still my preferred way of prompting is asking for what I'd like rather than telling it what not to do. Its personal I admit 😉

I am making my first variations on existing art... I did Art Nouveau stuff earlier with Dalle, worked beautifully. But now with variations on existing Art Nouveau, it disappoints... What am I doing wrong? Or how can I help Dalle do this better?

[brisk thistle]

I tried a few times adding 'no text' or 'no words' to get rid of the random words appearing, but then it would just amend the characters used in these words with the characters used in the negative part of the prompt :/ It seems like it also depends on how 'certain' dall-e is about the graphical representation of a concept, and if it can't come up with something it will decide to just spam text or smth...

[brisk thistle]

I think variations ignore the prompt completely and just tries to make random variations of the visual style and elements in the picture. The few times I tried it, it disappointed in the same way.

[charred imp]

You're right. I just tried ' a poster with no text no words" and doesn't work, dalle probably cannot imagine a poster without text. But it worked when I asked for abstract painting about math . I would be great if we could simply use weights for prompts, including negative weights.

[lucid ravine]

always love the surprises - the creation with the interlocking puzzle is wild.

[sand basin]

Yes, my experience too

[sand basin]

The first generations of variations are simpler than the original: less complex, even turning faces into dolls' faces, next into Pixar style faces, next into Playmobil faces 😂 But back in the day, when we had more than 50 requests, I had Dalle delv deeper and deeper into generations of variations. And: Dalle introduced new #complexity on its own initiative, randomly in the images, for me to decide what looked promising, creatively, so it seemed. Almost like how #fractals work. That way Dalle and I embarked on creative journeys through generations of variations together then, with fresh, #unexpected creative outcomes. I still fiercely hope, that sometime later, we can do this again, like biological #evolution did/does it too. Going deeper and deeper. (Is that why #Google named their AI #Deepmind?)

[open badge]

I think Deepmind comes from the term deep learning

but ey, i might be underestimating the name

[fervent yarrow]

Is there a way to prevent cropping with modifiers in our prompts? Outpainting is tedious. I had some success by adding “full body” to this query of a pug.

[fervent yarrow]

Has anyone noticed that it seems like Instagram's algorithm penalizes DALL•E output? I'm definitely getting fewer views and likes than I do with other artwork I post. Trying to figure out why. I wonder if Instagram recognizes dalle hashtags, tagging @openaidalle, or if it even recognizes the colored watermark on the images? I can think of several reasons why Instagram might do this.

[plain plume]

insta algorithm

[snow jolt]

What happens to the images we generate, are they all saved on their end? If so would selecting "garbage" image results for deletion on our end help them at all if we could?

[dense mortar]

After the maintenance will our counter be reset? because I lost track day 1

[peak sky]

If you upload an image to dalle is it added to the training dataset?

[deft silo]

It looks like it from the warning message when you upload but I'm also curious about this. I wouldn't mind uploading a picture of myself to do some inpainting if it didn't but I can't be sure.

[deft silo]

Unfortunately this most probably wouldn't help, the main constraint is compute ressources and an image stored somewhere doesn't impact it

[rugged wing]

Hi! I'm trying to understand the faces policy, if the face was not generated by Dalle like in this picture, can we share the content online?

since i'm here.. why can I access old generations over links but not through the UI(limit of 10)?

[plain plume]

Just curious, how the DALL-E "censoring" works? Is it something, which is put in place afterwards from a second instance? or is it "in the original Model". And will something like "banning mechanical weapons" also will have an influence on details of other mechanical objects?

[south fox]

A question to 10>6 images reduction: what about our already created images, will they still remain in amount of 10, or will 4 cut out, if we come back to our older completions via bookmark?

[lilac ledge]

@neat bridge sent as a DM , thanks!

[pale shell]

The history sidebar will only show 6, but if you view an older bookmarked submission you will see all 10 images.

[pale shell]

We're still iterating on the best UI here. Paginating through history would certainly help, but we're also hoping that the ability to explicitly save the images you like (coming later today!) will reduce the need to dig through your history to find things.

[south fox]

Thank you!
And (sorry again) I have another suggestion: is it possible to have a prompt shown in the one-image-mode? At the moment we can see the Prompt only in the overview with 10/6 images, but not in the single-image view.

Thank you!

[pale shell]

Out of curiosity – why do you want the prompt visible there? Is it because you are bookmarking the individual image URLs?

[rugged wing]

thanks for the response, i like that! Will check it out when it rolls. I also like the idea of complete paginated history with the availability to erase generations or "like" generations onto a special "likes" collection. So users will have complete curated history and a separate likes collection. But this should go onto suggestions 😅

[pure pond]

Hi, anyone having luck with bioluminescent mammals/creatures? Can’t seem to drill down to the prompt to get one to generate.

[deft silo]

Same here (sorry not very helpful 😅 ) but yeah I tried it some time ago and couldn't even get it to make satisfying bioluminesctn fish

Or was that transluscent bioluminescent fish? hmm

Usually DALL-E seems to have a harder time with stuff that's less "life like"

For instance I have a much easier time making a sloth climb something than I would with say a dog

[simple sigil]

Are prompts including guns or cannons, or other instruments of war, such as tanks or fighter jets, allowed?

[deft silo]

I don't think they are, I know that for guns at least it will prevent you from generating those

[pure pond]

This makes a lot of sense. It’s difficult to push it to do abstract takes on life like or true to life concepts. However… I did have a tiny bit of success just now…

[deft silo]

Oh that looks cool! Yes you might be on the right track

[pure pond]

Thee prompt was “transparent skin”, and then the light color from within. It took like ten tries to get this so imma pause for the day on this effort lol

[deft silo]

Ahah interresting! It seems sometimes just being descriptive is better than giving higher level, more abstracted concepts

[south fox]

For better orientation. Sometimes I'm working with different bookmarks and different, slightly modified prompts, for example for my animation experiments. There, if I open images in differnet tabs, I after a while don't know, which prompt belongs to which image (I always have to come back to the overview). This is nozt a curcial issue, but would help 🙂

[patent sequoia]

Is dalle down right now? I can't get in

[south fox]

Yes, but OpenAI is on it

[patent sequoia]

It's time for the waitlist anxiety all over again 😂

[junior ferry]

Hello everyone. I suspect someone has already asked this, but how do you change the aspect ratio to 16:9 or simply something wider?

[pale shell]

👋 You can't change the aspect ratio within the user interface – the model only outputs square aspect ratio images. But you can use inpainting cleverly to achieve different aspect ratios (see #inpaint-outpaint)

[simple sigil]

How many prompt declines does it take to be banned?

[junior ferry]

👍👍👍

[half tartan]

Anyone know why I got 6 images instead of 10?

[candid kelp]

new update in order to let more people use DALL-E :)

[half tartan]

Thank you - now seeing the update channel

[sacred valley]

Would it be possible to retroactively add any of our pre-update generations to “My Collection”? 👀 Would be cool to eventually view a gallery of every single one of our generations

[fervent yarrow]

Lol, I just read this after posting the same thing. I will delete mine.

[tender bane]

Copyright question: Since OpenAI owns the copyright at the moment, I'd love to see some consideration on the rights of the individuals involved in the human element and what they create. As I'm sure we are all learning, it still takes a lot of creativity to come up with some of these prompts! So I know I'd be disappointed if someone else started sharing "my" images around, not giving me any credit because I don't have any copyright claim to them. (Actually, maybe this is more of a suggestion than a question?)

[quiet hollow]

When I press "save" I might lose images without downloading them?

[stoic gale]

Why does Dall-e struggle so much with text? It's fascinating. The words are almost right 80% of the time. Did you handicap it's ability to generate text for some reason or is this just a recognition issue?

[stoic gale]

Copyright the prompt 😂 jk

I'm also wondering about the copyright situation. What's the legal framework like? I didn't think you could copyright computer-generated images in the US? Asking out of curiosity more than anything.

[neat bridge]

+100, it’s really important to us that everyone understands that DALL·E is a powerful tool meant to help humans, not replace them – and that it’s only as creative as the users using the tool.

We love regramming (?) what our users create if users first publicly share their generations, with attribution on our Instagram -- e.g. https://www.instagram.com/stories/highlights/17939266553053203/

Please let us know if you encounter cases that make you feel otherwise?

[neat bridge]

precedents say that there needs to be "human authorship"/involvement for anything to be copyrighted -- we're in uncharted territories, so we're trying to figure things out when it comes to copyright & ownership!

[stoic gale]

Thanks for the reply! I really appreciate it. As a bit of a tangent, I'm curious, how do you use GPT-3 and Dall-e internally within the Open.ai offices? Is automation a big part of your workflow? Thanks

[half tartan]

Is anyone else struggling with rendering reliable photorealistic images ?

Also having an absolute blast testing the limits with different film cameras & types of film. Absolutely inspiring photographs.

[dense mortar]

It'd call it a prototyping tool to help with composition, inspiration, as well as a better image blender (like a more dynamic 'content-awareness' than photoshop has.)

How you use the tool is up to you.

[broken kernel]

definitely. it can be very frustrating when you fail to get any useable images but then something comes along that blows your mind

[visual gale]

Is anyone else having trouble logging into DALL-E today?

[sullen silo]

Yeah, I have been experiencing some server issues in the last hour. But intermittent ones

[visual gale]

I cannot log in at all. I wrote for support. I didn't want to bother them if other people were experiencing this problem. Thanks.

[fervent yarrow]

Exactly. I would humbly suggest that, technically, no one really knows who owns these copyrights. I would also argue that in our civilized society, someone should. And I would also say that machine generated works can be novel or transformative, and therefore should be eligible for copyright. As an artist, I’d like to see a future where Dalle is considered a tool like Photoshop, and I own my creations. That said, I can envision different copyright assignments, especially if openai adds a novel layer on top of the core image generation. Finally, openAI might want to assert that there are certain ways that images can never be used, and that might require an accommodation.

[clever mist]

i would argue that some of these content violations are ridiculous. i also received this for having godzilla watering marijuana plants.

is preventing the generation of art depicting the passion of the christ intentional ?

we're not talking about things that are anywhere outside of a G rating but something that can be found in every single church.

additionally marijuana plants are not illegal in most of the united states.

[indigo rune]

has anyone felt like recently the results have been sort of off? like im getting a lot of prompts that aren't related to what I'm asking (a lot more failures), mainly when asking for photographs or photorealistic things that are not realistic (like the tiger with horns one)

[patent sequoia]

Hey, my DALL-E is outputting less good results, I tried a prompt I know looked good last time and it generated something worse, what's happening?

Uhm?

Is DALL-E nerfed or something?

[neat bridge]

super weird, I’ll flag it to our team!

[quiet hollow]

if we do not download the images, there is a chance to lose them? Or once we save them are safe?

[open badge]

all images are in cache rn

[open badge]

after generating you can go back into your search history and click back to your old prompts and the images generated should be there

you can ofc also save the links to the generation you want to keep

[clear aurora]

Quick question about posting again publicly - I would be creating PBR materials presented in a style like this on a model - and underneath I would post the unedited DALL-E image with the Logo - is this okay? (just being cautious!)

[torpid umbra]

Anyone found a reliable way to generate a new image, but re-use a character or object from a previous image? Say I generate an illustration of someone in a forest. I want to generate a new image of a very similar looking character on a busy city street. It looks like the only way to do that at the moment is to use the edit feature and try to erase everything around the character?

[sullen silo]

Yes, the only reliable way I found is to download the image, remove the background, save as PNG with alpha transparency, upload the image and edit with a new prompt

[torpid umbra]

Okay good to know, thank you!

[candid kelp]

This is my screen time on iphone. I've never seen this happen with anything except spotify playing music in tbe background and counting it as screen time.. it's showing that I am using and looking at that webpage 24/7.. This is just from having dalle open in Brave mobile in the background and i guess not closing out of the app?? Does anyone have an explanation for this?

[young flume]

Is there any way to get DALL-E to understand relative size or scale, or to make objects bigger or smaller than they normally are? For instance, "a mouse that is the size of a car" or "A mouse that is bigger than a car", or "A mouse sitting next to a car. The mouse is bigger than the car". Is there any way to get uncharacteristic sizes of objects? Thank you!

[glad crypt]

have you tried generating one object (e.g., to the left or right), then edit the image of your choice and change the prompt to have both objects?

[young flume]

I did, but it wasn't very effective when I tried it. I probably need to do more tests. I found that "towering over the houses on the street" seemed to work in one case. Thank you!

[lusty python]

When will Dall-e be available to the general public and commercially to companies?

[sturdy vine]

Has anyone had the experience where inpainted content almost always looks significantly worse (less coherent, more spotchy, etc.) than the original content? If so, do you have tips on how to fix this?

[clever mist]

do you know if it's possible to generate valid flowcharts in DALLE2?

i am assuming the answer is 'very unlikely'

[clever mist]

incredibly enough it does do semi-proper time complexity visualization

[clever mist]

if you look at those diagrams, the top resembles sudoku grids and traveling salesman problems where as the bottom image looks like diagrams of algorithms that can be linearly computed (multiplication, sine, cosine)

not shocking because it's associating those tokens with the images

however the top left diagram of the bottom image resembles https://en.wikipedia.org/wiki/Star_of_David_theorem

[neat bridge]

lol definitely not, but i actually wonder if codex could do it through a data viz library (if you know how to code)

[clever mist]

i have done software dev for >20 years.

does the codex presently have a visualization library?
if not is this something that could be easily added?

[visual gale]

If I ask DALL-E to create a face with one eye covered by hair, does that count as a face that cannot be shared? I'm not sure if half a beautiful face is considered to be a full face. I hope that made sense. ❓

[patent sequoia]

When will text be unobfuscated?

[civic wharf]

I was curious, is DALL-E mini (/maxi) in any way associated with OpenAI? If no, how are they able to lean on the DALL-E moniker?

[autumn mountain]

No it’s not, I think openAI has asked them to not use the DallE name.

[civic wharf]

They apparently haven’t listened lol. Their current model is quite comical results-wise but my worry is that the general public is now going to conflate mini and the real thing lol. Oh well.

[autumn mountain]

Agreed. It is concerning that they have a similar name.

I wonder if legal action is possible or being considered

[sturdy vine]

It is quite common in ML for folks to use variations on previous model names. DALL-E mini was built a while back, on a similar design as the original DALL-E, not this new DALL-E 2 API. I doubt OpenAI intends to litigate that practice, would look pretty bad

[autumn mountain]

Makes sense. I was unaware of that practice.

[visual gale]

If I ask DALL-E to create a face with one eye covered by hair, does that count as a face that cannot be shared? I'm not sure if half a beautiful face is considered to be a full face. I hope that made sense. ❓

[charred imp]

I really wonder why OpenAi haven't reacted sooner to that. There is dalle mini, ruDalle and Dalle mega model is being trained and will be released soon. And the company who created Dall-E doesn't care about 'copycats' using the name of their best known product? Dalle mini has been relased a month ago, training that model wasn't a secret either. I just guess openai didn't worry beacause they didn't expect the copycat will be so good (for an open-source, free to use model) and that it will become so popular.

Although, this is the first time that a name of a neural network model became known to a wider public, so this is unprecedented territory

Researchers wouldn't care if someone named their model 'resnet mini' bacause it wasn't a consumer-facing product

[stoic gale]

@clear aurora Given your background in texture & material design; I'm wondering if you've found any good ways to generate normal or height maps using Dalle2? Do some prompts work better than others for texture generation? Are there any keywords that we should be aware of? Just curious about your findings. Thanks!

[clear aurora]

So far - no XD
In both DALL-E and Midjourney so far, height data has been barely useable with heavy edits at best (Please link me to any successes if you have found!)
Normal maps I have not tried in DALL-E, though TBH just normal maps themselves with no other context have limited uses IMO, or only in very specific situations - I would prefer good height info as can create normals from it, but also edit further and use for displacement.

So far the main success is with very good front on photo like pictures, in which Substance Sampler then has a great AI to generate decent normals from and delight albedo. (Height is still not great in sampler yet).

I'm going to be doing a write up later with feedback but so far DALL-E has not been very good with grunge maps either for me, and I had more success with Midjourney for those (Both prompt and image variations - I have an artstation post on this) . This could be a resolution issue, but I dont think thats entirely it.

However, for photo like images and editing for seamless images (Including making my holiday photos seamless :D) DALL-E is currently unmatched 🙂

I get the feeling that texture specific things like height and grunge maps (and normal maps?) Might be due to not having those data sets? I'm only guessing but I'm picking one of the data sets DALL-E is trained on does not include creating specific texture maps. If thats the case I would love if thats something that can be done / added in the future, and I'd love to open a discussion about what we can do to help DALL-E produce the desired results in this category. Tagging @trim kite for visibility on that talk 😄

Would love to hear other opinions / experiences on this topic!

[stoic gale]

I haven't given it a real attempt yet. This was the best out of 5 prompts that I ran (I just threw the leaf together quickly in blender). I'm not totally satisfied with it, but I'm definitely going to keep pushing into this area of exploration.

[clear aurora]

Thats awesome! Can you share your prompt wording for the leaf height?

[desert ice]

Hi everyone, I’m trying the un-cropping technique using Dalle. If I have input an artwork into the system, does Dalle own the copyright of the artwork?

[quiet hollow]

a diff q on my end: uncropping works by importing a transparent image? Does Dall-e understand trasparency?

[open badge]

Yes

You just have to click once on the transparent part and then you can go ahead and click generate

[quiet hollow]

Mmm i dont seem to get it, so i am here. What do i press from here?

if go to edit, even if i click on transp part, generate is greyed out, unless i type something

[open badge]

Oh yeah. So If you wanna Edit You should describe the image in question so DALL-E knows What to Do in the transparent area

Don’t forget to describe the whole scene. Not Only what you want in the transparent area

[quiet hollow]

i just wanna fill it though with the rest of the image

shall i type "fill it"?

[open badge]

Yes. Then you just write What your original is. Like bottle product 3d render

Like describe the original and then it Will extend it

[quiet hollow]

so here i need to redescribe the whole image?

[open badge]

It doesn’t need to be that exact. Just like 3D render of bottle

[lucid ravine]

should the edges of the inset image also be inpainted ever so slightly?

[open badge]

Oh shit Yeah. That helps

It helps remove the seam

[quiet hollow]

so i just typed without painting anything, extend transparent area, much more easy 😄

worked

[open badge]

👌🏻👌🏻

[open badge]

This is How you make composites aswell btw. So you can generate images. Crop them and then extend them

[quiet hollow]

yes i mean re describing a thing that is already there and extending is weird approach. thanks for the help!

[open badge]

No problem 😄

[stoic gale]

"A leaf, in the style of a Unity height-map, orthographic camera, Unity Engine" I also tried Z-depth shader, Z-depth only, displacement map, displacement only, height-data, height pass and z-pass, but none of them worked particularly well.

[dull furnace]

What’s the timeline for an API for Dalle? Would be really great to see the text to image capability offered in a more developer friendly way such as is done with GPT

[candid kelp]

for inpainting, is there a difference between making the area transparent in photoshop vs within dalle? like, if i give the full image to dalle and use their erase tool, does it retain the knowledge of what was previously there when generating something new in the transparent area and does that affect the outcome? vs if i give it already transparent it will never know what was there

[dull furnace]

I don’t believe it matters from what I’ve seen

You can easily test this to make sure

[torpid umbra]

Are we allowed to share photos that contain a face if the face is very low quality? Or no faces at all if it's depicting a photo? For example:

[simple sigil]

No. OpenAI owns the output based on the image, but not the original file itself that you upload.

[whole cove]

Wasn't sure if I should ask in here or tip & tricks, but has anyone had luck making characters with unusual skin tones? I've tried a few different colors (red, green, blue, + more specific colors of those) and keep getting normal skin tones

[spring wyvern]

i've heard it said that for such 'unrealistic' things, try prompting for non-photo media (e.g: 'unreal 3d engine', 'photo-realistic' render) as opposed to photos

[whole cove]

I'm actually sticking to digital art and painterly styles as I'm wanting this as more character concepts

[spring wyvern]

ahhh i see. in that case i'm all out!

[whole cove]

This is the closest I can get to even semi red, but the character's skin tone is supposed to be like the color of her dress

[vestal widget]

I came into the same issue. If anyone has any prompt ideas that produce fantasy colored skin (purple,blue, etc) I would love to know!

[livid stream]

Looks like the “Openaidalle_2” IG account was deactivated…?

[serene bronze]

Yeah, that one was reported and deactivated, can see info in General-chatter about it : #openai-chatter message

[stoic gale]

Include the words “art-station character concept art” in your prompt. That will help. Also, you could try starting with a hand or a foot (in the colour you’re looking for) instead of a blank image.

You could also try adding” humanoid alien” because they’re usually off-coloured humans.

[whole cove]

Definitely gave me nicer looking art, but sadly, still getting only normal skin tones 😦 thanks for the suggestion though! I love the style that the first suggestion gave me

The top two rows were the ones I saved after trying it. I managed to get that one reddish one, but only cause of lighting. I'm going to try playing with prompts about lighting

Also will definitely try seeding it!

[stoic gale]

Here’s another idea: download one of these, tint the skin in photoshop using the hue/saturation slider, re-upload, run variations.

[vestal widget]

I am getting some results with trying "(color) (character) with (color)body paint" but they are inconsistent to say the absolute least

The intended color usually ends up on the clothing (see that yellow horror right there 😂 ) or with a pattern added. Fun, but might have mixed results with the subject you are creating.

[whole cove]

THANK YOU THAT WORKED WONDERFULLY

[stoic gale]

You need to point it at an area in latent space where it’ll be able to find these references.

“Infinite Vibes” used a great example earlier: “Cat riding a bike” wields poor results, because that’s not something that cat’s (as a whole) do.

[whole cove]

Honestly I was suprised it wasn't able to give me a red-skinned woman cause I tried devil, succubus, and tiefling (all of which are commonly red) combined with my description, which I would've thought would help it

[rugged wing]

is each and every prompt you do timed, and if so is that why you have to sometimes wait until a certain amount of time to do another generation?

Your limit doesnt get refreshed every day?

[serene bronze]

correct, each of your allotted 50 is on its own 24hr timer