💬｜general-chat

[young hatch]

@haughty wasp Does Fooocus meet your need? Or am I misunderstanding your question?

[haughty wasp]

Fooocus is a web UI like A1111?

if yes, then It wont help

I need an outpainting model that gives good results without using web UIs

[young hatch]

Yeah - I think Fooocus uses the same models - yeah, its a webUI - I read your question wrong

[red scarab]

Why would one of the images created be blurred

[warm junco]

On Civitai you can find many inpaint models (these are used for outpaint)

Or you integrate controlnet outpaint that doesn't need an inpaint model

[fervent thunder]

Use gui
Install https://github.com/huchenlei/sd-webui-api-payload-display
This will display the api call made when generating image.

[dim spindle]

Im about to undertake trying to create a lora. I know the basics of creating one but have a few questions still. (I cant upload images to this channel 😦 )

The lora is to create images in a unique/distinct style of a certain artist.

They are of top down D&D tokens/minis. The figures have/cast a shadow. For the purposes of a lora, will this matter?

The other one is: What to include in the descriptive text. How descriptive does it need to be and is there such a thing as 'too much'?
Currently Im thinking of: [art style] [perspective] [gender] [race] [class] [wearing] [items held]

Does it need more/less?

[fervent thunder]

Trying to build a nudifier arent u

[paper lily]

Hi. How can I get the ruler button?

[bleak matrix]

Good morning, everyone! How are we all this beautiful day?

[sudden ruin]

What ruler button and where

[robust sail]

hi

[solemn kettle]

Hi there! Has anyone downloaded stable diffusion locally onto an external hard drive?

If so, I just ordered one and was hoping you could either walk me through it or send me a YouTube link on how to save it to my external harddrive

[mild grove]

Could someone who does animation at Comfy help me?

[sudden ruin]

#🎥｜animation

[sudden ruin]

#🤝｜tech-support

[burnt shoal]

why do people hop in vc muted, and aren't there to talk or anything?

[marble steppe]

what are the requirements for running SDXL on a local machine? I have 3 videos I could utilize.

[marble steppe]

they're good listeners?

[burnt shoal]

but noone is even in vc

apart from them

[marble steppe]

they're the best listeners?

[warm junco]

Nvidia GPU with 6gb of vram

[burnt shoal]

does it need to be nvidia?

like can 7900xtx run it

[warm junco]

Nope for AMD on Windows it would be 16gb
AMD on Linux 12

[burnt shoal]

oh that's what you meant

so 7900xtx would run it just fine?

[warm junco]

Yes should work fine

[burnt shoal]

and one more thing, where can i find the docs for xl?

I looked in resources, as well as the channel pins for tech support

[warm junco]

Upscale could error out

[burnt shoal]

okay, no worries

[warm junco]

I will write a AMD Knowledge guide on Github someday

[burnt shoal]

any idea as to what i can follow for now? be it nvidia or something?

[warm junco]

For prompting?

[burnt shoal]

correct

I'm kind of new, and the info about xl is quite unclear from what ive been able to find

[warm junco]

Then this is the perfect guide from our Team member Andrew.
https://docs.google.com/presentation/d/1HEcE3qOAGVujcDaNQbiLXyx7zwKHQkXEILsYBhsot7A/mobilepresent#slide=id.g1e409d50d6c_0_177

Its the SDXL Promptbook

[burnt shoal]

this is awesome, Thank you so very much.

[shut moat]

does anyone here use SD-Turbo (2.1)? do 2.1 lora's work with it?

[burnt shoal]

uhm

the promptbook does not actually say how to run sdxl on my own machine

am i missing something?

[warm junco]

You asked for prompting xD

Not for running it

[burnt shoal]

ah, im so sorry, perhaps there was a misunderstanding

[warm junco]

On windows directml you just need to drop the model inside the models/Stable-diffusion folder

Then you also should download the sdxl vae file and put that into models/vae

[burnt shoal]

oh that's easy enough, Thank you.

Where would i find the model/ vae?

[warm junco]

The SDXL 1.0 model:
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main
The Vae:
https://huggingface.co/stabilityai/sdxl-vae/tree/main

[burnt shoal]

You're the best.

Thank's so much.

[warm junco]

You can also find community SDXL models on Civitai.com

[burnt shoal]

I see.

[versed yoke]

https://fxtwitter.com/echo_pbreyer/status/1747272304671998094

[marble steppe]

Can you use multiple cards together or does each card need to have 12GB of RAM on Linux?

oh wait... I do have a card with 16GB lol

I have an RX6800, and two 5700 XT's

what RAM?

or rather how much RAM

[solemn kettle]

Ty

[warm junco]

16 on Windows, 16-32 on linux

[spice plank]

Anyone recommend any good models that looks like people took it with a phone camera/etc? The ones I've tried, the skin is too perfect, lighting is perfect, it doesn't look like a real picture at all its just... idk... weird, I have like 50 models and they all have this problem

[shell hazel]

Have to describe the imperfections in the prompt, use the word like 2 or 3 times to make it weighted. I use things like Dark spots, uneven skin, rough skin, blotches, pores, sweat, etc

[daring mountain]

can a SD checkpoint apply photorealism/ skin/ body to a LORA trained on high poly detailed mesh head?

[magic tangle]

OutOfMemoryError: CUDA out of memory. Tried to allocate 160.00 MiB (GPU 0; 8.00 GiB total capacity; 6.92 GiB already allocated; 0 bytes free; 7.24 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
Time taken: 0.8 sec. how can ı solve this guys? I am not god at hardware stuff. Thank u

[scarlet moss]

You could try search on Civitai for images that produce your desired result then click on them to see what model they used and what prompts they used

[warm junco]

What's your GPU and what's inside your webui-user.bat?

[marble steppe]

I assume it's probably better to setup on Linux?

any particular kind of CPU ?

[warm junco]

For AMD GPU yes,
CPU doesn't matter for SD

[marble steppe]

any particular distro? I've never used Linux before except for the mining rig I

im running that has a custom linux os on it

[sly eagle]

Hello! anyone know a notebook or code for runpod that train voices and can use trained voices?

[warm junco]

Ubuntu or Arch Manjaro

[fervent flame]

Hello, I am currently looking for a paying consultation with an expert in photo editors/applications for creating AI avatars (photos). The user case for the application is as follows: the user uploads their selfie and receives AI-generated avatars in 20-40 styles, featuring their face, of course. An example of the application can be found here (link to the App Store: Reface: Face Swap AI Photo App).

Currently, I have two tasks:
1/ Figure out how to technically make the AI avatar work (cases described above).
2/ Understand what prompts to write for generating AI avatars - assignment below.
Right now, in the App Store, there are various photo applications with AI that generate cool avatars in different styles based on your photograph. To generate such images, you need to write prompts. Now, I need to find someone who can write cool prompts. We plan to use Stable Diffusion with the XL 1 model.

if you are an expert, please write in the general chat or in private messages

[dry knot]

hey guys! what would you use if you had to face swap but it has to be very realistic?

[warm junco]

I would use Reactor or Controlnet IP-Adapter, the later one can also do freckles without smoothing the face to much.

Mostly the thing you want

[dry knot]

interesting!

thanks for your answer

will try that

[warm junco]

Reactor uses face restore and that will smooth the image to much

Its still good but has it's limits

[dry knot]

i see

ipadapter is a bit more complicated no?

[warm junco]

IP-Adapter is very powerful ans requires testing

[dry knot]

yeah i saw that

reactor is only one node

ip adapter seems more advanced

[warm junco]

Yes

[spice plank]

woops, wrong AI chat, thought this was text gen AI

[vagrant raptor]

If I were to buy a desktop at around $1300, and I don't feel like building a PC again (been there, done that), what would you all recommend? Brand? Website? Good deals? Gaming and AI.

I don't need monitor and such

[warm junco]

Its hard to recommend websites because there are not much sellers with world wide shipping ^^

[vagrant raptor]

True, I'm in the US.

[vapid tangle]

why does sdxl ignore prompts sometimes. Like I tell it to make a parrot with A BIG BEAK, even in parentheses, it never makes the beak big

[final fog]

is there a way to set cadence scheduling on deforum? after looking at this it looks like being able to schedule it would be very useful https://youtu.be/oRNCK_0w2uc?si=OoOwqXv3E5pKMECY

[vagrant raptor]

Consider a toucan prompt and modify the color?

[vagrant raptor]

But to answer your question, I don't know, my response is to try a different model usually.

[placid hatch]

Is the rx 7800 xt a viable option for SD right now?

[vagrant raptor]

I have Radeon RX 6800M GPU with 12GB GDDR6. I have to run lowvram with 12gb and have to fight with everything because it's all made for Nvidia and Radeon is still trying to play catchup with basically everything. Just walk away.

Note my question earlier? I'm looking for an Nvidia tower because my radeon is working almost as good as a low end Nvidia, but not quite as good.

[scarlet moss]

I was using an RX 5700 XT with 8GB vram and it was awful lol. Was taking about 2-3 minutes to render 1 single 700x1200 image and would constantly fail right at the last second due to insufficient vram and the whole image would be lost. When I switched to RTX 4070 with 12GB vram, it was like literally 40 to 60 times faster (taking only a few seconds to generate a 700x1200 image), and I have not had a single vram issue since, even with 4 to 6 times bigger resolutions.

So if a project took me 1 week to do with an RTX 4070, it would take me like 1.5 years to do with an RX 5700 XT, due to the constant Vram failures and extremely slow speed.

[trim trench]

Hell

[scarlet moss]

I would say an RX 7800 XT would be viable based on the Stable Diffusion benchmarks here: https://www.tomshardware.com/pc-components/gpus/stable-diffusion-benchmarks#section-stable-diffusion-768x768-performance

But I've read that the Stable Diffusion implementations that AMD gpus run with are much much less efficient with vram, so despite having passable speeds, you'll still effectively have much less vram compared to similar Nvidia cards, so you may not be able achieve as high resolutions before running into problems.

[vagrant raptor]

I second this. Sometimes I turn off --lowvram and go with --medvram and enjoy the extra speed until it fails.

[tidal agate]

We are looking for an experienced developer who have worked with Stable Diffusion (Fine tuning Dreambooth) before

We have an existing training and inference codes that runs on A40 NVIDIA and we would like to switch it to NVIDIA T4 and work with latest torch 2.1 and cuda12.1.

Please post the input the and output of the samples dreambooth photos that you have achieved before before bidding on this work

[junior nest]

hi how to set up stable diffusion on macbook pro m3

any guide please

[languid helm]

DALL-E in chatgpt like sd in llm when ?

do we have any research papers on that atleast ?

[warm junco]

Hey, check the pinned messages of #🤝｜tech-support message

[still copper]

hi

[halcyon willow]

What is the development currently focused on?

Efficeny, quality, speed, length,....

[scarlet moss]

Yeah I just tried a few prompts and still the same issue. The closest I could get was 'sword submerged into ground' but that ended up having the sword stabbed into a puddle lol. I think this might be an issue with the model being used, they just haven't been trained to recognise swords stabbed into the ground.

[empty star]

Best upscaler for landscape photos?

so far I'm liking LDSR and NKMD Superscale, Esrgan seems too soft and Ultrasharp creates "fuzzy" texture

[hexed timber]

Hey everyone

I am new to AI

Where do you people learn about the terms like LoRa, SDXL etc

Basically I want to get started in image generation using AI

Where should I start?

[candid wedge]

How to upscale an particular image?

[hard orbit]

what are you guys with a 4070 running as arguments?

im running xformers half vae currently but i think thats what i was running with the lower gpu

[warm junco]

Thats the only thing you need.
But you can try add --opt-channelslast
As it could provide some benefit or none.
How long does it take for an SDXL image on 1024x1024 with 30 steps?

[hard orbit]

8.7 seconds

no sorry that was 524...1024 is 26

[warm junco]

Thats good

[hard orbit]

yea its fast i just wasnt sure if those arguments might be hindering how well it works ...speed is not an issue for me now

[warm junco]

xformers is the performance boost and no-half-vae is for sdxl vae compability

[hard orbit]

ok so leave it alone

its not broken by any means I just like to ensure its optimal

[warm junco]

hey, on some websites there are good explanations like this one:
https://stable-diffusion-art.com/how-stable-diffusion-work/
also take a look at #1080946152318443610 for some helpful links around SD.
If you have a capable PC to run SD localy feel free to check the Install Guides in the Pinned Messages of #🤝｜tech-support or ask there for any technical help.

[warm junco]

yes, your args are the best for your GPU rn

[hard orbit]

Thank you!

[fervent thunder]

If i wanted to turn images of clothed men in underwear + 6 pack and is very realistic, what would the workflow look like

Would it be just inpaint, do i train lora, are there other steps to make it more realistic etc

[empty star]

what is the best sdxl canny controlnet to use? the diffusers one?

[marsh vigil]

hello! when constructing captions to train models, are newlines treated the same as other whitespace? for example, is a, b == a,\n b?

[frank rampart]

Hey hive mind. ive been on stable diffusion for afew weeks now and loving it. wondering how to get started with creating anime/videos with it? Can you recommend a youtube guide that i could follow along too for the weekend?

[fiery fox]

Hi, I'm using A1111 and the epicrealism_pure evolution V5. But why my model always have grey clothes on? I can write, orange dress, wearing orange dress, ((orange dress)) and so on. Nothing works. very rare cloth stuff like this works. I think jeans is the only thing that more or less works. But why? Is there only certain types and colors of clothing that works? Or Am I doing something wrong?
And another question.. my model is always looking in the camera. No matter i say, view from front, view from behind. She turns her head like an own. How can i say not to look into the camera. nothing works.
Hope someone can help me with those stuff. Thanks

[frank rampart]

ive downloaded multiple models and they dont always work as desired. try a different model.

send us a screenshot so we can see your model, prompts and resulting image

[warm junco]

Hey, can you show an example of grey clothes?

Because prompting for yellow shirt
Should be enough to get one

Both of your questions are perfect for #📝｜prompting-help
Prompting for "Looking away" sometimes work

[inland egret]

Guys, an off topic questions. Does anyone happen to know a good free online PDF ai summarizer?

[warm junco]

Hey, https://online2pdf.com is nice

[molten sage]

Hi everyone! Is there a way to test multiple Lora strength (weight) at once in ComfyUI?

[warped isle]

I dont know if this the right channel to ask this but does anyone know what app to match colours from 2 different picture? As example putting someone else picture to my picture, it does have different lighting, shadow, color, etc.

[crimson dagger]

Hi! I installed comfyUI locally and am trying to use "comfyUI manager" but after installing it the comfy launcher stops working. Anyone know a fix to this?

[chilly vector]

hey, I haven't generated anything for a while now, is 1.5 still good enough or are there new tools to work with?

[fervent thunder]

Still good enough for avg use. But you can get much better result with comfy ui and custom workfow if you dedicate some time

[chilly vector]

I just want to integrate image generation to SillyTavern, is that good enough?

[brave jetty]

In the #1100484581037195384 How do people succeed of doing so perfect, defaultless, animated images???

In particular the last animated flowers ones

[fallow summit]

boommmmm

[fervent thunder]

Guys if I have a design I like of a character but I really don't like the face of the specific character, and I just want to replace the face using another image, what do I do?

[chilly vector]

I used photoshop and inpainting to get rid of something i didn't like in my gens

[silent grove]

hey guys gm

quick question, can we use the images we generate here for comercial purpose? Example: ads, books, magazines, etc?

[warm junco]

You can use an extension like Reactor to faceswap it

[abstract pecan]

Hey does anyone use gif2gif with an AMD GPU? I'm running into this issue after reinstalling and reupdating the extension several times

AttributeError: 'StableDiffusionProcessingImg2Img' object has no attribute 'script_args_value'

im pretty noob so bare with me

[fervent thunder]

React extension , easy to use

[low forum]

how to create custom size ?

[modest shuttle]

Hello. I've got an RTX 4090, and automatic1111, and I'm only getting about 13its/sec. Is this expected? I'm on windows 11.

[vast zodiac]

hello how to change the ui language?

[plain pier]

https://tenor.com/cmRKcPBojiL.gif

[hot lily]

anyone know how to make something like a pokemon more realistic?

[frail prism]

There is an image browser built to navigate through generation metadata: https://github.com/cocktailpeanut/breadboard. But it's support ended year ago. Any recommendations for similar soft?

[frail prism]

Nevermind, I found https://github.com/zanllp/sd-webui-infinite-image-browsing, and it's awesome.

[nova turtle]

is somewhere a comfyui chat or server? I have question... how can i reproduce img2img output in sd also in comfyui? in comfyui i missing option like chose preprocessor for softedge (softedge hed) and also active pixel perfect and that my controlnet is more important. Or did i overseen this options? thank all in advanced for help 🙏

[warm junco]

for that you need controlnet extension and nodes

[nova turtle]

are this nodes already preinstalled?

[warm junco]

i dont think so

[warm junco]

you need comfyui manager and there you can easily install them

Video Guide: https://www.youtube.com/watch?v=hTSt0yCY-oE

[nova turtle]

yes i know thx, i already have the manager 🙂

https://ibb.co/fnQFw2T
https://ibb.co/cJvRDbs
https://ibb.co/zGFBpJD

check screenshots pls

this hed option in sd is what iam looking for in comfyui

but in manager i cant find and also in preinstalled processors

[inland wigeon]

hello

i am using stable diffusion via stability matrix. i am new to AI, but have an IT background. my company has a scholarship program coming up and i need some advice.

i am interested in AI, but think i can learn that as i go (conveniently with the help of AI 😄 ) i would like to understand what i'm doing with image creation, as opposed to feeling like i am just knob twiddling. i think what i need to understand better is digital art. does that sound correct? if so, my question is what programs/certifications would point me in the right direction such that i could use the scholarship and achieve a better understanding of using stable diffusion? hope that makes sense...

let's take a music player app for example. i don't want to learn (yet) about music theory, the science of sound, etc. i don't want to, necessarily, learn a particular music player app. what i do want to learn is the information such that when i change to a different music player app, i can intuitively find my way around as a user. what i want is the ability to understand how to make an idea for an image, understanding why a particular image comes out bad, and how and what setting(s) to go to. what all the tweakable settings mean, do and are for.

once i can sit down with an idea from my imagination and have it come out at an acceptable level, then i can continue on to learning about the guts of AI, science, math, etc. 🤓

[nova turtle]

https://github.com/Fannovel16/comfyui_controlnet_aux this is what i was looking for 😉

[warm junco]

Ah okay 👍

[warm junco]

So you want to learn how to use a Stable Diffusion webui

[inland wigeon]

i guess i'm not sure. that's why i figured i'd try to ask. i don't feel i have a good grasp on this stuff so far. maybe i'm misunderstanding how much art plays into this...?

i can follow tutorials, but w/o am lost

i guess i'm wondering if i should try to learn one of the adobe products, or an open-source alternative like gimp or inkscape to understand what the AI is doing??

what it's helping with

[warm junco]

Your using a local webui?
For beginners auto1111 webui is recommended to use.
Then play around with the settings.
Learning by doing.
Dont set the goal to high at start.

[inland wigeon]

sounds like good advice 🙂

[warm junco]

You dont need to learn Photoshop to learn Stable Diffusion.
If your PC has a 4gb nvidia GPU or 8gb amd. You can run it localy and try it out

Image generation is different then drawing images by hand with Krita or by modify images in Photoshop.
Its an own subject

Get familiar with the basics. Then understand what models, loras, steps and cfg etc are.
Generate a few images, try out different settings, and then go to the next subject like upscaling or inpainting images with Ai

[inland wigeon]

gotcha! thanks for the advice 🙂

[nova turtle]

https://github.com/kenneth2001/comfyui_controlnet_preprocessors/blob/main/README.md

looking for soemthing like that,but does no more work?

[warm junco]

Np, at the start all the features can be a bit overwhelming and the community made a lot of extensions too. So for any questions on that you can ask in #📝｜prompting-help or #🤝｜tech-support

[nova turtle]

@warm junco could you help? 🙏

[nova turtle]

here...

[warm junco]

cant help much with comfyui related stuff. But preproceesors should be added by ControlNet

[pliant ermine]

anyone running a 1080ti?

[warm junco]

only a 1080

[pliant ermine]

im trying to figure out a cooling solution, card is new to me but i realize older tech

having a hard time finding out what the standard is

[warm junco]

for cooling the best is to not overclock the gpu, also you should have more than 3 case fans

[pliant ermine]

im looking for an aftermarket kit for the card itself

the fan on it works, but cleaning it yesterday i could feel a rub when spinning it

[warm junco]

and how is the temp?

[pliant ermine]

no idea, havent loaded the machine yet

but if its failing, id feel better replacing it to start with, just havent heard good thing about the stock blowers

[warm junco]

is it a custom model or reference ?

[pliant ermine]

unsure, but maybe reference? pny 1080ti

[warm junco]

then its custom

my 1080 runs with 2 fans since 2018

no issues yet

[pliant ermine]

dropped an image in #🏞｜general-with-images

[warm junco]

the card support 0fan mode, and when it starts to spin you can here a little bit of an noise

but only at the start

[warm junco]

hm, you can still find second hand waterblocks for 1080tis

[pliant ermine]

since you mention it, originally i bought a used 1080ti, which eventually went code43. i found the same exact card and ordered another, this one came with an ekwb waterblock attached. looks really nice actually, and i considered trying to find the rest of the system to use. in the meantime, i swapped over the fans from the old

totally open to another cooling system though

[warm junco]

ahh okay, yea ekwb has nice waterblocks

idk if it makes sense for any other cooling method other than air or water

[pliant ermine]

looks like its missing the actual block that the tubes would connect to though unfortunately

hard to find info on it all now as it says its end of life

block is slick though

[hybrid fox]

how can i add bot to my server so that i can see all of my generations at one place

[cold pecan]

Looking for the same solution. I have seen it one time in a yt video, but cant find it now...

[cold pecan]

Hast du schon was gefunden??

[hybrid fox]

no i havent found anything yet

[cloud lintel]

Hi im new, is it possible to give a picture to make it better, a better style ...

[wind ingot]

Sure, its called img2img , but in order for AI not to trip too much you gotta restrain it from creating too much (low denoiser) , if you know the direction you wanna go the better, you can guide it with words (prompt)

[pale latch]

controlnet helps maintain image consistency too

[inner lintel]

No chances for 1.6 public release?

[pale latch]

i haven't seen any yet, but so many of the community refinements have extended the base 1.5 so well, that it's not really a landmark release

[drowsy robin]

I wonder how many people in this discord has ordered the rabbit R1.

[shrewd grove]

Where can I get "DPM++ 2S a Karras" sampler for A1111?

[pale latch]

i'm using the developpment branch and it's there. i think it's there on stock 1.7 too. it's just not in alphabetical order

[potent spire]

Hell is unleashed on their server as you see lol

Because of Perplexity

[drowsy robin]

As someone who bought it

THey just unveiled that R1 buyers get a seperate full year of perplexity pro

And no I am not advertising it, this literally just happened 65 minutes ago

[potent spire]

Yeah, for 1 year

Which is okay, thats a marketing and business strategy

[pale latch]

product like that is going to be dependant on how phoned in their software stack is

likely is just investor/acquisition bait

[nocturne jackal]

hello

[brisk lark]

Hi, does SD have a feature like MJ where you can give it an image and it will output a description of it (to use as a starter for a new image)? TIA

[brisk lark]

yep worth $200

first 100k buyers

[drowsy robin]

Well we don't know this for certain. I bought in early to see.

We'll see.

[brisk lark]

they're charging $200 for a year of perplexity pro, I just got my credit

so that part is known

[drowsy robin]

No, I know the credit is real

But I mean the product. The LAM's capabilities.

[brisk lark]

Oh, you're questioning the value of a $200 charge, not that they're charging that for it. I guess you can do that too.

All good.

[brisk lark]

Any answer to this?

basically reverse-engineering a prompt from an image

[drowsy robin]

No, I mean...

The device.

The R1.

I'm not even meaning perplexity's credit bonus, I mean the product itself.

There isn't enough usage information of the device

We have specs and small demos

[brisk lark]

ok, your'e responding to my comment stating they give you $200 credit to Perplexity. That's the context I'm responding to

I agree we have no idea on how the R1 will be

[drowsy robin]

Conversations don't just stay on one page.

Lol

But 👍

Yes.

[brisk lark]

All good man, I'm done with this thread, I get your statement. 😉

Regarding my question I asked for help, I found this - https://www.youtube.com/watch?v=jCr9tr-iX_0

There does not appear to be a /describe feature in SD itself.

[trim magnet]

there is one,in automatic1111 webui > img2img tab > interrogate clip

[brisk lark]

Yeah the video mentioned this for AI generated images with metadata embedded in it, thx for the reply.

There is also this https://huggingface.co/spaces/pharmapsychotic/CLIP-Interrogator

[dire sleet]

what do I need to download from here to be able to use this inpaint model? there isn't a model in the root of the file section

https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1/tree/main

and my understanding is that the encoder/unet/vae are all parts of a model process so I don't think one of those alone are what I need

like as an example the 1.5 non-XL has a ckpt right in the root

https://huggingface.co/runwayml/stable-diffusion-inpainting/tree/main

nm I guess they aren't offered in that format from that repo but someone converted it elsewhere

[brisk lark]

Ok bro, they still charge $200 for a year. That's all I meant. 😎

[compact charm]

hey everyone, new here

i just tried to download stable diffussion

but i never seem to get it right, it doesnt work

can anyone help me?

[brisk lark]

This helped me today - https://youtube.com/watch?v=Z6E41eXStsU&si=GQMM_Xhv2QPtI8A1

[compact charm]

i just downloaded the program but it makes me weird pixelated noise instead of pictures

anyone knows why?

[warm junco]

Hey, can you make a screenshot with the txt2img settings and the output?
Best would be in #🤝｜tech-support

[warm junco]

You should follow my install guide in the pinned messages of the #🤝｜tech-support channel

[quartz night]

night shade release here? https://nightshade.cs.uchicago.edu/

[keen parcel]

When can we expect SDXXL?

[hearty gyro]

So I have a question about renting graphics card/subscription models. I am running SD 1.5 pretty okaish (but very slow) locally, but I want to try out SDXL and AnimateDiff. I am looking for a website that ideally can be used for A1111 and Krita and ideally I don't need to set up /download everything with every launch. I don't mind paying for the service, if it is reliable and customizable (and allows NSFW 🙂 )

[covert beacon]

If I host Stable Diffusion WebUI locally, can it be used for Commercial Purposes?

[warm junco]

Hey, what's your GPU?

[hearty gyro]

GTX 1660

[warm junco]

Ah okay, do you have --xformers --medvram --no-half in your webui-user.bat?

[hearty gyro]

Yes I do, the 6 gb VRAM is saving my ass here

[warm junco]

True. But sdxl isnt fun on that one

[hearty gyro]

Exactly, I tried SDXL a few times out, but even a simple picture takes over an hour to render and that is without ADetailer and ControlNet

[warm junco]

I dont think there is a Webservice that you can combine with Programms on the PC like Krita.
But there are many GPU renting services and auto1111 hosts.
Google collab, Paperspace, Runpod.io, Vast.ai,
or you use websites like Leonardo ai

[noble bramble]

any 1 knows where i can get more stable diffusion modules from?

[warm junco]

Modules or models?

For models checkout Civitai.com,
its the largest database

[noble bramble]

thanks man that's perfect

[warm junco]

Np, also noteworthy to say. Models are always 2gb or bigger.
Only these files go into models/stable-diffusion folder

[hearty gyro]

Yeah I going to try out Colab Pro soon, but I am a bit confused, what they mean with compute units, like the 10 bucks tier has 100 compute units. I am just worried that when I start to make Idk. 10 pictures and one video, suddenly all my units are gone

[ivory nexus]

This share link expires in 72 hours. For free permanent hosting and GPU upgrades, run gradio deploy from Terminal to deploy to Spaces (https://huggingface.co/spaces)
How to "Gradio deploy"

[fervent thunder]

What about SDXXXL

[fervent thunder]

Ive tried colab pro, its a scam

Even with pro many times i couldnt run SD, also they are purposely refusing to run it or lowering performance in many cases for SD

Open source is too dangerous for them -_-

Instead i suggest running SD by renting gpus. I use runpod but there are many other sites for this too

[hearty gyro]

Thanks for the insights, I honestly just need a reliable service in which I can set up A1111 with my own Loras and if possible something easily configurable, so I don't need to set up everything with every launch ideally

[crystal thunder]

guys im trying to use a SDXL model and a SD 1.5 model combined but its just turning everything yellow. i tried the refiner extension but its not activating

ive tried using the SDXL as base and 1.5 as refiner and vice versa, one gives really garbage results and the other just turns it all yellow

the plugin says "It's Base model, use Refiner, extension disabled!"

https://github.com/wcde/sd-webui-refiner this one

ive tried swapping VAEs, ive tried all different sorts of configs. ive tried taking out every word of my prompt. im running out of stuff to try

why does the refiner extension think i dont have a refiner model?

[warm junco]

You can't use 1.5 to refine sdxl and sdxl to refine 1.5 models

These are different bases

They also need their own vae

[crystal thunder]

yes you can, there was even functionality for it months ago

[crystal thunder]

https://github.com/vladmandic/automatic/discussions/1687 look

[warm junco]

The refiner extension you used is old and not updated.
Auto1111 by default includes the refiner mode

[crystal thunder]

also i use another program called fooocus and it uses SDXL as a base with 1.5 as a refiner and it works great. im just having trouble getting it working here

[warm junco]

Where is here?
The link you mentioned is for SD.next

Not for auto1111

Fooocus dont use 1.5 models as refiners as far as I know, as they only let you use sdxl models

[crystal thunder]

on automatic1111

[crystal thunder]

ive been doing it for over a week i can send you a screenshot

[crystal thunder]

i cant send screenshots here but im using an SDXL checkpoint called realisticStockPhoto with the cyberrealistic checkpoint thats 1.5

i really want to get it working in the other program too. im just trying to duplicate my results from fooocus to automatic1111

i know it can be done cause clearly im doing it

[warm junco]

Can you send it in #🏞｜general-with-images ?

[crystal thunder]

yes

[terse plume]

the collab it could be good ?

[bleak matrix]

Good morning, everyone!

How are we all today?

[crystal thunder]

could be better lol. all my images are coming out yellow

[bleak matrix]

Oh no! Well, you could say it's full of morning sunshine, hahaha!

What seems to be the issue?

I see CS has ya covered!

[crystal thunder]

lmao! well im not sure. i dont think its my VAEs. see im trying to reproduce my results that I had in fooocus. im using automatic1111 and using an SDXL base model with a SD1.5 refiner. it works PERFECTLY in fooocus. in automatic1111 as soon as the refiner kicks in everything turns yellow

i liked the results so much i NEED them back in automatic1111 haha. im only using it for the openpose

[inland wigeon]

morning @bleak matrix

@crystal thunder that happened to me yesterday

[crystal thunder]

i posted a few examples #🏞｜general-with-images message here and here #🏞｜general-with-images message i flipped the base model to 1.5 and refiner to SDXL and its actually worse. at least the other way around i only get 1 frame of yellow and then the picture absorbs it into the shirt lol

[crystal thunder]

did you manage to fix it?

[inland wigeon]

@crystal thunder in the sense that my current level is - twiddle knobs until it looks different- yes 🙂

lots of knobs to twiddle

looking forward to my weekend, to sit down and try to start making some sense of all this...

[crystal thunder]

you had everything turn yellow and blue when the refiner kicks in, but you managed to fix it? what did you do? do you remember?

[inland wigeon]

funny you mention blue, that's exactly what happened. as i changed things everything was blue one round, yellow the next, etc. unfortunately i don't understand it well enough yet to know (nor do i remember) what fixed it

[crystal thunder]

dammit lol well that means it can be fixed though!

[inland wigeon]

lol 👍

[bleak matrix]

Good morning! How are you?

[deep kraken]

How long does it take the image to generate and does it give an error message if it’s not going to create the image?

[fervent thunder]

Thats what everyone wants but its not easy. There is the serverless option in runpod, which i havent tried yet, but the idea is that you are charged by second for gpu usage (VERY COST EFFICIENT) and for the storage all the time. But you need some technical knowledge to set it up

[warm junco]

Depends on your GPU

And if it crashes there will be no image yes

[terse plume]

what to do if the Loras doesnt appears?

[steel dome]

I created a new Instagram account where I post only AI generated imagery with a specific theme (Street Photography + Robots), can I share it here or is it considered spam?

[hearty gyro]

I am def. willing to learn and expand my horizon, I got pretty good, pretty fast with SD 1.5 and running it locally, so setting up a server based solution with a good enough tutorial shouldn't be much of a problem...I hope ^^

[warm junco]

Switch the model to 1.5 then it will show all 1.5 loras

The same for sdxl

[terse plume]

yes that's what it was,thanks

[bleak drift]

Is there a place where you can put in a request for someone to make a lora?

[north pelican]

Hello, what version of SD is this? I would like to access sdxl, how do I do that?

[haughty wasp]

Guys, can anyone tell me any good service like Replicate? So we dont need to worry about infrastructure...

[earnest blaze]

SD is consistently bad at understanding a little intricacy in the prompt. Dall-E seems to be much better. I tried creating a scene where a girl is walking in the rain except the rain drops are falling cats. and it just won't do it.

[wind ingot]

Don't wanna be that guy, but then: why are you wasting your time writing that in here instead of Dall-E Discdord? 🤔

[cedar salmon]

we are in the do it yourself isle

[twilit tapir]

hello

[warped helm]

Hey there, does anyone know of some SD/SDXL like models that are tuned for outputting Illustrator style made images? i.e. flat colours, solid lines, etc?

[floral dagger]

Hi! Are there weekly challenges?

I want to participate every week to level up my skills and see what others do with the same concept/idea

[forest trout]

I would focus on #🔆｜dailies for now.

[inland wigeon]

any good youtubers you all would recommend for starting out with this stuff, that you'd like to send a new subscriber to?

[earnest blaze]

@wind ingot I am wondering if anyone knows how to get SD to replicate Dall-E's results. That's why.

[foggy zodiac]

Hi there. Anyone know a good place to report people trying to sell stable diffusion art on patreon etc?

[inland wigeon]

is it illegal to do that? crappy, definitely, but illegal?

[foggy zodiac]

Yes actually. Goes against their terms

All art is for non commercial use.

Sale is commercial.

[inland wigeon]

ahh, good to know

[trim magnet]

not regulated yet so you can still do it

[foggy zodiac]

Only reason I'm even bothering is because the person behind it was super sketchy and creepy af in an actual art server I run for people to commission hand drawn or digitally drawn art.
Just being weird and creepy to people, and also trying to sell their stuff.

[wind ingot]

That's fair, but you won't find many Dall-E people here, it sure takes some using but SD best part it's the flexibility and community, not ease of use, I'm sure if you ask many people will try to help you achieve what you need

As long as you don't try to piss on their lawn 😁

[vapid tangle]

Anyone know how to take zoomed out images with SD? Say a bear, I can't get a photo taken from 50 feet away. it always makes a close up photo

[pale latch]

describe the scene and then add on "a bear in the distance"

[warm junco]

Hey, you can ask in #📝｜prompting-help for good tags or how to get a specific image

[terse plume]

somebody could tell me a good anime model? because i used the yesmix_v15, this version doesnt exist anymore in the civitai.

[somber inlet]

Can anyone help me with realistic skin texture?

[scarlet jolt]

Anyone know how I can find my past images easily? When I run my name or keywords from my prompts I don't get any results

[sage sky]

can someone please help me?

updating Stable diffusion ruined my SD

[obsidian adder]

Cinematic is hosted

*goated

[edgy ridge]

heyy yall, anyone know how to generate images with two distinct characters? is there a lora or something i can get?

[crystal thunder]

yes i just read an article actually about that

[edgy ridge]

ohh can you send me the link to it

[crystal thunder]

what youre looking for is the regional prompter https://stable-diffusion-art.com/regional-prompter/#Installing_Regional_Prompter_extension

[crystal thunder]

i just did, its a great read check it out

[edgy ridge]

thats perfect, thank you :))

[crystal thunder]

no problem, it also helps seperate color bleed if you're having trouble with that as well as being able to perfectly setup your composition 🙂

[edgy ridge]

uhh is there a way to install this on easy diffusion? im not using the automatic1111 webui :(

[crystal thunder]

uhhhh im not sure, it looks like they call their extensions plugins, and theres not really a lot of them

i dont know if it would be compatible

i would highly recommend taking the time to learn automatic1111, you can get the same results with way more control

and easy diffusion doesnt have a lot of devs so progess is slow

[jade bear]

Yoooo

[crystal thunder]

i was hitting my head against this wall too cause i wanted to pose my characters but couldnt without control net. i just finished learning it and i have to say i regret not doing it sooner

[edgy ridge]

ill look into automatic1111, thank u for the help <33

[crystal thunder]

no problem 🙂

[jade bear]

What’s the difference between regular SD and SDXL?

And should I upgrade

[shell wasp]

I am getting an " application not responding " error . What could be the reason.

[warm junco]

For an auto1111 install guide checkout the pinned messages of #🤝｜tech-support

[warm junco]

Do you have a screenshot of that error? What does the cmd shows?

[warm junco]

SDXL are models trained on images that are 1024x1024 in resolution and have a better word understanding.
If you use a local webui to generate images. You can use SD 1.5, 2.1 and SDXL models

[crystal thunder]

are there any good prompts for selfies? ive tried 'selfie shot, taking a selfie' but the phone always comes up in her hand like shes doing a mirror pic

lol turns out i just had to remove the prompt for the shoes

[daring hearth]

Hey, could someone point me towards some good guides to get started as a newbie?

[crystal thunder]

is there an easy way to calculate regions for regional prompting? maybe somewhere i could just paint them on?

[placid hatch]

Debating between an RTX 4070 and RX 7800 XT. Is the RX 7800 XT worthwhile to pickup if you are running it on rocm on linux?

[delicate marsh]

is anyone facing issues with stable-code

for me it generates random code when i ask something

is there any java specific model in works

[copper shell]

follow on insta snaptastic_zesties

[warm junco]

yes, on linux it can use rocm and then the speed is very good, on windows its usable too, but slower. also AMD works to get Rocm to windows so that will happen this year (maybe in the next months)

[stable pawn]

I have many images I made locally and online, most of them contain hidden information (used to save prompts aso) but there is no standard, this is not EXIF or IPTC nor JPEG Comment, but I sat that Bing generated images was saved so EXIF and IPTC could be identified they where empty when viewed in IrFanView.
So my question is if there exist any app that can scan images and identify the prompt info and then extract it from images made with (aa example) Autimatic1111, Invoke and ComfyUI to make it easier to copy and paste and find what Checkpoint that was used?

[warm junco]

all the information is stored in the metadata (exif) and yes, automatic1111 itself has that tool inbuild. its the PNG-Info tab. where you put in the image and will get the information from it. That only works if the uploaded platform or the user didnt deletet the Metadata before.

Also a tool for getting the information is the Exiftool.
it can be included to the ImageGlass viewer

[inland egret]

Hi guys good morning! I am writing to possibly ask for your help on a project I have in mind. I have been trying for days and days but unfortunately I still haven't succeeded and I would like to ask for your help.
In stable diffusion, I would like to try to represent like a 2D black silhouette in the real world, interacting like normal people,but I get the picture always in black and white or with light reflections that make you understand that it's actually a real person (like on the nose, vesiti, feet etc..). Would you possibly have a solution? I would really appreciate it.

[terse plume]

2D?

[inland egret]

yes

I just sent an example in general with images

[terse plume]

intresting

[somber inlet]

Does anyone have luck with ai hair and skin looking less plastic?

[terse plume]

i put a image on general image

[haughty bolt]

hey, how can i make the haidilao dance using stable diffusion?

[terse plume]

what is a haidilao?

[inland egret]

I appreciate the help, but I need the silhouette in a normal environment, interacting like normal people, without any particular light behind it

[terse plume]

but do you need make a intire serie of black silhouette or just some image?

[inland egret]

a intire serie

[terse plume]

a put another image

[terse plume]

tottally black i couldnt make

[terse plume]

but do you know you can put this Lora and trying to make

[inland egret]

I would appriciate that!

[terse plume]

Lora: scene silhouette Fun

NSFW scene silhouette fun

[inland egret]

I'll try that! Thank you

[woven berry]

Anyone know if you can batch process a bunch of images using controlnet on the txt2img tab? I wanna do a style transfer but my results on img2img aren’t as good. I get exactly what I want on txt2img when I use one frame of my video in controlnet. I just want to batch process through all 150 frames on the txt2img tab…

[bleak drift]

Are pruned stable diffusion models lower quality than a full model?

[wind ingot]

I believe, but not 100% sure, you could use chaiNNer to drive A1111 in batch, just not sure how to store a serie of different prompts , maybe on a text file? You could try asking in ChaiNNer Discord

I know it can batch just not sure in case you need different prompts for each

[woven berry]

I’m just using the same prompt, and doing a style transfer to each frame of video in my image sequence

[wind ingot]

Then probably ChaiNNer is what you need

It can load a video as frames

Then query A1111

If it's to run other tasks like running PyTlrch models to clean/upscale video you can use the models there with no need for A1111

[woven berry]

I’ve already split up my video into images, just need control net to function in batch

[warm junco]

then use it in img2img, with a low denois

[wind ingot]

But with different images?

Or same image every time?

[woven berry]

Meh img2img with low demos is keeps too much of my geometry

[wind ingot]

Lower controller weight?

[woven berry]

Weirdly (and I’m missing SDXL) I had to crank the settings up in order to get close to what I created in txt2img. Denotes and cfg scale has to be kinda high

[hallow plaza]

I have an image I want to remove text from. whats the best way to do that?

[woven berry]

Both at 9 and .9

[warm junco]

then use higher denois and lower controlnet weight

0.2 or 0.3 or even 0.4

for denois

[woven berry]

I did. That how I got it to be close to what I made in txt2img

I just wish I could batch on text2img using controlnet

[wind ingot]

Without knowing exactly what you're trying to accomplish: you can add more ckntrolnet models: depth, canny and IP adapter, you can even add more than 3 if you go to A1111 preferences

[wind ingot]

You can....using ChaiNNer

[woven berry]

Cool

I’m going to look this up

Thx

[bleak drift]

What is the best "Abyss" stable diffusion model?

Because I notice there are alot of them on civitai

[barren crag]

Hi all! I'm trying to generate cartoon styled ui icons for a pc game, but from my research it seems like this is a task that stable diffusion is not great at. Does anyone have prompt or alternative model suggestions?

[oak yoke]

guys i need help. where can i subscribe for dream ai??

[velvet fossil]

I'm back guys

where's Javi Lopez?

there's something I want to know

[storm bolt]

Heyy

[solemn sky]

anything in the works that's better than SDXL? i've kind of stayed on 1.5 cuz the models using 1.5 have been so finetuned over the years i actually get better images with them than SDXL.

i have a custom 1.5 merge which i don't even need negative prompts for.

[hallow plaza]

I have an image that I like, is there an AI that will design the words for a book cover on top without changing the picture?

[fervent thunder]

Looking for an editor who can switch faces

[fiery fox]

how much you pay?

[fervent thunder]

5euros

[fiery fox]

guys when i activate controlnet and use reference and wanna render 5 pictures, it talkes almost 1 hour to render. is this normal? When i just render 5 normal pictures without controlnet (reference), it takes like 5 minutes???

[fiery fox]

you need it just once for 1 picture or long-term cooperation?

[fervent thunder]

1

At the moment, no idea about the future

[warm junco]

Reactor for auto1111
Or faceswaplab

Or the tool Facefusion

[fervent thunder]

Which one

I don t know it

@warm junco i don t have pc bro can u help me

[warm junco]

Oh then I dont know which service that offers.

Maybe you need a cloud solution for auto1111

[fervent thunder]

U cannot do it for me please? Will pay bro

[fiery fox]

AMG Mercedes, but no PC.. God damn bro😅

[fervent thunder]

I broke the monitor

Yesterday

💀

[fiery fox]

I do that too sometimes when i play valorant. But often i break my keyboard. just got a new one this week xD

[shadow oak]

hey guys, i'm trying to install stable diffusion but when i run the WebUiuser.bat file i get this error:
--index-url https://download.pytorch.org/whl/cu118
Defaulting to user installation because normal site-packages is not writeable
Looking in indexes: https://download.pytorch.org/whl/cu118
ERROR: Could not find a version that satisfies the requirement torch (from versions: none)
ERROR: No matching distribution found for torch

[fervent thunder]

Damn

[shadow oak]

any ideas??

[fervent thunder]

Idk bro i m tryna do a faceswap

[fiery fox]

how fast you need the picture?

today?

[shadow oak]

me?

[fiery fox]

nah catalin

[warm junco]

Yes, you have installed the wrong python version

You should follow my install guide in the pinned messages of #🤝｜tech-support

[shadow oak]

mhh which one should i install? and should i unistall older versions?

oh sorry i didnt know it

let me check it

[fiery fox]

I'm still using Version 1. something, but i know like 3 months ago version 2. something dropped. Is the upgrade worth it? I'm just scared there is 0 improvement and i will just run into bugs

[warm junco]

version of what?

[fiery fox]

A1111

no idea where i can see my current version

[warm junco]

1.7.0 is the latest version

you can see that at the bottom of the webui in browser

[fiery fox]

oh i see.. much appreciated. i have 1.5.1
So is it worth to update it to versiion 2. whatever it's the newest

[warm junco]

i would always go for the latest version, because it has bugs fixed and better compatibility with extensions etc
But you dont need to. You can also read the Changelogs since 1.5.1 and check if there is something you miss in 1.5.1
https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases

For example its also recommended to use 1.6.0 or 1.7.0 if you want to use SDXL

[warm junco]

but whats also important is your torch and xformers version. check it also at the bottom of your webui. it should be 2.0.1 and xfoemrs 0.0.20

[pale latch]

oh shit my xformers is 0.0.23 what do i do??

[wind ingot]

I would argue it should be 2.1.2 it's the latest stable, and if you know how to change then you would know how to set up a temp venv and try a nightly build on 2.2

[pale latch]

[wind ingot]

The performance differences are quite big

Sometimes nightly builds screw up thought

[pale latch]

are they? i didn't notice any big performance changes from 2 to 2.1. i got fp8 enabled now, which is a little slower but is a trade for bigger batches

a lot slower on older than ADA hardware i hear

[warm junco]

how do you test fp8 if there are no models for it?

[wind ingot]

2.0 to 2.2dev on Mac was like 20%

[pale latch]

autocasting i think is how it's done. it's just a switch in the auto1111 settings

[wind ingot]

There is printed Turbo

*prunned

[fiery fox]

it is. version: v1.5.1  •  python: 3.10.6  •  torch: 2.0.1+cu118  •  xformers: N/A  •  gradio: 3.32.0

[pale latch]

i think there are some fp8 models too though. made for phones

[warm junco]

whats your GPU?

[pale latch]

i'm probably wrong i have no idea about the mobile models

[fiery fox]

RTX 3060

[wind ingot]

Seu, misread, prunned turbo is fp16

[warm junco]

okay good xD i was a bit confused

but yea i dont think auto1111 can use fp8 for fp16 models, but im not sure

[solemn sky]

it can

[warm junco]

fp8 models should be even smaller and faster than fp16

[solemn sky]

fp16 is the default when using cuda

what's the problem?

[warm junco]

we are talking about fp8

iceycold said he has fp8 support enabled in auto1111 but didnt saw a performance increase

[fiery fox]

up

[solemn sky]

fp16/8 doesn't seem to increase speed but rather decrease vram usage. ofc, this might indirectly affect performance.

[warm junco]

then your torch version is okay, but if you want to use xformers you have to edit the webui-user.bat and there add --xformers to the Commandline_args=

[warm junco]

will that change when we have fp8 models to test?

[solemn sky]

i haven't seen any fp8 models around so can't say 😛

[warm junco]

me neither xD

[solemn sky]

if i want to increase speed i'd rather look at LCM models

could also try using SDP instead of Xformers

--opt-sdp-attention --no-half-vae --upcast-sampling --autolaunch
RTX 3060 12GB - Automatic1111 v1.7.0 | Python 3.10.9 | Torch 2.0.1+cu118 | Xformers N/A

could also install the TensorRT package but that's a sea of trouble i haven't delved into yet.

[warm junco]

i would use xformers over opt sdp, because xformers uses slightly less vram, also you normaly dont need --upcast-sampling as you alread have --no-half-vae

tensorRT is fast but not compatible to all auto1111 features

[solemn sky]

oh, that's SDNext using it then?

[warm junco]

dont know if sdnext supports it

[solemn sky]

anyway, i use the no-half-vae to avoid blank images

apparently it was the cause.

[warm junco]

yes using --no-half-vae is good, but --upcast-amspling is just a different technique for it, no need to use both

[solemn sky]

oh that's nice, didn't know.

[midnight vigil]

Hello

I am trying to download this https://huggingface.co/SG161222/Realistic_Vision_V6.0_B1_noVAE/tree/main

which version should I use

I have 12 gb of vram

also should I use VAE

[solemn sky]

if you go to Civitai you can preview the results.

[warm junco]

the 2gb (fp16) one for generating images, the inpaint versions are for img2img

[warm junco]

yes, you should use the 84000-mse vae for that model

[midnight vigil]

I would need to switch between versions with the images and inpaint?

[warm junco]

yes for inpaint use the inpaint model

but you can also inpaint without it too

[midnight vigil]

I see there are two versions for this ckpt and safetensors

https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main

which one should I use does it matter

also how do I install the VAE

[solemn sky]

always safetensors

[midnight vigil]

ok

[solemn sky]

ckpt models can contain bad code you don't want

called "pickles"

[midnight vigil]

yes

the pickles looked weird

I am familiar with safetensors bc of ooga booga web ui

[solemn sky]

download and use a pickle scanner before using ckpt models.

[midnight vigil]

can you help me install VAE I dont want to mess up my install

right now I saw this tutorial online Download the ft-MSE autoencoder via the link above. Copy it to your models\Stable-diffusion folder and rename it to match your 1.5 model name but with ".vae.pt" at the end. In my example:

but

its a safetensor

[warm junco]

put the vae into models/vae

[midnight vigil]

not a pickle so how would I rename it

[warm junco]

dont rename it thats and old guide

[midnight vigil]

ok

just simply put it in the vae

[warm junco]

@midnight vigil then go into settings -> User Interface -> Quicksettings, there add sd_vae then hit apply and restart ui

[solemn sky]

\stable-diffusion-webui\models\VAE
is the location

[midnight vigil]

I dont have a quicksettings option under user interface only these

Gallery
Infotext
Live previews
Prompt editing
Settings in UI
UI alternatives
User interface

[warm junco]

user interface

then quick settings

[midnight vigil]

ok I see it

this is how I would manage multiple VAEs as well?

just replace it

[solemn sky]

yes, gives you a dropdown menu at the top.

[midnight vigil]

ok

[warm junco]

dont replace it, just add sd_vae and then you hit apply to get the dropdown on next reload

[midnight vigil]

ok

[solemn sky]

while in there, you might also add > sd_lora and CLIP_stop_at_last_layers

[midnight vigil]

yes I wanted to get into loras

is that possible on my machine I have 12gb of vram

on an nvidia rtx 4070

[solemn sky]

ofc

[warm junco]

yes thats very good for sd stuff

[midnight vigil]

any good tutorials?

[solemn sky]

get a good negative embedding

i suggest once again you use civitai instead of huggingface cuz civit gives you previews of everything, and, it's more popular for image models.
huggingface is a good search for LLM's but not for images.

some models also lets you generate images on-site without even downloading the model.

[midnight vigil]

what do you mean images?

image generation?

generators

wait I see

[solemn sky]

go to civitai and see for yourself, it'll unlock a new universe for you 🙂

[midnight vigil]

I tried it before

can you help me with this last thing

I did the automatic installation through the zip of a1000

and this is the response upon launching

Launching Web UI with arguments:
no module 'xformers'. Processing without...
No SDP backend available, likely because you are running in pytorch versions < 2.0. In fact, you are using PyTorch 1.13.1+cu117. You might want to consider upgrading.
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.

It says I dont have xformers

[solemn sky]

open "webui-user.bat" with Edit, or open it with a regular .txt editor

in the COMMANDLINE-ARGS, add --xformers

or, upgrade your pytorch and use SDP instead of xformers.

[midnight vigil]

which one do you recommend

I dont want to break my SD but I also want to have the most optimized

[solemn sky]

i always use SDP, can't speak for xformers.

[warm junco]

oh pls dont install it via the zip file

i have a made a very good install guide in the Pinned Messages of the #🤝｜tech-support channel

[midnight vigil]

I have issues when installing via clone

[primal nexus]

Hello! I need help with something... I made a video with deforum on automatic1111, then I batchprocessed the pngs of that video with adetailer to have better faces. now i have 2 questions, 1st: if i wanted to create a video from them...(ffmpeg doesnt work for some reason since the filenames have 2 number increments in them and ffmpeg refuses to accept that and i dont want to rename all the 200+ images) can i feed those images into a comfyui workflow to create a mp4 video with them?
And more importantly question 2: can i use those few hundred pngs and further process them in comfyui to make the video less flickery and have less changing details, maybe interpolate them or smooth them out somehow? i cant feed them into a ksampler node right?

can someone with experience maybe help me out a bit?

[warm junco]

i would install the Powertoys, it features a powerfull renaming tool, i use it for renaming hundreds of images so that ffmep and rife (interpolation) works on it

[warm junco]

you should take a look at my install guide, or ask in #🤝｜tech-support for help on git clone

[midnight vigil]

well

[warm junco]

because with the zip you wont be able to update the webui

[midnight vigil]

the issue is

I already have python installed

[primal nexus]

thanks alot! will look into it

[midnight vigil]

an updated version

because anaconda puts 3.12.1 as sys python

and I cant change it

so I need to use a venv to install it

[warm junco]

you can change that easily, by editing the webui-user.bat at the line Python=
point it to the right python.exe
then you need to delete the venv folder

[midnight vigil]

but I dont know how because no tutorials say so

ok

i dont understand

the venv is only used to install

[warm junco]

for any install help feel free to ask in #🤝｜tech-support

[midnight vigil]

so the python verion doesnt matter after that

[warm junco]

no the venv has all the core files needed for SD to work

[midnight vigil]

why would I delete the venv folder then

[warm junco]

it will be recreated with the right python version, after you changed the python path in the webui-user.bat

[solemn sky]

cuz a1111 installs its own venv folder using whatever python currently installed at the time.
if you don't direct it to python 3.10.6 or .9 it will install with your python 3.12.1

[primal nexus]

do you also know if I can batchprocess those images in comfyui, maybe with animatediff and/or controlnet or something else for 1) better smoother animation, 2) upscaling

[warm junco]

not for comfyui, im doing it in auto1111 currently, but for comfyui you need the img2img batch process

smoother animation isnt available in any webui currently

[primal nexus]

automatic1111 is so slow sadly.. 😦

[warm junco]

whats your gpu?

[primal nexus]

4080

[warm junco]

then you should edit the webui.user.bat and at the line Commandline_ARGS=
add: --xformers --no-half-vae
that should make it faster

[primal nexus]

i already did that

[solemn sky]

trust me, if you haven't used CPU for generating you haven't experienced slow yet.

[warm junco]

then auto1111 shouldnt be slow at all

but i know comfyui is still faster

[primal nexus]

oh wait

i only have --xformers --api

should i delete the --api and add -no-half-vae?

[warm junco]

thats okay too, no-half-vae wont increase speed, its for compatibility

but yea remove --api if you dont need it

[primal nexus]

i dont even know what that is lol, i just added it bc it was mentioned in some video to increase speed

[warm junco]

lol it wont xD

[solemn sky]

what a fishy guide

[warm junco]

--api is a bridge for external programms (photoshop, Krita) or some extensions like OpenOutpaint

[primal nexus]

ok good to know thanks

im using gimp right now but im manually saving and loading images to/from there

so the best way to make a mp4 out of a bunch of pngs is still ffmpeg right?

[warm junco]

then take a look at that, a really cool standalone krita tool to generate images inside it:
https://github.com/Acly/krita-ai-diffusion

more powerfull than gimp, also opensource

[primal nexus]

yeah i have krita installed already, but not using it much rn

i dont even know why, i dont know much about those paint tools

i usually used paint for my stuff lol

[warm junco]

yes, ffmpeg, ive made me some .bat files to auto convert images from a folder
also a bat to interpolate

[primal nexus]

thats cool

so im installing powertoys now and trying the renaming thing

thanks again for all the help!

[warm junco]

no problem here is a help for powerrename:
Indicator: .* Replace with: ${padding=8}

[primal nexus]

how to delete the whole filename and replace it just witha number?

it just adds the number now i have 3 increments in it 😄 😄

the apply button is greyed out now hmm

padding=8 means it fills it up with zeroes up to 8 digits rights?

[warm junco]

yes and counts upwards

[primal nexus]

for some reason its broken now, i cant hit apply anymore 😦

ahh i got it, i had to check use regular expressions

ty!

[warm junco]

yea its a bit "try around"

[tribal pasture]

has anyone been having issues with controlnet?

since the last update it wont do what json i send it

[spice plank]

Is there someway to sort by date THEN sort by model used?

[carmine herald]

Whats the deal with some embeddings being .safetensors? It seems like a1111 doesnt recognize those as embeddings when i put them in the folder too

[warm junco]

do you have an example of that? you can post in #🤝｜tech-support

[warm junco]

make sure to update your webui

[tribal pasture]

i seem to got it working

[spice plank]

To anyone who wants to also know the answer to this, the setting you are looking for is in Settings > saving to directory > Directory name pattern > add [model_name] to the input bar.

[tribal pasture]

seems like an update fixed it

[carmine herald]

its updated

ok well now im not getting anything showing up in the textual inversion tab, i didnt even change anything 💀

[warm junco]

make sure the webui is whitelisted in any adblocker

also select an 1.5 model as checkpoint to see the 1.5 embeddings

[carmine herald]

this was it, i didnt know it hid embeddings for different models, thank

[haughty patio]

is there any channel for very beginner/noob questions ?

why my image appear in #1100484581037195384 its automatic?

[carmine herald]

Another question, what's the deal with the turbo model? I'm running it at 10 steps at 1024x1024 and it takes 5 minute to process. With other models it takes only 30 seconds or so to do the same size. using dreamshaperXL_turbo if that matters

[carmine herald]

looks like this one
I guess #🤝｜tech-support also works

[wind ingot]

Turbo runs on 512px

[carmine herald]

thats interesting

why is it called xl then

[wind ingot]

Because supposedly runs and responds to same prompts and contolnets

[haughty patio]

ist there any list of best prompt search, prompt databases like lexica website?
whats your favorite when you looking inspiration
or when you looking to learn new styles

[wind ingot]

I get around 1 image per second in Turbo,on a Mac

[carmine herald]

how many steps?

[wind ingot]

Around 5

[carmine herald]

i just look at what people put in prompts on cites like civitai and places like here

[wind ingot]

Turbo needs CFG around 1

[carmine herald]

so odd

[wind ingot]

Yes, it's very different but once you realize that, gets very simple, but it's very sensitive to sampling methods and mostly steps

5 and 10 steps get you very different outcomes

Start with CFG of 1

[carmine herald]

alright, ill play around with it

thank you for your help

[wind ingot]

Np

[midnight vigil]

?

[crystal thunder]

guys, im trying to make it so that my prompt can have multiple backgrounds, when i do it like this "{at the beach|at home}" the background comes out super distorted and blurry. same with jeans or anything else i put brackets in. how do i make multiple options like that without it killing the quality?

[spice plank]

My bad, forgot to switch discord channels, thought I was in a different one

[crystal thunder]

i could really use some help with this, i provided some images as examples here #🏞｜general-with-images message

[fading prism]

Why does the format automatically change to video? Is this a glitch or do I have to add the format to the image every time I write a prompt?

[knotty turtle]

Has anyone developed a system where Stable Diffusion interfaces with Discord to remotely receive prompts, generate images, and automatically post them back in Discord, similar to Midjourney?

[terse plume]

you can put "at home, a good vision of the beach"

[terse plume]

the bots

[terse plume]

bot-1,

[terse plume]

but its different of stable difusion,

[midnight vigil]

hello

[terse plume]

stable difusion you can add Lora,model,

[midnight vigil]

can anyone help me after installing both control net and webUIautocomplete i get these warnings

• "styles.csv" missing
• Tag Autocomplete issue

venv "C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\venv\Scripts\Python.exe"
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug  1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.7.0
Commit hash: cf2772fab0af5573da775e7437e6acdca424f26e
Launching Web UI with arguments: --xformers --no-half-vae
Style database not found: C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\styles.csv
Tag Autocomplete: Could not locate model-keyword extension, Lora trigger word completion will be limited to those added through the extra networks menu.
ControlNet preprocessor location: C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\extensions\sd-webui-controlnet\annotator\downloads
2024-01-20 19:20:58,619 - ControlNet - INFO - ControlNet v1.1.432
2024-01-20 19:20:58,683 - ControlNet - INFO - ControlNet v1.1.432
Loading weights [15012c538f] from C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\models\Stable-diffusion\realisticVisionV60B1_v51VAE.safetensors
2024-01-20 19:20:58,991 - ControlNet - INFO - ControlNet UI callback registered.
Creating model from config: C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\configs\v1-inference.yaml
Running on local URL:  http://127.0.0.1:7860

mr fernando please

[terse plume]

what are you trying to do?

i dont get it

[midnight vigil]

I have these issues after installing both control net and webUIautocomplete

extensions

@terse plume

fernando

???

[terse plume]

do you have space in the driver to install this?

[midnight vigil]

yea

[terse plume]

sometimes you have to try again

[midnight vigil]

ok

thank you fernando

[terse plume]

you are doing on collab?

[midnight vigil]

no on my computer

it works fine I just get two warnings

• Style database not found: C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\styles.csv
• Tag Autocomplete: Could not locate model-keyword extension, Lora trigger word completion will be limited to those added through the extra networks menu.

[terse plume]

ask these guy @exotic sphinx when he is on

he knows better about stable difusion intalation

[midnight vigil]

ok

thank you mr Fernando

[crystal thunder]

that doesnt work

that would make it so you're at home with a sight of the beach

i think i fixed it with dynamic prompts but it generates like 4 images with different combinations instead of it just being a random one each time

[midnight vigil]

Upon installing controlnet and sd_webui_autocomplete, I get two warnings when launching webui-user.bat

• Style database not found: C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\styles.csv
• Tag Autocomplete: Could not locate model-keyword extension, Lora trigger word completion will be limited to those added through the extra networks menu.

Entire output:

venv "C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\venv\Scripts\Python.exe"
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug  1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.7.0
Commit hash: cf2772fab0af5573da775e7437e6acdca424f26e
Launching Web UI with arguments: --xformers --no-half-vae
Style database not found: C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\styles.csv
Tag Autocomplete: Could not locate model-keyword extension, Lora trigger word completion will be limited to those added through the extra networks menu.
ControlNet preprocessor location: C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\extensions\sd-webui-controlnet\annotator\downloads
2024-01-20 19:20:58,619 - ControlNet - INFO - ControlNet v1.1.432
2024-01-20 19:20:58,683 - ControlNet - INFO - ControlNet v1.1.432
Loading weights [15012c538f] from C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\models\Stable-diffusion\realisticVisionV60B1_v51VAE.safetensors
2024-01-20 19:20:58,991 - ControlNet - INFO - ControlNet UI callback registered.
Creating model from config: C:\Users\USER\Programming\AI\Photo Generation\stable-diffusion-webui\configs\v1-inference.yaml
Running on local URL:  http://127.0.0.1:7860

can anyone help me

[karmic brook]

#🤝｜tech-support

[somber inlet]

anyone have luck with realistic skin and hair textures?

[tulip shadow]

Hello the Artists

[midnight vigil]

hello TheLegatus

[inland egret]

Hi guys, a bit of a special request for help. I can never generate such specific things. I should make a picture of a bald man lying on a white sofa watching TV. The problem is that: 1. It doesn't give me the right shot. 2. In the empty white and infinite room I would only like a TV and the sofa on which it is lying. The view should be 3/4. I appreciate any kind of suggestions

[wind ingot]

@inland egret have you tried sketch/scribble controller? You just need a rough drawing to guide it, use a low weight like 0.3 and have it stop about 0.3 to 0.5 so AI gets creative

[inland egret]

I suck at drawing, that's why I've never tried xD

[wind ingot]

You don't need a masterpiece, just a very very rough lines

[wind ingot]

Just draw a rough for the perspective you want

Or...find an image close and use canny

[unkempt depot]

Hi, I wanted to create youtube videos like this https://www.youtube.com/shorts/qGW3oQJT50o . Here chatgpt is used to create a base image and then the image keeps getting modified using commands. I am a stable diffusion noob, so I would love to know if I can also do this with stable diffusion, so can I make stable diffusion give me an image and then modify the base image based on what I say? If the answer is yes, then how can I do this? Help would be very much appreciated thanks 🙂

[solemn sky]

i have a couple of questions...
why did this server go from Stable Diffusion to Stable Foundation and then back to Stable Diffusion?
also, how come there's no rules against deepfakes and/or the use of real people? Unstable Diffusion has, and they actively ban anyone even attempting such a thing.

[hasty hornet]

probably cause unstable does nsfw and that's a bit different, then just doing funny images with Elon or something

[solemn sky]

hm, problem isn't just knotted around the idea of nsfw tho, it's that ppl/media of any rank or skill can spread missinformation or defamatory rumormills.

[hasty hornet]

They are doing it everyday, at least here people probably can recognise it.

like...idk how banning it here would help

[solemn sky]

it's to lessen the knowledge of deepfakes and the likes, the more difficult something is the less likely ppl are to spread damaging media.
idk about you but the normie on youtube does not recognize real from AI.

[hasty hornet]

normie watching tv or reading internet articles doesn't too, yet what does it have to do with banning AI images \ video with other real people on this discord?
How is that supposed to help said youtube , tv or overall media bs? They aren't gonna stop.

[solemn sky]

you don't catch my point... it's the thresh hold, making the step into deepfakes more tedious thus slowing or in best case scnario reduce the number of ppl doing it.

put yourself in one of the celebrities shoes.

[hasty hornet]

Yea, I don't catch your point, there are enough sources on the internet where people could learn how to do that if they wanted...
I also don't think that one additional step of "oh, people on this discord don't want to teach me, I guess I have to use google" - will stop a-holes from being a-holes lol

[solemn sky]

ofc, but officially saying no to actively help people do it would give this community a lesser toxic image, from the publics point of view.

[hasty hornet]

idk, I think we just need to learn how to live with deepfakes, instead of trying to "hide" them from publicity and not talk about them...

They aren't going anywhere

[solemn sky]

again, point is to raise the bar.

ppl like things simple, if it gets too tedious, some gives up.

[solid thicket]

where can i ask someone to create a image for me?

[rain torrent]

Stable difusion hate octopus, i've tried create a art to use on my RPG campaing, and i'm getting frustreted

[solid thicket]

how descriptive are you being?

[rain torrent]

dream prompt:medival, giant octpus, camuflage, in the deep ocean, Megalofobia, animal, no light, art by greg rutkowski negative_prompt:a exagerate number of tentacles, only 8 tentacles format:Image

bot 10 you can see my attenpts in the las 40 minutes

[solid thicket]

What in the things you dont want?

[rain torrent]

exagerate number of tentacles

[solid thicket]

Just keep trying man

[rain torrent]

That's life

[warm junco]

you can use the #1080261341362786384 to create images for example

[marble knot]

Do you guys know if we can capture mouvements with stable diffusion to create animations and export the animation as a bvh file? (in order to animate 3D models)

[stark sierra]

How do we join?

[fading crypt]

Anyone knows if dalle3 api has history enabled? Using it via my app, and I tell it do create iron man (which it does) but when I continue with create a dinosaur it creates iron man again

[placid pecan]

Is there just thing as checkpoint model for food?

[pale latch]

probably not since most base models do food sooo well already

though maybe. it is a huge community

https://civitai.com/search/models?sortBy=models_v5&query=food lots of loras it looks like

[placid pecan]

Hm, I just went to check there after asking here

[pale latch]

smart idea

you'll notice that many base model checkpoint example photos, there are a lot of portrait images. I think it's unfortunate tbh, since base models have the knowledge of 2billion images, not just portraits

[midnight vigil]

hello can I ask can I use sdxl in a1000

do u guys recommend that or not

I heard it just combines image generation with AI airbrushing

[nocturne nymph]

What's the difference between SD 2.0, SDXL, and SDXL Turbo?

Other than being different models. Is there some discussion on why you might use one over the other? What are their technical differences?

[slate dust]

SDXL is trained on larger resolution images and understands prompting more accurately than 1.5. Turbo is a version of XL that allows very fast generating at very low CFG. 2.1 is largely dropped by now I think 🤷

They can all be used in any UI

I think a lot of poeple still use 1.5 because it has a lot of checkpoints/Loras built for it and it's less resource intensive. XL is where it's at, imo, though

[midnight vigil]

really @slate dust

do you use sdxl in a111 or comfy ui?

[slate dust]

a111 and SwarmUI now

Which is a great gui with Comfy backend

Comfy is so much faster. I only go back to a111 if I need specific extension

[nocturne nymph]

Thanks!

[fervent thunder]

Hi all is there any image to cartoon converter bot or free software available? Please help me

[midnight vigil]

Ok yes SwarmUI seems very good and easy install from the docs

[solemn sky]

have you tried weights?

random noob input? this discord ain't for nsfw.

[midnight vigil]

who is familiar with coqui tts?

[placid pecan]

chat ded

[inland egret]

guys it's been 2 days already and I still haven't found a solution...please help me

[pale latch]

like you JUST want the black cutout shape ?

segment anything, color fill #000

https://civitai.com/models/86568/kate-or-shadows-house-lora this lora sort of looks like what you mean, in anime

[nimble shard]

hi trying to do video generator for product, for example i have an image of a shirt as input, the output would be a 1 minute video of a fashion model wearing that shirt, anybody has done that before

[young ivy]

Gigabyte GeForce RTX 4070 Super 12GB Eagle OC for 840 dollars or MSI GeForce RTX 4070 Super 12GB Ventus 2X OC for 775 dollars...or neither....or see if I can snag a 4070 Ti Super if one shows up under 900 dollars on the 24th

I don't think we'll be seeing series 5000 this year

probably fall next year

fuck so many people trying to add me as friend to...talk to me....like not even about anything but I get suspicious when someone starts talking to me like we're familiar but they don't tell me who they are or how I know them

[spring eagle]

where can I find people to hire for creating lora?

[tender cove]

Everyone that tries to add me waits a couple days, then tries to send me invites to their discord server. I just reject invites now unless it's someone I'm talking to in channel.

[fervent thunder]

how can i use lora with stability.ai? Api

[young ivy]

yeah, it just seems weird when someone very random adds me and says "let's be friends" or something. Unless I've talked to them in here that is...but then they should know their server nickname too...

[terse plume]

check the creators on civitai

[terse plume]

you just need see if he make the Lora that you want

and then talk with him

[bleak matrix]

Good morning, everyone!

How are we all today?

[terse plume]

nice

[terse plume]

how to make a good background?

[young ivy]

what?

I don't know? Or at least the knowledge is locked away right now since I haven't done any synthesis in weeks

[terse plume]

ok

[tulip shadow]

the same question bro

[fervent thunder]

Hello

[haughty wasp]

Hey, any good image deblurring AI model ?

[tulip shadow]

i'm also researching the same topic bro

[haughty wasp]

got it man

do you know any good outpainting model?

[tulip shadow]

finding bro

[bright ember]

Hi guys, sorry to bother you, but I always have problems. I wanted to create a gif with Animatediff but it gives me this error: OutOfMemoryError: CUDA out of memory. Tried to allocate 496.00 MiB (GPU 0; 11.99 GiB total capacity; 9.75 GiB already allocated; 0 bytes free; 10.98 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

how can I solve it?

[signal flare]

any good songs?

[trim thunder]

Can someone tell me how to use starlight filters in stable diffusion?

[wind ingot]

I'm no expert and never heard of Animatediff till now but reading their GitHub noticed they have an A1111 extension, you could try with some of the A41 memory management arguments

[bright ember]

where can I find all this?

[wind ingot]

The extension for A41?

https://github.com/continue-revolution/sd-webui-animatediff

[wind ingot]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations

Start with --opt-split-attention

[pseudo wyvern]

its not generating my image >;3

[knotty turtle]

Is it possible, in A1111, to use SDXL as the base generation, and then use SD1.5 for the Hires. fix?
I was just trying this but the image came out corrupted, presumably b/c I can switch VAE's in the middle of the process

[wind ingot]

Misread question,

You can load back the image and use another model

[fervent thunder]

Can someone explain to me why this software doesn't work on AMD GPUs? I tried to install it, but it doesn't work cause I have old GPU.

[wind ingot]

AMD needs ROCm and support for that ATM in mostly on linux

Unfortunately AMD is a mess, even for something as Blender most old AMD cards got unsupported meaning no GPU support, because ROCm is mostly for recent cards, funny thing is you can still get support on (very) old Macs with AMD cards because of Metal

[warm junco]

whats your GPU?

[neon hawk]

What ai do you guys recommend, for creating 2D sprites, textures and tilemaps for top-down games?

[cosmic cypress]

Pitch of a fictionary dystopic movie about the Roman Empire, in the style of Warhammer 40k, for AI lovers (Morgan Freeman cloned voice)
https://youtu.be/N9itVxjBE7U?si=YB6awmwNRTeDddv4

[upbeat pine]

bot is down?

[bright ember]

Hi CS1o, sorry to bother you, but I always have problems. I wanted to create a gif with Animatediff but it gives me this error: OutOfMemoryError: CUDA out of memory. Tried to allocate 496.00 MiB (GPU 0; 11.99 GiB total capacity; 9.75 GiB already allocated; 0 bytes free; 10.98 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
how can I solve it?

[upbeat pine]

@karmic brook bot is down?

[warm junco]

hey, whats was your GPU again?

[pale latch]

AMD Drivers is why. RocM is their implementation of the CUDA api and they've only started working on it for windows since stable diffusion came out

[bright ember]

RTX 4070 OC 12 GB

[warm junco]

can you share your animated diff settings in #🤝｜tech-support ?

[upbeat pine]

can someone resolve the bot please?

[pseudo wyvern]

@upbeat pine what's going on?

is the bot down?

[bright ember]

ok,thx

[upbeat pine]

yes..

[pseudo wyvern]

oh

[upbeat pine]

nobody resolve

idk

[pseudo wyvern]

alr

i hope the result of my ai image has a gyatt

JKJK

LOL

[north pelican]

no funciona?

it does not work?

[pseudo wyvern]

yeah

its down

sadly

[junior harness]

GPU is more important than CPU for faster image generation with Fooocus or stable diffusion right?

[spring inlet]

hi

[warm junco]

Yes GPU and vram amount

[junior harness]

CS1o, so a 16 GB RTX 4060 ti is better than a 12 GB 4070 ti?

[warm junco]

Thats a tough question because the 4060 ti has a smaller Memory bandwidth (bus) than the 4070 ti

More vram is always better for higher resolution or better training (or other ai stuff like llm)

But the 4070 ti would be still faster

[junior harness]

i see

But RTX 4080 will be much faster than 4060 ti, as it has the same memory but 256 bit bandwidth vs 128

right?

[warm junco]

Exactly, plus its additional cuda cores

[junior harness]

i see, thanks

does it really matter if you have, say, 13900k vs 13600k?

[warm junco]

For SD ? Nope

[junior harness]

thanks

[warm junco]

No problem

[lone spire]

idk if this is the place but does anyone know of a model, lora or way to create spritesheets (or just sprites in general) akin to 3rd or 4th gen of pokemon? creating the battle character works amazing with some models + loras but creating the actual overworld sprite is a nightmare

[primal nexus]

hands are the endboss for sd

they are driving me insane

working for 4 hours now on a single hand on a picture, didnt even start with the other one yet 😅

[pale latch]

people playing twister is the special hidden boss in ng+

[pale latch]

thats a lot of time for a hand.. not using controlnet or what?

[primal nexus]

I am using everything and then some

[pale latch]

could paint one in manually faster if you used that old fashion grid transfer technique

[primal nexus]

i did

[pale latch]

not a common hand position or what?

[primal nexus]

i used a photo of a perfect hand, colorgraded it correctly to match the hue of the subject in the image perfectly, pasted it over, and it still cant figure out to blend it in the picture correctly 😂

its a fist from the inside

[pale latch]

thatlldo.gif

[primal nexus]

no sadly it wont do

[pale latch]

throw some jpeg compression at it

[primal nexus]

as soon as i try to blend it in img2img sd fucks the perfect hand up, even with very low denoise

[pale latch]

take a cellphone camera screen shot

call it a day

[primal nexus]

i simply dont understand whats the problem, in other situations it painted correct hands in much more complicated poses

[pale latch]

whats your inpainting prompt look like?

[primal nexus]

i want to know whats going wrong there, i want to learn something from this

i tried fist, hand, fist from inside, then all those combinations together with the standard prompts of the image and without them

also tried no prompts

doesnt seem to make any difference whatsoever, it just puts the hands in a mixer and presents them proudly

[pale latch]

What model are you working from?

[primal nexus]

i could just leave the fist i pasted in with gimp and not change the picture anymore, it looks already 10 times better than anything sd produced

its called realpixelMix_v10

i tried others too

[pale latch]

might help if you showed examples in #🏞｜general-with-images

[primal nexus]

ok

[pale latch]

so we working with 1.5 ok. not bad we can do it still. you got a1111 or comfy or what?