💬｜general-chat

[chrome dove]

nope it's still doing it after hitting apply settings

[supple glen]

okay so is it just me or is that LinkedIn event that got posted in the recent announcements just straight up using the Super Mario 64 font

[woeful quartz]

hello friends i am a complete beginner at ai generation. i am looking for tips on impainting actual photos. are there any models that are specifically good at this? im currently using foooocus and their default models

[weak tusk]

Here is the latest Prompt Technique: https://workflowpedia.com/algorithm-of-thought

[plain crane]

I was reading about sd for MLx , i dont believe it , it says 8.79im/s , meaning 8.79 images per second ?? Even 6.32 /s for pytorch , no way , that’s 7 times faster than a rtx 4090 at 50 steps. That is insane speed for image gen even for an M2 Ultra . Even for training , that is fast. https://github.com/ml-explore/mlx-examples/tree/main/stable_diffusion

[visual pier]

Moin ^-^
How do I choose an Upscaler while Img2Img-ing in A1111? Obviously when resizing the target image

I mean obviously it needs to use an upscaler... somewhere right?

[unkempt hatch]

So what will be the successor to SDXL?

[old lion]

No official word but probably SD3.0

[visual pier]

Didnt SDXL just release :D

[elfin cove]

[finite cloak]

we might find an opportunity to release some tunes/variants of XL before moving on to a new arch

[unkempt hatch]

Just that even the newer fine tunes of SDXL are still worse than fine tunes of SD1.5

[visual pier]

I get that, I just think we should give it some more time in the oven so to speak

[unkempt hatch]

Maybe, but also it's so much slower, and not because if the resolution. SD1.5 at 1024 is way faster than SDXL... so also unless it gets hugely optimized...

[visual pier]

Yeah I cant even use it because its too demanding for now :D
And on top of that I am still waiting for proper checkpoints to come out. Currently I (think I) can generate more fitting images with 1.5 still

[warm junco]

if you select an upscale script at the bottom you will get the list of upscalers

[visual pier]

Ah thanks ^-^

[visual pier]

I decided to try out SDXL to see what the hype is about but the model refuses to be even loaded into my A1111 because my graphics card is appareantly too old now :D

Ah it just decided to be selectable. Lets see whether it works or if my pc starts melting :O

[signal trellis]

is there a tool that works like this:

input video of a person moving around
the tool analyzes the movement frame by frame
for each frame it aligns the controlnet model with the person
you now have a sort of "video file" of the controlnet model mimicking the motion of the person
this "video file" can then be used as a template for animating generated characters
does something like this exist or do we still have to manually align the controlnet model frame by frame? please @ me if you know, thanks

[feral sundial]

can anyone tell me that if stable diffusion did not respond any old prompt than how to generate new prompt

[sudden ruin]

You can extract a controlNet Video with comfy pretty easily

It can read out the pose frame by frame

[sudden ruin]

What do you mean?

[sudden ruin]

What GPU are you running?

[visual pier]

Nvidia GTX 1660 Ti

Its working now. But veeeery slow

[signal trellis]

any tutorials on how to do this? just learned about comfy today so havent ever used it

[sudden ruin]

Yeah thats kind of on the lower end nowadays

[visual pier]

I added the --medvram-sdxl line thats recommended for 8 gb vram

[visual pier]

Pff for 1.5 it works surprisingly well

[sudden ruin]

If you have ever worked with a node based software, youll find it easy to use. Otherwise you probably gonna need some time to adjust

You definitely should get the comfyUI manager

[visual pier]

Pff I personally think Comfy is technical to a fault. You can barely call it a UI in my opinion

[sudden ruin]

I love comfy, but I also enjoyed working in Nuke and with blender nodes

[visual pier]

Idk I personally feel less offended by blenders approach to a node based UI :D
Obviously I dont have a lot of time in comfyUI but so far it only complicated my workflow. A1111 spoils you with its usability ^-^

[sudden ruin]

Thats totally fair tbh

[visual pier]

I am glad 1.5 exists. If I had to start with XL I would not have gotten this far

[sudden ruin]

With all the fine tunes it still holds up for sure

[visual pier]

It took me about 30 mins to generate one image with XL, where usually it takes about half a minute with 1.5. And for whatever reason the image is botched :D

[compact kayak]

Just got my 4090, you know the first thing i did was download SD

[visual pier]

I finally have a reason to upgrade :D
I like SD as a hobby and a better GPU would help. Need to save up first though

[sudden ruin]

4090 helps a lot

[finite cloak]

Might I introduce you to #🐝｜swarm-ui ? Has the underlying technical prowess of Comfy, and a UI that spoils you even more than auto does :D

[visual pier]

Yes and I am currently using it... half the time. I still have trouble replicating some of my workflows but Im sure one day ill figure things out

[compact kayak]

I upgraded my PC after maybe 5 years, I use it for work and personal (self employed). Bought the Neo G9 ultra wide too 👀

[visual pier]

I am using my PC around 6 years now. But luckily I only would have to upgrade my GPU ^-^

[compact kayak]

Yeah that's much better 😅

[visual pier]

Still they are expensive and I dont have much money right now :O
I even thought to sell some comissions for SD but I dont like the idea of that. And extra income is kinda difficult in my situation and country

Like I like SD as a hobby. And I am impressed by the images I get out of it. But selling them would be a whole different can of worms :D

It has all kinds of ethical, legal and practical boundaries I am not ready to cross

[compact kayak]

Yeah I've been approached from some contacts for SD work too but I'm not really sure on the legal side of it either. I'm just doing it as a hobby for now

Does anyone have preferences here for UI, Automatic1111 vs Fooocus? I've tried both and perfer Automatic1111 for the amount of customisation possible, however Fooocus is quite good with it's ease of use and being able to tick different art styles

[visual pier]

I like A1111 most. And am slowly learning StableSwarm for some nieche operations

Like using different models for two pass text2img

[compact kayak]

oh cool, I'll check it out, I haven't tried StabelSwarm yet

[visual pier]

Its fiiine (at least for me). It has more options for customization but I dont think you need them all that often. Like it wont magically upgrade your model or anything

[compact kayak]

Yeah fair, I just need to get some more ideas of things to generate 😅 that's the hardest part

[visual pier]

Yes and the very basics. Like a good prompt is 80% of a good image

[candid bronze]

is the local sd better than the ones websites? I been using tensor art but that thing cant create an image without deform finger or body part.

[compact kayak]

The speed is great though with a 4090

32 seconds for the beolow settings

52 sample steps
1024x1024
batch count = 2
faces fix

[meager kayak]

I solved my problem, ty anyway

Hi guys, do you know how I could find the settings I used to train something with Kohya_SS?
I made a friend's face, I don't remember how many epoch, Batch size and steps I put. Do you have any idea?
EDIT: found log files

[ashen galleon]

anyone here worked with RLHF?

[velvet wind]

Is there a way to take photos in real life and use stable diffusion to change it

[ashen galleon]

you mean edit it?

[velvet wind]

yeah

[ashen galleon]

there is a work done lately by google, check it out: https://arxiv.org/pdf/2210.09276.pdf

[pastel nexus]

are sdxl forks even worth it?

like having separate models for anime/realism etc?

from what I've seen, juggernaut xl can do pretty much all pretty good just from styles

[sterile plinth]

Do people prefer ComfyUI or Automatic1111?

[sudden ruin]

Depends on who you ask

[empty iris]

Does anybody knows some open source Text to 3D AI???

[hallow plaza]

Is there anything better than kohya out there? I remember using everydream a while back.

[weak tusk]

Interesting! Mistral is less than a year old and is now “worth” 2 Billion $ 🚀

https://workflowpedia.com/mistral-funded

[pale latch]

stability are the company leading the open model releases. their text to 3d model is in research preview right now. contact them to ask for access

[pale latch]

naw. Inflated valuation is all. normal for startups. Thhat link doesn't even say anything about 2 bils. Often, bots will hype a startup on popular servers. Pump and dump stock bros love doing this. I really hope thats not going on here.

[floral umbra]

Do you guys know of a workflow for comfy/extension for automatic to make a 3d model out of a picture? Not fully 3d, just the visual part. So it will be flat on the backside.

[pale latch]

you can do quick and dirty ones with depth estimation and some javascript code, but it won't be a 3d model. I haven't seen extensions like that but i've seen it done in blender

[floral umbra]

That's the thing. I need it to be 3d. As in not flat. Unless i can easily make a depth model into a VDM for blender and make it 3d that way?

[jade tangle]

Question: I know I can use :1.3 and such to increase an AI's focus on one prompt, but is there a limit to it?

Does it go from 1.3 to 1.100? Or something?

[pale latch]

Blender would be the tool to use but i'm not sure what the process would be. absolutley there should be some sort of tool to convert depth + pixel info into a model. You'd likely need to do some trimming

[pale latch]

1.100 is < 1.3 😉

they're not "increased attention" but rather a multiplier on the token's influence

[jade tangle]

Soooo increased attention

[pale latch]

not really. these things don't really have much attention to begin with

[regal nacelle]

Hi folks 👋 I'd like to learn more about text-to-image models and had this idea of fine-tuning a model to generate memes but I have no idea where to start 😅

I've read the "Ressources" and found this intro https://github.com/Guizmus/sd-training-intro for SD training but not sure if that's the right place to start. Does someone have any ressources/tutorials to indicate?

[pale latch]

https://stable-diffusion-art.com/tutorials/
https://stable-diffusion-art.com/train-lora/

[floral umbra]

Aye, trimming and cleanup i can easily do. Just need the result to work. Sadly i tested a gen and made a depth map in photoshop, but was quite flatand lacked every detail. So sadly i need to wait until a proper image to 3d arrives. As i don't need 360 around subject 3d, just of the visual part.

Blender is what i'd use regardless to make a extruded "poster" out of it

[pale latch]

https://youtu.be/35cD7nHBh7U?t=111 i'm not going to pretend to know how to use blender. i used to do a lot of Maya3d work but stopped around 2008. this looks like the juice though. timestamped past all the "get an image with a depthmap" steps that puff it up

side note: hate that about youtube tutorials! "Learn advanced techniques!" 10 out of 15 mins are setting up a project and doing basic selections and file management steps

[floral umbra]

Just watched that one, but sadly, that's the problem. The depthmaps ain't detailed enough. https://image.duckers-web.site/hEja1/QoVEgaxi28.png

What i want to find however, is if there's a A.I/model that can turn a image's color codes with adjusting sliders perfectly to a normal displacement or depth map :D

https://docs.unity3d.com/2019.3/Documentation/uploads/Main/BumpMapColourMapStoneWallExample.jpg Like this, but from generated images.

[pale latch]

thats a normal map. there are absolutely monovision normal map estimations

not exactly great though

a1111 controlnet extension has one built in

[floral umbra]

Aye. Do you know what "2d to 3d" displacement texture "method" that adds the most details? Would that be normal map? Or some other material type? Similar to the technique star citizen uses for their "3d" image assets

[pale latch]

i've seen some really great "extruded posters" as you call them. and they're definatly not just a image with a depthmap displacement. there's some polishing for sure going on

[floral umbra]

Indeed. All i need is a displacement texture of sorts, like a depth map, but doesn't just capture smooth shades, but every detail of the image. Then i can clean up afterwards

[pale latch]

https://github.com/thygate/stable-diffusion-webui-depthmap-script this also is an extension specific to depthmaps. worth noting that normal map esimations start with a depthmap and combine pixel color with depth info. so even that appraoch needs a good depthmap

[floral umbra]

Aye. I'll check it out :)

[pale latch]

https://youtu.be/yfWHVigsOk0?t=445 this guy used it. his vid has some points that the other didn't. such as a lot more points on the plane. ahaha 3d editor puns

[floral umbra]

Aye. Compared to others, i tend to go a bit overkill on my subdivisions anyways to get the point out

[pale latch]

wait... off topic.. but youtube recommended this 3 hour old vid to me. we're in the golden age of gaming i swear https://www.youtube.com/watch?v=NC40SRNADAQ

[floral umbra]

There's so far been 4 golden ages with games

1: computer based games where you control the little toon on the screen was born
2: Computers became strong enough for games to be made in 3d space
3: VR gaming was born where YOU are the controller
4: Ray traced games to have actual simulated light rays in realtime.

[static herald]

hi guys, is this general-chat channel a place where i might ask a question about stable diffusion, or is there another channel for that sort of things?

[floral umbra]

Here, or any of the stable diffusion subchats depending on what you wanna make

[static herald]

thanks, here goes: i want to ask if anyone knows what kind of problem i seem to be running into

i'm trying to do that anime photo to realistic photo conversion thing

and the results keep coming back very strangely shaded

like the controlnet lineart is keeping the shapes correct, but the colors are just not following the shapes at all and are just random

[wide shoal]

Basically. If you want 3D "objects" from an image, that will need new kinds of models, but I've seen some "image 2 3D" stuff.
Using just regular image models and depth, you can only get projection-displacement kind of stuff which is quite okay for some small parallax effect on scenes and environments.
For the latter, take a look at his: https://github.com/thygate/stable-diffusion-webui-depthmap-script
I see they have options to output an actual 3D mesh.

But the mesh will have projected texture with no fancy materials or shading

[floral umbra]

Neat! It doens't inherently have to be 3d, just a depth map that retains the same amount of details as the OG image

[wide shoal]

Then any depth controlnet preprocessor can give you that

Although that extension goes deeper and employs some advanced techniques to provide better results

And outputs 16bit rather than 8bit depth

[floral umbra]

Tried those. Didn't look too good. Gave quite foggy results last i tried

[wide shoal]

Try it with the Multi-resolution merging option, it's pretty impressive, although slow

(that extension I mean)

https://github.com/compphoto/BoostingMonocularDepth/raw/main/figures/merge.png

[floral umbra]

Slow is no problem provided result is good :)

Also rip, 3d mesh one could only do fast, and not slow and accurate

[wide shoal]

If you get the good depth map, it's pretty trivial to displace it in Blender yourself

I see people have linked some videos already

[static herald]

any chance anyone had any issue similar to this? I posted an example of this sort of coloring issue in the tech support channel cos this channel doesn't allow photos...

[wide shoal]

Oops, somebody already linked this. My bad, nothing to add then :>

[floral umbra]

https://image.duckers-web.site/hEja1/tamERUzE40.png Thanks, i hate slow ass servers..

[pale latch]

be me with a 4080 still playing hollowknight

[pale latch]

looks like the model has no idea what to put into the lines. did you prompt?

also pro tip. you can link messages. #🤝｜tech-support message

[floral umbra]

Or be me with 3090 playing dave the diver in 4k Or avatar in 4 hours

[pale latch]

i want that diver game. looks like it vibes like echo

[floral umbra]

thanks to dave the diver, i now want more indie 2/3d games :P Even got the idea to learn coding and pixel art making to make a subnautica clone, but 2d

[pale latch]

i actually got a steam deck oled so my 4080 is getting a lot less gaming heat these days. mostly diffusion

[floral umbra]

Which echo?

[pale latch]

subnautica in 2d could work well. i love subnautica so much. expertly crafted into 4 acts around the depths of the ocean

echo the dolphin of genesis fame

[static herald]

thanks so much, don't use discord that often and learned a new thing : )
I used the prompts and model from a tutorial which showed good-looking examples while mine mostly turned out strangely colored

[floral umbra]

I got the old deck with 2TB nvme and 1TB sd card :P Gonna upgrade the ram to 2x at some point next year

[pale latch]

i've beaten subnautica half a dozen times but the sequel i haven't even finished. it's just not compelling to me

[pale latch]

no no no. trade up. get the new oled and upgrade that!! the screen is sooo saucy

[floral umbra]

Same. Felt bland. And sequel is even subnautica 1, just replace assets, that's it. Same old bad bugs the first game had

[pale latch]

upgrading the ram could be fun surgery. i didn't open my original 512 deck at all though. wasn't planning on till the warranty was out

[static herald]

so if it's the model(checkpoint right)'s problem, installing a new model should solve it, right...?

[floral umbra]

I'm not gonna pay 550 bucks again for only 8% increase 25% and we'll talk. That's my demands in any upgrades. But with nvidia being nvidia, i demand my upgrade that costs 2.5 grand to have 4x performance gain lol

[pale latch]

well, no. my theory is you didn't prompt at all

or have your cfg set to 20

[pale latch]

it's a 100000000% increase in contrast ratios

[static herald]

prompted a lot actually, cfg 40

[floral umbra]

Still not worth it. Plus i live in norway, getting hold of one is the difficulty lol. As we ain't allowed to buy straight from valve. Was bad enough with 97 ship and 130 import fee for base 64GB deck

[static herald]

installed a new model and it seems to fix things now, things work

but the old model i used is a reliable one i think, it's realistic vision

[pale latch]

true nuff. no warranty to run out in that case. all in all, the deck, no matter what model, is worth it.

Hollow Knight is a beautiful game

[floral umbra]

40? sheesh.

[static herald]

higher is better right...?

like with diminishing return

[floral umbra]

Aye. I'll wait for a new deck with a new stronger cpu and gpu, not just a die shrink and oled screen

[static herald]

oh wait did i think about the wrong thing

yeah my bad sorry cfg 7

[floral umbra]

No. Lower means "do what i say damnit!", higher means "do what you want damnit!"

[static herald]

was thinking about step

yeah yeah was thinkg about the wrong concept

cos i think on my ui cfg goes only to 10

oh wait no

i only ever pushed it to 10 but still ,yeah see no point in pushing to 20

[floral umbra]

So start at say 7cfg, and increase by 2-5 until it looks bad. Then you know that model's range. SDXL iirc can do 20-30 with ease.

[plush night]

where can i see random pictures yall created? or post mine?

[static herald]

ohhhh

[pale latch]

technically it is an apu upgrade. 6nm from 7

[static herald]

it's that high?

tutorial i found says it should stay on 7 so i thought that's like the range...

[floral umbra]

#🏞｜general-with-images

[floral umbra]

Same apu, just shrunk. I want new arcitecture. Zen 4 cpu damnit!

Or hell, increase APU from 8CU's to 12. If asus ally can, then deck can as well :P

[pale latch]

i know. i just love technical pedantry 😉

i used an ally. the ui is shit lol. it's just some asus spamware gateway slapped on top of windows 11

[floral umbra]

Aye. I want the performance increase, not the ally itself. It lacks too much, and after asus's fuckups with boards with bios's that literally fries cpu's, and bios updates that breaks your warranty, i avoid asus these days lol

@pale latch https://image.duckers-web.site/hEja1/MuKEGOvo38.png Could it be the model i used that made the mesh turn to shit?

[pale latch]

hahaha i don't know what is going on there. 3d mesh estimations are one of those fields thats a lot of "YMMV"

[wide shoal]

Did you take a good look at it.

It looks like that's a view from the top

And looking at depth projected meshes from the top like that always looks funky

[plush night]

any1 making money with stable diffusion here?

[regal nacelle]

Hey folks! I've just fine tuned my first model following this tutorial https://stable-diffusion-art.com/train-lora/ The Model is saved on my Google Drive but now I'm wondering how can I run the inference on it? I don't have Colab Pro, so can't use AUTOMATIC1111 Stable Diffusion WebUI

[maiden parrot]

can someone explain me the difference between "Stop at" and "Weight" in the img2img panel inside fooocus?

[weak tusk]

I get what you’re saying about startup valuations - However Mistral is for-real! I’m a professional Prompt Engineer and I have extensive experience with Mistral 7B

[pale latch]

feels like hype to get investors. what's it's real world application?

"it can do what chatgpt does" okay, what's chatgpt's real world application?

[weak tusk]

They are open-source, and I’m not promoting them, I’m promoting competition to the OpenAi near monopoly

[floral umbra]

It's shaped like one very long spear of stretched mesh :p

[lavish current]

If anyone here does #1072016199837290536 or 3D #1072014484807356526 gens you guys gotta try Pika Labs

Fantastic for animation

[copper crystal]

closed source. pay 2 use. meh

keep it on their discord server

another thing i dont appreciate about them. Just like OpenAI. They're a non profit that is running a for profit company. It's obvious tax evasion and they're obviously not giving any service that justifies non profit status. They're building everything proprietary and exclusive to their business.

[lavish current]

You mean Mistral right?

Pika is free

[copper crystal]

no. pika. 1.0 will be charged for . they don't intend to keep it free

[calm flax]

if someone can help me in #🏞｜general-with-images , I'd greatly appreciate it

[earnest prairie]

Discord's new Mobile UI is trashy

[unkempt forum]

Hallo, how to invite BOT on own server? Thanks

[earnest prairie]

On the officiel website

[unkempt forum]

Oouhh thanks i'll try, i think on discrod 😁

[earnest prairie]

I succeeded, but I couldn't generate

[dire cedar]

Hello!

I have some error rendering images

can someone help me fixing it?

[fickle jewel]

Hey, i was wondering if anybody knows if it's possible to extract the original training data from a lora file?

[waxen berry]

what kind of training data

https://www.reddit.com/r/StableDiffusion/comments/13efkwn/do_you_want_to_see_how_your_favorite_lora_was/?rdt=55178

[cinder wing]

Can you expand on the issues you're having?

[ivory nexus]

How to find key prompts for styles?

For example poster style

[misty pollen]

ANyone struggling recently with random infinite wait times when the image is on 95% completed?

[waxen berry]

a couple times maybe on a1111

if your system was 100% dedicated to SD, you may not have such an issue

[fervent thunder]

Just wondering something, my pc has 4gb ram and that would be enough to use SD?

[fossil phoenix]

are you using faceswap? lol on A1111 those things love to do that, or other last step scripts

[fossil phoenix]

this is a joke, yes?

are you using windows 3.1?

[fervent thunder]

10

[fossil phoenix]

you'll have about 2gb free, with win10

[fervent thunder]

This isn't a joke. People says I need at least 16 gb

[fossil phoenix]

vram or ram

[fervent thunder]

Ram

And which level python learning would be enough

[covert dirge]

Hello guys, anyone can help me to obtain prompt of people without face distorted? Is there a method to not obtain this distortion

[bleak matrix]

Good morning, everyone! How are we all today?

[odd maple]

how do i fix error code 2

[sudden ruin]

#🤝｜tech-support and only post your question into one channel

[calm flax]

People

Is there a Hires. fix for img2img?

[eager phoenix]

another non-commercial 😔
and no response from reaching out to try and get access through the contact form last week

Does anyone have any updates around how we can use these models if willing to pay / create a partnership?? @rigid bough @karmic brook since you guys are only stable staff on rn tagging you in case you know

[pale latch]

Some authors leave the training meta data in. I find that's messy and unfortunate though, so i like to build my own json of metadata and attach that to the model instead

outside of that, there's no way to unbake a cake (get training data from a model). you need that recipe card (training meta data)

[rugged bough]

what the hell are these nubmers real

[pale latch]

nubmers are not real. abort. ABORT

[rugged bough]

must test must test

[pale latch]

square of 2. 1/0. This statement is false! ABORT

https://tenor.com/view/bart-simpson-lisa-simpson-the-simpsons-gif-8731042

^ posted a bort

[subtle pumice]

Whats the best model for really detailed anime/cartoon gens?

[radiant lantern]

what's the context size of stablelm-zephyr-3b please ?

[warm junco]

Yes the SD upscale script

[calm flax]

thanks!

[humble ravine]

back

[obtuse lake]

Where is image2image option?

[crystal edge]

who has basic negative prompts

[pale latch]

the automatic1111 hiresfix is basically just img2img compacted down for use in text2image tab

[earnest dirge]

Do someone know a cheap way to get background removal through api? 🙂

[pale latch]

my go to is "dead, deformed, disfigured, broken, old, wrinkles, corpse, jpeg low quality, compression artifacts, banding, grainy" i modify it all the time too. i stopped using presets and type it everytime to give myself more experimentation room

https://removal.ai/pricing/ @earnest dirge

[earnest dirge]

This is the cheapest i have found so far: https://picwish.com/pricing?tab=business but i wish to get cheaper, as the project i am currently building will require hundreds/thousands of requests 🙂

[pale latch]

At that scale it might be better to run AWS with an open solution

https://huggingface.co/spaces/ECCV2022/dis-background-removal

[earnest dirge]

Have not heard of it, but i will check it out thank you very much! 🙂

[pale latch]

or azure. whatever cloud provider works best for you

[earnest dirge]

The one you linked might actually work wow! thank you so much man! 😄 Appreciate it very much!!!

[pale latch]

yeah if you're doing thousands and know enough to run your own model solution, you're golden

[limpid root]

Hi all!
I'm an Engineer that is new to Stable Diffusion. I am looking to use it for a commercial project, primarily looking at hyper-realism, and consistent avatar use (so create a character/face, use it in many pictures/videos).
I was wondering, is there a "best UI" to use? I've used Fooocus, but felt limited that I couldn't copy builds from CivitAI, and the lack of tutorials.
I've seen ComfyUI and A111 come up a lot, so, is one of them superior?
If anyone can suggest some advanced tutorials, that would be awesome.
Thanks all!

[warm junco]

hey, most people use auto1111 or comfyui. Auto1111 is easier to get into if your new. Comfyui is for people who like to control every aspect of the creation

[limpid root]

Thanks mate!
Comfyui it is! Do you know of any good tutorials??

[warm junco]

i think this can help:
https://www.youtube.com/watch?v=REdk5UuvTyE

[limpid root]

Great, thank you!

[bright nest]

2 hours ago Emad said something about "Stabilty AI membership" on X. im not sure what this means? whats this "membership" and is it something we have to buy into?

[south steppe]

Hello! I am having issues installing stable-diffusion webui

more specifically, i cant run the webui-user batch file. Error code 128

[smoky osprey]

can anyone help me my extensions are not showing up even tho they are installedI have webuireactor downloaded and adetailer but only adetalier shows up

[earnest dirge]

@pale latch Do you know any more websites that do background removal or any github? I wanna implement it to my website, the one you linked before huggingface, turned out the API did not work and many had troubles with it importing it on node.js

is there any plans for stability to integrate background removal for their api? that would be awesome.

[pale latch]

clipdrop is owned by stability and i believe they do bg removal actually

https://clipdrop.co/remove-background

https://clipdrop.co/apis/pricing 1 credit per bg

[hard steeple]

does anyone know how i can upload a reference image with my prompt?

[raw dagger]

What UI

[raw dagger]

https://www.remove.bg/api

[hard steeple]

in the bot channels?

"/ dream" prompt

[raw dagger]

Oh I've never generated something with the bot here before, so I couldn't tell ya, sry

[hard steeple]

oh no worries! thank you 🙂

[wanton karma]

ERROR:root:!!! Exception during processing !!!
ERROR:root:Traceback (most recent call last):
File "E:\ComfyUI\ComfyUI_windows_portable\ComfyUI\execution.py", line 153, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI_windows_portable\ComfyUI\execution.py", line 83, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI_windows_portable\ComfyUI\execution.py", line 76, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI_windows_portable\ComfyUI\comfy_extras\nodes_clip_sdxl.py", line 44, in encode
if len(tokens["l"]) != len(tokens["g"]):
~~^^^^^
KeyError: 'g'

Please Help!!!!!!!!!!!!!!!!!!
Relates to COMFYUI

[old lion]

screen shot ur workflow?

[wanton karma]

everytime when i use cliptextencodesdxl this massage appears

jst sending you the image... wait pls

[wanton karma]

#🏞｜general-with-images message

[old lion]

ur feeding the positive prompt into the negative?

[wanton karma]

i tried everything but when render reads CLIPTEXTENCODESDXL it stucks and showing error msg

[cedar salmon]

Hi, In ComfyUI can i save images in a plot while Auto Queue is running and I'm typing real time Prompts. Digging around some packages trying to figure it out, so far i only figured out how to plot batches and not Auto Queue singles.

[subtle pumice]

is there a good tutorial or guide for good anime images

also good models?

[dry tendon]

Guys, do you have any good new upscalers to recommend? I must be outdated on this.

[subtle pumice]

whats an upscaler?

[cedar salmon]

a make it biggerer, work with smaller than upscale to bigger

[pale latch]

its used to weigh things by pushing up on them

[subtle pumice]

ohh ic, whats a good one? is it worth it? i am new to this

[dry tendon]

https://www.youtube.com/watch?v=or2xaftZXvE

[subtle pumice]

thank you

[pine cape]

is there a way to make SD choose between multiple parameters? for example, if I want either green, blue or brown eyes, I want to give the option of either of them but let SD choose what to put in. Can I do that reliably?

[smoky swan]

hi, I'm new here! I'm facing difficulty installing SD on my mac. I got stalled at the first step of installing homebrew. Is this the right community to come to for help?

[earnest dirge]

i want to use https://huggingface.co/spaces/ECCV2022/dis-background-removal and use their api on my website is this even possible? Have anyone done it? when i try to import it npm i -D @gradio/client and then try the api it just fails, if anyone have an solution please dm me.

It is free too right?

[bleak matrix]

Good morning, everyone! How are we all today!

[bleak matrix]

I would check #🤝｜tech-support for info if you need help! Also, if you want to install SD, I would recommend that you actually install #🐝｜swarm-ui You can find the page here: https://github.com/Stability-AI/StableSwarmUI

[chilly quest]

Hey all. Does anyone know of a list of optimal resolutions not only for SDXL, but SD15 and SD21? I've found some SDXL ones easily enough, but not for the other two, unless they're both as simple as '512, 768, and all combinations of them'.

[cinder wing]

It is recommended to generate at the size the model you're using was trained on. For 1.5 this is 512x512, for 2.1 this is 768x768. SDXL was trained on 1024x1024. You can always upscale/hires fix after, but generating at larger sizes than training data can yield more unwanted artifacts and repeating textures.

[chilly quest]

Shouldn't there be a list of base model trainings though? Like 1024x1024, etc for SDXL.

[cinder wing]

I listed the training resolutions for the 3 major base models in my previous message. Finetunes will use the same resolution as the model they're based on. For non-Stability models trained from scratch, the developer will usually include the training data resolution on the model card.

[cinder wing]

Each model is only trained and optimized on a single size, so there's not really a "list" per se. In theory, you should be able to use any multiple of 64, however doing so with high denoise can lead to tiling.

[wind kraken]

how do i stop it from making videos?

[pale latch]

SDXL and 2.1 were trained using bucketed resolutions. While the very most base resolution is what you've shown, they also have many other resolution capabilities. Fine tunes don't always use the same resolution either. Many SDXL models have effective renders below their base resolution and many 1.5 models can pull off 768 images well too. Since refined models often use higher resolutions.

There is absolutely a list of bucketed resolutions that SDXL is trained on. You might want to read through the technical report because your knowledge has gaps. https://arxiv.org/pdf/2307.01952.pdf

https://i.imgur.com/ZZ3JdeB.png

@chilly quest relevant post for you too

When i train loras for SDXL, i train them at 768 or 896 resolution. The base resolution isn't a hard lock.

[last estuary]

does stable diffusion follow instructions better than Dalle?

[pale latch]

lol no

[sacred jay]

have we gotten any word on when the Stability membership will come out? I'd really like to use sd turbo for commercial uses

[pale latch]

youve got things like guidance models to use with stable diffusion, but prompting SD is a lot more of a skill than it is with Dalle

It's like operating a crane where there's 100000 pivot points and some of them do random stuff when you're inputting controls. Can you use it ? sure. If you're good at using it you could call that "skilled". Most times it'll be disasters though

[pale latch]

https://twitter.com/EMostaque follow

[sacred jay]

I do, just hoping there might be more info here than 'soon'

[abstract night]

https://youtu.be/_p2KW6olwIw

[pale latch]

here gets all it's news like that from emad. he's here too you can search his user name and see all his posts

huh nevermind. discord is hiding him from search bar now

side note: i think lately, sites like twitter, reddit, discord, have been changing a ton of their back end stuff to prevent future scraping. It's been fucking with little shit all month long. like .. unable to find emad in the search bar for instance lol

can't link discord images anymore either. image links will need to have a special session token that needs to update all the time. image links will expire from now on

cant see reddit posts from over 2 weeks ago anymore. huge limits on their api now

elon has been pretty vocal about all the limitations he's dropping on x.

wouldn't be surprised if meta is up to shit too but tbh, they probably led the pack years ago to prevent data scraping

[inner lintel]

Why scrape anything if modern-day AI can generate decent data itself? 😛

[pale latch]

can it though?

[inner lintel]

"Synthetic data" is a thing for a long time

LLAVA can generate perfect captions for images

[last estuary]

way cheaper

[pale latch]

Yeah 100% i agree. People make a few distilled models and generative content to train on. a base model though? hmmmm. Most you'll do is refine the capabilities of the first model into the new model at best. It won't learn anything new.

It'll only get the best of what the model is capable of, and only if some curration is done

Synthetic data is useable but there's still a ton of value in training on real world data too.

[dry moss]

hi there, is there any way i can edit a generated image by adding other elements on top ? i need it to create a parallax effect

[hallow plaza]

Whats currently the best image to caption tech out there?

[pale latch]

llava from what i can tell. I usually just use WD vit tagger though in the dataset editor extension.

[median oxide]

Anyone use DreamShaper 8?

[hexed cosmos]

Newbie to dreamstudio.ai here - I kind of went wild and created many hundreds of images. It seems like the older ones are no longer accessible. Are they gone for ever? No, I did not download. 🤦‍♂️

[delicate mural]

Anyone know of a know of a prompt tools that isn't Clip Interragitor?

[pale latch]

3 messages ago i was talking about dataset editor extension https://github.com/toshiaki1729/stable-diffusion-webui-dataset-tag-editor

[delicate mural]

is this something you have to download?

[pale latch]

yup. its an extension for a1webui or stand alone

its meant for captioning large image sets so you can do it to a folder , and each text file works as it's own prompt

https://github.com/AIrjen/OneButtonPrompt theres also extensions like this which are made to generate prompts at the push of a button

[gloomy grail]

Hey guys, is there a channel where I can find people doing stable diffusion gigs?

Looking to pay someone for some assistance on the platform

[inner lintel]

On a serious note, can any of these new open-source models surpass SD 1.5?

Pixart seems to not

Even 2.1 and XL turned out to be worse, in fact only a model of DALL-E 3 quality could replace SD 1.5

[potent spire]

Replace/surpass in what area*

? *

Image quality?

[inner lintel]

Accurate "understanding" on text prompts. It's not the only thing for t2i quality, but is the most important imo.

[potent spire]

Oh, for that D3 is still the king

[finite cloak]

2.1 was kinda not great yeah, but XL is incredible compared to 1.5

have you actually gone back and tried SD 1.5 recently?

Not your favorite community model built in 1.5, but 1.5 itself

that model sucks

(compared to XL)

XL 1.0 base beats most 1.5-based community models except in extremely specialized cases. Eg community 1.5 models tuned for perfect photoreal run through hires fix will do better than XL 1 base at photoreal to get the same image at native res or whatever. In terms of photo quality at least, (not prompt following, XL is raw king at that no competition whatsoever tbh).

There are community models for XL that are getting really really good. For example, Albedo, which is a giant merge of tons of community XL models, produces some incredible results (though it has a bias towards photoreal, since that's naturally the focus of a lot of community finetuning)

[inner lintel]

The point is that 1.5 is valued for its fine tuning, which is the point of an opensource model

And, there are some things that old models based on NAI leak still doing better, than any of XL finetunes

[finite cloak]

if you only ever use the model that has the best finetunes, you'll never stop using SD 1.5. No model will be the best at finetunes until it has had a year+ of care put into that as 1.5 has.

Even now XL shreds 1.5 on most points, especially prompt understanding as you were asking about lol

[inner lintel]

For photorealistic style, yes maybe XL is better

[finite cloak]

had a relevant conversation in r/SD discord a few days ago:

in a conversation about how to best prompt 1.5-based anime models i offered my first attempt at grabbing a random xl model and just handwriting the first natural prompt that comes to mind and the result is way better than anything else shown thus far

[inner lintel]

That 3d nose tho 😅

I mean, it's 2.5d

[finite cloak]

(that was more specifically using albedo, which i mentioned earlier as having a bias towards photoreal)

notably with a prompt that got specific about the background, XL was the only result that even added a coherent background, much less got it correct to the request

[inner lintel]

I tried CounterfeitXL and found it to be overfitted while still have poor overall style, but perhaps it's only this finetune that is bad

[finite cloak]

had a conversation in the same discord months ago comparing anime models at the time

blue pencil was the best from my short testing there

counterfeit was particularly wonky

[inner lintel]

It doesn't help that devs removed all NSFW data, so XL is worse at understanding characters anatomy

[finite cloak]

it's pretty good at anatomy, definitely better than 1.5 base was

not the best at naked people but you can find community finetunes for it if that's your taste

[atomic marlin]

is there a way i can make real looking people and photos look less staged? more improv and real situations

any advice on prompt wording?

[last sierra]

I have a 7900 XTX, I will be able to get a MSI 4090 for the price of 1600 USD, is it worth it?

[finite cloak]

You can try literal descriptions, eg "candid photo", or you can try adding contextual hints eg "gopro shot" naturally makes it lean towards the type of images you'd record on a gopro, or you can try adding weirdly specific phrases that would strongly associate towards natural situations, eg facebook pic of a dude at a party will give you a poorly taken bad picture of some freakin' dude posing in the middle of a party surrounded by random background people n stuff

[finite cloak]

Whether it's worth it is a choice heavily dependent on your own life circumstances. If you get paid 6 figures and love AI? Yeah absolutely. If you're struggling and going to ask friends&family to help pay for it? Absolutely not worth it no.

In the Current Era ™️ having any nvidia card will just make life easier with using AI software, and probably run faster/better too. Online benchmarks say the 4090 is about 50% faster for gaming, so it's probably twice as fast for AI (Because, again, AI stuff is just built for nv and not for amd these days).
That's getting slowly better every day though - so if you're planning on the scale of whether it's still worth it a few years from now? Probably not so much

[last sierra]

Yeah money isnt that massive of a issue, but if its way better then I'd purchase it, 7900 XTX here is around 1400 USD.. I am able to return it also, but AMD wont be improving much AI in the future?

[finite cloak]

If you can return the old one for $1400? Yeah the upgrade to a 4090 is absolutely! worth $200

that's an easy huge win in speed and reliability

AMD might / probably will improve on AI in the future. Who knows how long that'll take or what roadblocks or edgecases might take longer than the rest.

At $200 when money isn't a big issue, the years it'll take to get there ain't worth waiting for

and a 4090 is still a faster card regardless, by at least the 50% on gaming benchmarks

[light dove]

why this error keep on showing 'Nonetype' object has no attribute? while start render the stable diffusion
can anyone explain why and how so that i can rectify. pls help

[slate hare]

https://imgur.com/yp73BXe

[fervent thunder]

Hii! I'm not sure if I can ask here directly about help. But I'm trying to achieve a very specifc manga style and can't seem to find any model that is able to reproduce what I'm looking for:(

[steep star]

i think 2x3090's would be a better option
yea, you cant you both at the same time for SD

[warm remnant]

NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

Someone had this error?

[warm junco]

whats your gpu?

[warm remnant]

4090

[warm junco]

okay, and when do you get this error?

[warm remnant]

When I try to faceswap using reactor

[warm junco]

okay strange a 4090 normaly dont need --no-half

[warm remnant]

Its 4090 24GB, where to add this --no-half?

[warm junco]

is the model larger than 1.98gb ?

[warm remnant]

Ye, its juggernaut, 6GB

[warm junco]

so its sdxl?

[warm remnant]

standard sd automatic1111 webui

[warm junco]

i mean the model, is it for 1.5 or sdxl?

[warm remnant]

I'm not sure where do I check that?

Oh yea, its for XL

The model

https://civitai.com/models/133005/juggernaut-xl

[warm junco]

okay, can you try with an 1.5 model to test if it works?

[warm remnant]

Downloading

works now

thanks G

Although not the perfect faceswap

[cosmic summit]

please tell me what Refiner is for?

[cedar salmon]

someone have a favorite x2 general upscale model, ive ended up with BSRGANx2 so far out of what i tried

[crystal edge]

why i always get distorted face and hands when using full body prompt

sdxl 1.0

[nova basalt]

Hey. So does anyone have a link for a google drive colab stable diffusion thingy? I was using one for some period of time but I'm not a colab pro user, so I can't use It anymore.

[livid crystal]

Hey guys

[livid crystal]

I want to create a paper cut craft style of an image, how can i do it?

[wary belfry]

it enhances details of images generated by the base model

[spiral inlet]

@warm junco around?

[warm junco]

yes

[spiral inlet]

sorry i didnt get back to you the other day, had to leave

i just did the command line thing you mentined but now getting the error ERROR: Could not find a version that satisfies the requirement torch==2.0.0 (from versions: none) ERROR: No matching distribution found for torch==2.0.0

[warm junco]

Can you post the full cmd log in #🤝｜tech-support ?

[spiral inlet]

ok

[warm junco]

With gpu

[hot ridge]

Hi Guys,
I'm new to the world of DL,
i'm having problem with my code.
i was refering to https://github.com/justinpinkney/stable-diffusion/blob/main/notebooks/imagic.ipynb
and i don't have ckpt file,
is there some way to download this file, like npm in webdev?
or do i have to code it?

[zinc sun]

Hi, how can use some terms like bigger, smaller ? That don't seem to work for me

[hot ridge]

I'm sorry, whom are you refering to here?

[zinc sun]

I'm just asking help to use stable diffusion

[cosmic summit]

Do you happen to know if the new fooocus interface can support sd 1.5 models as the main one?

[tough oriole]

Hi!
Does anyone know any news on the SDXL engine update for Clipdrop etc?

I mean, on clipdrop sdxl looks much worse than on discord bots.

[tidal magnet]

Hey guys, what do you use to prune the models except for A1111?

https://github.com/lopho/stable-diffusion-prune
I have tried this script, but it doesn't compress the model at all.

[pearl raptor]

Hey everyone! My Stable Diffusion is running super slow. Here are my specs:
ASUS ZenBook Pro Duo 15 OLED UX582LR-H2002R-BE

Intel® Core™ i9-10980HK
32 GB DDR4 RAM
1 TB SSD
NVIDIA GeForce RTX 3070
Thoughts on whether these specs are sufficient for better performance?

[warm junco]

hey, do you use auto1111 ?

[pearl raptor]

yes, on stability matrix

[warm junco]

ah okay, then activate xformers

for sdxl you also need --medvram-sdxl --no-half-vae

[pearl raptor]

done and medvram too

[warm junco]

with medvram your slowing yourself

[pearl raptor]

ok i'll remove it

Launching Web UI with arguments: --medvram-sdxl --xformers --skip-torch-cuda-test --no-half-vae

[warm junco]

nooo, dont use --skip-torch-cuda-test

with that it wont use your gpu

[pearl raptor]

alrightµ

[pearl raptor]

thanks bro btw

[warm junco]

no problem 🙂

[cosmic summit]

Do you happen to know if the new fooocus interface can support sd 1.5 models as the main one?

[covert dirge]

Is there a method to avoid distorted faces?

[spiral inlet]

is there any sign of sd becoming as good as dalle3 or midjourney any time soon?

or am i using the wrong model? im using what installed with automat111c, 1,5

[old lion]

u need to run SDXL, also it depends on what you fancy....

[fervent thunder]

question, can impainting be used to generate art with a non square/rectangular format, such as a triangle or circle, while yeilding the same results from those sqaure/rectangle shapes, beside the drop in resoltion since ur making a smaller shape inside the square

[spiral inlet]

@old lion where can i get the model?

An artistic design I can use for a skin of a Kalashnikov in CS go with the theme of finished redwood and shiny platinum trim

dalle 3 gives me....

wait cant i add an image in here?

i put a comparison in https://discord.com/channels/1002292111942635562/1004159122335354970

[old lion]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main

[devout pewter]

Make a model that can generate a city, geometrically smooth and detailed, as in real life, not just buildings, but consciously, the AI must understand what steps, windows, trees, and other details are intended for. in order to be able to design cities in SD, any historical time, from historical to modern. perhaps you will need a special controlnet for this, to build buildings and streets on bones and guides

[worthy lynx]

Hello, I am looking for someone very comfortable with AI to create a drawing in the style of GTA. I will provide photos of the person's face and their dog. The person in the drawing should resemble the real person in facial features. I am willing to pay around fifteen euros/dollars. If you are comfortable with this, please do not hesitate to send me a private message.

[keen patrol]

hello guys. my 2 year old really loves books and i was thinking of making our book own story with him, me our animals. i tried using a cartoon checkpoint and it generally it works fine but i need some caracter consistency. i was thinking of doing a lora but since its cartoon it doesnt need to be that perfect. just to be similar to us. Whats the best approach of doing something like that?

[ivory nexus]

Hello!

What is the best model to make small human figures against the backdrop of nature (i'm using vector style Lora)

?

[median oxide]

How can I stop SD auto loading prompts on startup? For some reason, there are some prompts that load and a seed number too

[inner lintel]

Any thoughts on the "TokenCompose" solution for SD?
Looks like it can improve quality of text encoding to Dalle3 tier. It is said to be done by fine-tuning of already existing checkpoints.

https://mlpc-ucsd.github.io/TokenCompose/

If it works, I would like to see more models refined by that method

[lapis wigeon]

Why doesn't this server have any channel for lora making? Just did my first lora :)

[lofty ingot]

Hey, im very new to stable diffusions, ive just been generating very simple things of general interest to me (Traffic cars, big work trucks, highways with tons of cars)

can someone explain sampling methods to me like i am 10?

its just overwhelming the ammount of them youknow? i dont even know where to start with picking the correct one

[lapis wigeon]

Sampling methods are like different people. Every Sampler is an different person. A few know each other and are pretty similar, and a few are completly different. A Sampler reads all the information you gave him (prompt, resolution, cfg, ...) and draws a picture. Every sampler has a bit different style. Most commonly used are: "Euler A" and "DOM++ 2M Karras". You can also do an XY(Z) Plot to compare these samplers and see for yourself what you like.

Oh well, Dreambooth can create lora. So the list doesn't really make sense. Loras can be compared against, Textual Inversion, Hypernetworks, Lycoris, and a few other kind of these things I don't know much about.

[lofty ingot]

I will do this

(the xy graph)

Im going to run a prompt on a few different ones with the same seed

[lapis wigeon]

But I don't know to much myself. It is also pretty small (KB, MB), it trains pretty fast, is for teaching the ai new things It doesn't really understand, but has some kinda similar things in its dataset.

[drifting maple]

Hey guys, just wanted to know of stable diffusion video works , I can't get it to, or atleast it takes a lot of time

[lapis wigeon]

Yep. You can give it an character, style, concept, pose, ...

Textual inversion isn't used much anymore. It is mostly only still used for negative prompts.

[lofty ingot]

Getting a somewhat consistent understanding of a few models rn thanks to @lapis wigeon
If you want i can share with you
🙂

but i think the resolution i had was a bit too low, because i got some weird feedback

[fathom inlet]

Quick question. I'm not heavily into the theory of controlnets. but, theoretically, if I were to split up the process and create a gray scale of my original image that looks exactly like the output of sd-controlnet-depth, then do the image2image using only the output of the first process - will I get the same result (theoretically) as using a controlnet

[quartz seal]

Good day folks, can anyone quickly tell me what the correct tool is to merge 2 images together? I have one main perfect image, and i want to insert an element from another image into the main image. I have isolated the element into a blank black single layer copy of the mainimage matching the dimensions. I have broght the main image into the inpaint tool, and loaded the element layer into control net locking in pixel perfect with full 2 strength. I have then created a mask area on the main image where i want the elemnet inserted and selected masked area only to inpaint. the results are roughly correct, except the quality of the element pulling in from control net side is 1) Not the correct colours and 2) has a distortion blur surounding the element brought into the main image. Is there something im missing here on how to control bringing the elemnt layer in controlnet into the main image painted mask in inpaint?

[lapis wigeon]

Probably Lora with maybe some help with a Prompt

[sly fiber]

hi wanted ask if someone could point to me best model for spritesheet generation

i found not bad this one
SD_PixelArt_SpriteSheet_Generator

[uneven halo]

Sometimes when I enter my prompt it shows up with "Format:Video", but I can't figure out why. I don't want a video. Also those never seem to show up anyway. What can I put in specifically to avoid this?

[lusty plank]

can anyone recommend a good Stable Diffusion model for generating abstract art? I have a lora that I've created that's based on an artist whose style is unusual sort of abstract art, and most models I've tried tend to make it too "realistic" and try to create more detail than necessary. The only model that seems to work is the 1.5 base model but that's not ideal as it doesn't seem to understand some of the concepts I give it.

[eternal tangle]

Hi! 🙂 Quick question.. do I have to use a refiner together with custom SDXL models like RealVisXL v2?

[zealous crest]

im just putting out this call for help here to see if i can find someone knowledgeable to walk me through the process of getting stable diffusion set-up as every time i have followed a yt tutorial i end up getting this error at the very end and i cant get it to work but i bought this 4090 for AI so lmk

[fading dune]

hello

i have been having issues when trying to download and setup stable diffusion

can someone help me ?

[quick holly]

HELLO, CAN ANYONE HELP ME? I WANT TO GENERATE AN IMAGE, OF HANDS HOLDING A WHITE PAPER, HOWEVER I AM NEW AND I DON'T KNOW THE REASON WHY THE BOT PLACES THE IMAGE PLACE IN THE VIDEO FORMAT

[warm junco]

You can set format:image

[warm junco]

Hey, you dont but you can

[fading ravine]

has anyone come across a good documentation on how to use SDXL for short animation/video ? As in text2video, or img2video

[fossil osprey]

anyone in here thats good with animate diff in ComfyUI

[median oxide]

How can I get SD to save prompts?

[pale latch]

with automatic1111 all the settings are embedded into the images. comfy does this too but it embeds the entire workflow.

gallery extensions will let you view old prompts easy

[lusty plank]

Does anyone know why when making a lora model, my sample images look quite good, but once I test the model in A1111, they all have strange artifacts on the face that weren't there in the samples?

It's odd because as the image is generating, it looks quite good and then at the last moment the face suddenly becomes deformed/has strange lights on it.

it's very disappointing as the sample images seemed to have quite a bit of promise (albeit a bit strange looking as most stable diffusion 1.5 base model images are).

[twin kernel]

Guys, on Clipdrop pricing it say something like 1000/24 what does it mean?

[glad jolt]

Did anyone notice chatgpt change the way it talks? I noticed it's gotten way dumber

It now often says "I'm really sorry" instead of I apologize

[smoky galleon]

does automatic1111 webui still have instruct pix2pix? how do you access it?

[narrow tendon]

hello guys!
is there a way to get from a Photo A to Photo B in Deforum ?

[median oxide]

Oh okay, for some reason when I first boot up SD, there is an old set on prompts already in the boxes. There's also a seed too so I was wondering if there is a function that saves them somewhere? Just so I can disable it since these prompts were when used like a week ago but keep appearing

Wtf

Discord mobile loves sending shit multiple times apparently

[zinc sun]

Someone know how if i can use stable diffusion (with automatic1111 gui) running in my computer in my house when i'm outside ?

[bitter olive]

you mean from your phone?

Does someone know which SD models are allowed to be used for commercial use?

[zinc sun]

Not mandatory but i think i will mostly using it from a laptop

[warm junco]

Yes, you need to select the pix2pix model in img2img and then the sliders for p2p will appear

[warm junco]

Try out an other browser. Maybe it got saved as "autofill credentials"

[warm junco]

Yes, you need to add the --share command in the webui-user.bat

Then you'll get an URL

[smoky galleon]

where do you select the model?

[warm junco]

At the top left dropdown

[smoky galleon]

the stable diffusion checkpoint?

looks like I need to download the right model then

[warm junco]

Yes its pretty big (7gb)

[smoky galleon]

yeah that is bigger than usual

[shy blaze]

Why is my prompt word not generating a picture? It's stuck there the whole time.

Why is my prompt word not generating a picture? It's stuck there the whole time.

[maiden parrot]

Hi, I have been playing with stable diffusion on my pc but I only have a 1070 which isn’t the fastest gpu out there. Is there an online platform with comfyui I can use instead of my pc at a reasonable price?

I had a look at cloud VMs with GPU but they are way too expensive

[warm junco]

Do you xformers for better performance?

[maiden parrot]

For now I have been using fooocus but lately I started looking into comfyui but I haven’t tried it yet, I have not heard of xformers

[warm junco]

Ah okay, in comfy and auto1111 you should have the best performance with your gpu

For auto1111 you'll need some additional launch arguments to get them --xformers --medvram-sdxl --no-half-vae

[maiden parrot]

How big of a difference should I expect for standard 30 iteration image generation 512x512? I think it’s currently taking a couple of minutes, but I haven’t timed it

[warm junco]

In auto1111 or comfyui a 512x512 30 steps should take Max 30 seconds

[maiden parrot]

Ok, in that case I will try comfy before looking around for better gpus, thanks

[warm junco]

No problem 🙂

[maiden parrot]

One last detail, I was using sdxl juggernaut, not sure it that matter

[ivory nexus]

Hello! What is the best model to make small human figures against the backdrop of nature (i'm using vector style Lora)
?

[warm junco]

Yes, sdxl models are trained on 1024x1024 and are heavier models (6gb)
Your GPU and others with 8gb vram will perform a lot slower with sdxl models.
The fastest time you can get is like 1:30 for 1024x1024 30 steps with that

[maiden parrot]

The difference between xl and standard is just the image size?

I don’t really need 1024 for what I’m trying to do

[warm junco]

The sdxl model also has a better prompt understanding. But there are a lot of custom 1.5 models too with that you can generate really good stuff

[maiden parrot]

Too many models to chose from 😅

[warm junco]

True xD

[hasty hornet]

it's never too many, I need moar

unfortunately too many of them are basically same thing or mixes that's not that much different

[maiden parrot]

That’s kind of what I expect, I think I will start from the 1.5 model and see what I get out of it, the other model are derived from it so I expect it to be the more versatile one

With all versatility of confyui does it still make sense to use loras or you can get the same result with a good workflow?

[lunar harbor]

Hi, I type my prompt and it says dreaming but gets stuck there. Why is that?

[slow forge]

Is there anyone who is looking for blockchain developer?

[hasty hornet]

base 1.5 kinda meh

[flint marlin]

I'm trying to download stable diffusion but I don't know how to do it.

[hot haven]

What is the exact problem?

[flint marlin]

https://www.dexerto.com/tech/how-to-install-stable-diffusion-2124809/

I have followed this guide but when I do step number 5 it does not work.

I can't send screenshots here

I cant send document

i don't know how to send the error message. where to send it?

[next ferry]

Is it error 1 or something?

[hasty hornet]

why do people follow random articles they find on the internet, while there is an official installation guide on automatic1111 github I don't get it

luckily this one is legit, but man, you are just asking to install some malware

[hasty hornet]

#🏞｜general-with-images you can send images there

[maiden parrot]

Any better suggestions as a generic starting point?

[hasty hornet]

depends what you want, I'd pick something popular from https://civitai.com/ or something that you're interested in

[maiden parrot]

The first goal/project is trying to convert a picture i something black and white (no grayscale) svg style, from what I have seen it should be doable

[magic falcon]

ComfyUI question: how to force same seed for all images in a batch?

[lavish lake]

below the seed you have "control_after_generate", select "fixed"

[magic falcon]

that one uses different seed for each images within the batch, which makes sense with a normal prompt because it would just create the same image many times. i want the same seed within all images

Trying to get Style Aligned to work well

[real zodiac]

the seed is the same for the whole batch but because of how the latent is generated the images aren't identical, I think you can duplicate the latent instead to achieve that

[magic falcon]

Ah... I see

[lavish lake]

it's the same seed yes

[real zodiac]

rebatch latents node should do that

[lavish lake]

"repeat latent batch" will copy the same as many times as you want

[hybrid siren]

How is the paid one different from this bot in discord?

[hasty hornet]

there are like...hungreds of paid sites with SD, which one are you talking about ?

[hybrid siren]

Dreamstudio

[hasty hornet]

you can set resolution \ ratio, they also add new versions to test there

[bronze basin]

Is the best way to create consistent characteres LORA? Like if I draw a character many times in different positions etc, and I create a Lora can then I put a keyword that will trigger that character within a prompt?

[hybrid siren]

Ok cool

[crystal edge]

what the benifits of hires fix?

[hybrid siren]

I've a 1060 so can't use sdxl

🥹

[smoky cairn]

hi where am I supposed to put the loras? I don't have the lora folder in the models for some reason?

[lavish lake]

kinda like that, you make the training data and specify maybe a keyword for that character

[dusty sparrow]

Hi, how can I make the bot respect the text of a logo for example?

Or do you have a link explain that ?

[prime condor]

hi guys, how can I make top/plan view images? it's always isometric, even if I set it as a negative prompt

[shy oriole]

Does anyone here by any chance know if Stable Diffusion at some point will be able to understand and deal with hex color codes as described here: https://www.w3schools.com/html/html_colors_hex.asp ?

[maiden parrot]

Should I use 1.5 models or move to 2.1? I read that 1.5 is supposed to be better but I’m not sure how old the stability article was, maybe it got better by now

[hasty hornet]

theoretically, but will need different approach. Doubt it just knows each possible value meaning \ has training data on it

You can inpaint something with your color and let it cook

[maiden parrot]

I don’t know what you are trying to achieve, but you can try to play around with latent noise to get the color you want

[cedar salmon]

versions are versions, doesnt mean better per say, like i use SDXL stuff which was based off SD 1

[fossil phoenix]

What is the clip vision and IP adapter vitg version. Google is being a pain since it just keeps pulling up other stuff. I guess the search terms have too much weight on the clip, vision and IP adapter. Search terms aside from the actual hugging face and GitHub repos. I really don't find it much information on them

I mean I've tried to use it but I'm not sure how it differs

[fossil phoenix]

It would actually not be hard to train that. You have the descriptions and you have the colors, as well as the color names, if you refined a model using that as your data set, which is pretty much already prepared, you would just have to properly annotate it and train it right. But I think that would be probably one of the easiest approaches. If you wanted to have something where you could use color hex codes

[fossil phoenix]

I mean you could quickly generate color swatches and hex tagged annotation texts, quickly with python, if you want to slurp the hex color names from the website, I'm sure that'd be quick work too, and now you have a data set. I'd tag it several ways, color, shade of C (top level color), dark C, Light C, and/or C1ish-C2 (orangish-brown etc.) and so on. but it would not at all be as difficult as creating a dataset of compelx imagery. tbh it may be not a bad LoRa form colorizer. though you'd prolly need to combine its usage with seg, or something... dunno, actually what you really want to do lol.

[shy oriole]

Thank you for the reply. Nice ideas to be considered.

[strong hornet]

i cant find

site

with Stable Diffusion

ai

anyone have link? i could make photos for free

[earnest fractal]

hey I was wondering how many other SD communities there on Discord and what other ones people know of?

[hasty hornet]

https://beta.dreamstudio.ai/generate ?

[iron imp]

Just use the bots in this discord? #1100170312106127410

[earnest prairie]

Do you know a way of correcting eyes after the image has been generated?

[fossil phoenix]

I use the horror lora, to remove them. They aren't so troubling then.

[fossil phoenix]

what exactly do you mean? a site with? To download and use, or a site that uses SD for you to generate?

[dawn bough]

anyone know of a model in adventure time style?

[earnest prairie]

It’s free ?

[fossil phoenix]

I was joking, don't ever take my advice, you'll find you end up with something clive barker meets george romeroish

If you want to fix the eyes, inpaint works

or img2img with low denoise

[earnest prairie]

But it's not free

[pale latch]

you can get some clipart of good eyes and paste them

low gpu memory solution

[earnest prairie]

Its trash

[earnest prairie]

.

No thx

[vivid palm]

What's the best model if you want to feed it some pictures of yourself and get some professional headshots out of it? Is fine tuning SDXL the best quality or is there something out there that's better out the gate/requires fewer headshots for the same quality? So far it seems like the ReActor extension with SDXL is pretty great for single photo model training, anything else I should look at?

[pale latch]

train a lora on the base model and then use it on a photorealistic one

[fossil phoenix]

lies.

foul, stench infused, lies. now sir, I expect you to apologize to img2img

Or I will img2img / inpaint any images you ever share on this server until you cry at the things I do the them... or I get banned >:O... though you might like what I do, if you're into that sort of thing... some people are.

[hallow plaza]

For some reason when I created my lora I keep getting women with two heads and two noses or sometimes two chests and 1 head. What could have gone wrong while making the lora? Could it have been some bad tags?

[earnest prairie]

@fossil phoenix When I said "trash", I meant I didn't get the result I'd imagined

[fossil phoenix]

oh, okay, that's fine.

you doing it wrong :p

[earnest prairie]

I can show u ?

[earnest prairie]

That's for sure

Software developers 🫡

Respect

[fossil phoenix]

what is your difficulty?

you wanted to fix some eyes ya said yeah? lol

can't show in here though, gotta go to general with images

[earnest prairie]

I have a laptop with a benchmark of ≈623, so I can't install SD on my computer, and I'm in this Serv, but after a few weeks, I see that the eyes are a bit messed up and I'm looking for sites to fix it

[earnest prairie]

I'm going to go to those sites again and make another image with the ones I've got

[fossil phoenix]

well you need something that allows inpainting, probably,

or beautiful eyes lora lol

just don't send me some weird hentai dragon in diapers... some guy did that the other day when he was asking how to fix something on another server... i sent it back peeled open and quite... not like it was before instead...

he did not like 😦

[earnest prairie]

That’s nasty

[earnest prairie]

When it's a site ?

[fossil phoenix]

I don't use sites

i dunno

[earnest prairie]

Dunno ?

[slate dust]

Try this one https://inpainter.vercel.app/paint
There are many free online if you search

But it's definitely suggested to invest in proper equipment if you want to get into it

[earnest prairie]

Watch my CPU #🏞｜general-with-images

[earnest prairie]

Except that I don't have the necessary means yet

[slate dust]

Of course! thankfully we have the bots in this Discord and lots of free online sites

Don't pay for any third party apps!

[earnest prairie]

Thx

[maiden parrot]

What is the difference between control net and ip adapter?

[eternal tangle]

If I use XYZ plot, there are only certain ADetailer options.. by far not all possible. Why? Is there a way to get more?

eg. ADetailer CFG scale and ADetailer checkpoint are missing. Would be nice to have those in XYZ plot..

[arctic coral]

How to fix the Torch can't use GPU error within Stable Diffusion Automatic 1111 Launch? thanks if anyone can let me know:)

[blissful oar]

hi guys, i was wonder, how to control the hand to grab a sword correctly

[normal plover]

Hi all. I create a youtube video regarding a model called Pixart-Alpha. Its amazing (the model, lol,m not the video). Check it out: https://www.youtube.com/watch?v=KcxssceXNXg

LMK WHat you think

[surreal goblet]

Just curious - been away for some time and just started fiddling again yesterday. Do images expire? I just checked in and all the stuff I did yesterday had vanished. Even the prompts. Except the ones that ended up getting voted on in showdown. Looked around but didn'

t find any notice on it.

Funny thing is, I can find old stuff (from like January and before), but everything I did yesterday was gone.

Should I be downloading as i go?

fortunately, I did download a chunk for immediate use, but geeez. Maybe give it 24 hrs before purging or something.

[mental chasm]

Do you guys knows that how to make the generated image present full body

[unkempt hatch]

trained 400 pics. 50 are art, the rest real. 2 pics have someone with rosacea. Now every image generated has the art style and rosacea, but otherwise very flexible, lol

[maiden parrot]

Is that a Lora model or a checkpoint?

[eternal pendant]

Question about the magazine is there a way to get a pdf version with are subscription to it?

[south steppe]

not enough GPU video memory available? i have 20gb of VRAM

[brisk hamlet]

is there a tutorial for training a new model on top of a pre-existing model? juggernaut xl is really good but i want to train it further with a set of images to further target it towards my company's needs. not sure if this is a thing or not but would love to see some tutorials if anyone knows how.

[floral umbra]

Hoi, would you guys know of a local A.I that can transcribe videos and spit out subtitle files? Preferably in batches.

[sly fiber]

yo this feature i saw in the addon for blender called pallaidium

very nice addon !

worked very good for me

it can do even sound and videos even than just image

[floral umbra]

What exactly does it do though? Didn't really show images of it making 3d out of images. More just using blender as a "controlnet for images"

[sly fiber]

https://www.youtube.com/watch?v=j9nV3hgHgQ4&t=13s @floral umbra

oh well maybe i missunderstood correctly what you wanted then

[unkempt hatch]

It was a checkpoint Dreambooth

[cursive gate]

Does anyone have good recommendations for the best online platform that can help a company make some amazing products with a logo?

We would upload a logo and start making products around it

[floral umbra]

This is what i want to accomplish. Like how CNC bores characters and objects out of wood, but i want to 3d print the generated image. either generate cool images, or use existing walllpapers, then make a "normal map/depth map" with A.I of the image, throw it into blender, apply that onto a "plane", then make a 3d "frame" around it and 3d print the image :)

https://fiverr-res.cloudinary.com/images/q_auto,f_auto/gigs/174222414/original/c072903614e4a2bbb49d9d68e3f79cfcf421a8b0/make-3d-carvings-models-g-code-for-cnc-router-machine.jpeg

[sly fiber]

i see , its called parallax mapping

there is addons for blender to do that too

so you can use comfy ui and import the images in blender and then apply the color image onto plane. dubdivide the place heavily, or use cycles microdisplacement. and then render it...

or dubdivide the image and use displace modifier with the depth map and render it or etc...

oh well yeah sry i mean then just use the depth map on the displace modifier... and then export the object to 3d printing programm

sry i shoukld read mroe carefully first before answer XD

you can generate nice images even in blender with opallaidium addon or there is super addon called dream textures

crazy cool

[deft bramble]

main, thats not a worm thats a fucked up alligator

[shell citrus]

What is the refiner used for? I am a little confused about what is does? Does it fix up details such as face or add details? Or does it influence the style?

[surreal pagoda]

the refiner is (ideally) used in the final steps of image creation and handles the formation of the smallest details

it basically adds/improves fine detail

[agile locust]

When it comes to training a lora on clothing, do people mix photos of the clothing in question with drawn ones? Will this negatively impact the output? If I wanted to use the same lora for realistic and anime models for example.

[surreal pagoda]

Hmm, I figure it would be fine, but if you want to be sure then it’s probably safer to not mix art and pictures

[agile locust]

I assumed it would come down to tagging.

[surreal pagoda]

Yeah that’ll probably help

[agile locust]

Like "A drawing of" or "A photo of"

[surreal pagoda]

Honestly I’ve never done any training but I bet clearly differentiating the different media types with tags should have a positive outcome

[shell citrus]

Could you like use another model's style as a refiner? Or is that what lora's as specifically designed for?

[surreal pagoda]

I think loras are better for that

the refiner is more for, well, refining the details, not changing the overall image

[scarlet valley]

hello, how can I use SD to generate pics? where do I make a promt?

[nova current]

How are you

[distant light]

Hi there, does anyone know how the networks / “filters” from the bots are construct or is it just plain sdxl?

[dawn crane]

Greetings! Question that might (should) be too common : where can I find 'courses' for stable diffusion?

[vague pine]

Yo chat, how bad is the low vram RTX 3080 vs the better 3080?

nvm found the answer

oh god why did i have to buy a 6000 series gpu

[fervent thunder]

On the same topic, i think workstation cards are at even better value than rtx

Like p5000 with 16gb vram should be competing with 3080

My ideal pc would be p5000 for SD and amd 6750 for gaming

[raw dagger]

https://civitai.com/models/176555/harrlogos-xl-finally-custom-text-generation-in-sd

ohhhhhhhh you want products WITH the logo

my bad I misunderstood

Probaby Redbubble, or Printful

[hot haven]

Can derivatives of the model (as in: images generated by using the model) can be used comercially? I'm writing a small ebook about AI and for the sake of example I want to (try to) generate every single image in that ebook with SDXL. Since I will be paid for writing that e-book, do I need permission to do what I'm planning to?

[raw dagger]

Check the licensing of the model(s) you'd like to use.

[hot haven]

@raw dagger I did, but I'm not fluent in legalese. The license is from here https://github.com/Stability-AI/generative-models/blob/main/model_licenses/LICENSE-SDXL1.0

The relevant points are (if I understand correctly): "Subject to the terms and conditions of this
License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive,
no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly
display, publicly perform, sublicense, and distribute the Complementary Material, the
Model, and Derivatives of the Model."

[raw dagger]

Yes you can use SDXL commercially

[hot haven]

And: "Subject to the terms and
conditions of this License and where and as applicable, each Contributor hereby grants
to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this paragraph) patent license to make, have made, use, offer to
sell, sell, import, and otherwise transfer the Model and the Complementary Material,
where such license applies only to those patent claims licensable by such Contributor
that are necessarily infringed by their Contribution(s) alone or by combination of their
Contribution(s) with the Model to which such Contribution(s) was submitted"

The first one doesn't tell anything about commercial usage and the second one refers "patent license to make" - that's why I was unsure 🙂

[raw dagger]

oh yeah totally its a bunch of mumbo jumbo 😂

[maiden parrot]

I’m collecting images to train a Lora, i red that you can use just 5-10 images but I’m a bit skeptical, do you guys have experience with training Lora’s?

[vague pine]

idk about dropping $900 for 2 cards when you can just have a 3080 for $100 cheaper

$200 if you go low vram

[marsh vigil]

i don't think low vram is a good idea. it's a pretty miserable experience

[vague pine]

on 12 gigs now

ofc i want 24!

whuy can't they just add it...

5060 on 8
5070 on 16
5080 on 24
5090 on 32

like cmon

IT DOESN'T COST THEM MUCH AT ALL

the only reasons i buy AMD is:

• not tragically overpriced
• more vram

[marsh vigil]

have you tried rebar? i've yet to try it, but i've heard generally good things

[full lark]

5060 should have at least 12 like the 3060 🥺

[vague pine]

top model innit

contact the EU to pass a law to have at least 32 gigs of ram on entry level GPUs

[broken dagger]

hey guys, i was wondering if someone can help me to identify a specific art style to train my ai to reproduce it

[maiden parrot]

I have been out of the loop for a while, is AMD good for AI? When I was working on ai nvidia was the only player

[vague pine]

directml but expect 6000 series to be 1/6th the performance of 3000 series

better on linux i assume

cause rocm

[maiden parrot]

So it’s better than it was but still nvidia is the go to

[vague pine]

yea

they got a massive lead with cuda

[maiden parrot]

Comprabile in performance or price?

[vague pine]

Comparabe to buying 2gpus? no.

[stray tangle]

How can I stop the bot from making videos?

[remote glade]

Guys I have a question

What's the difference between Latent Noise and Fill

Don't they do the same thing?

[unkempt hatch]

you know how sometimes compositions get weird and disorienting? Like the angle of a door doesn't make sense, or a windowframe blends into a broomstick? What would be the best negative for that?

[remote glade]

controllnet

just Inpaint it

[unkempt hatch]

but usually inpaint looks even worse. I'm talking about the entire image''s compositioon being a bit off

also

like, a wide-angle lens mixed with a panorama

not that bad but just... the composition is a little skewed

[remote glade]

You need to master the art of Inpainting correctly with controllnet

@unkempt hatch use MLSD it tends to make straight structure designs

[unkempt hatch]

oooh, right, good idea

I don't need the controlnet .yaml files, right?

[unkempt hatch]

hmmm, maybe. I never really tried ther inpaint controlnet.... I'm guessing it's way better then

[magic tangle]

hi everone

ı have problem

ı cant find in stabledif Settings/ userinterface- quicksettinglist > initial_noise_multiplayer

how can ı fix this

[signal dome]

24 gb vram yea rip my 2060

[last dagger]

the link in the announcement doesn't work

blog post 404

looks like it's missing /news prefix

fyi @karmic brook ^^

[random pumice]

I added it and still doesn't work

[barren kraken]

https://stability.ai/news/stable-zero123-3d-generation

[karmic brook]

the link is updated and should work 🙂

[random pumice]

Oh there in the link

aight

[snow spindle]

Is this real about the 3D ?!

This is huge

[simple heron]

hi is a NVIDIA GeForce GTX 1060 6GB powerful enough?

[marsh vigil]

@simple heron i'm currently working with a 1660 TI 6GB until i replace some hardware in a few weeks. a 1060 is "enough" in some sense that it won't be quick, you'll be limited to working with SD 1.5 (SDXL grinds to a halt on these cards and frequently just fails), and you'll probably be limited to using at most one LoRA at a time. you will be able to generate images, but it won't be all that comfortable of an environment

[simple heron]

thanks for the response. I guess I have to work in civitai then?

i'm new to this

[marsh vigil]

i think you could grab https://github.com/comfyanonymous/ComfyUI and try plugging in a few of the workflows and see what happens. i'd suggest https://github.com/AUTOMATIC1111/stable-diffusion-webui but in my experience it's a pain to get it working properly on the 10* or 16* cards

the stable diffusion 1.5 model should be enough to play around with, initially: https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main

https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.safetensors

primarily you'll be limited by the fact that 6GB isn't really enough for larger models like SDXL, and the relatively slow compute speed of the card means that you'll be waiting around for fairly long periods if you start going above 512x512 image sizes

[full lark]

I have a 1060 3gb and I can do 1024x1024 on sd 1.5 in less than a minute with lcm lora

[marsh vigil]

@full lark i'd consider "on the order of minutes" to be slow 😉

i suppose it's a matter of perspective

[full lark]

It is slow but still XD

It works

[marsh vigil]

yep!

[full lark]

I mean I used sd 1.3 or 1.4 on cpu only, I had to wait like 3 minutes or more

So im used to wait 😁

[potent spire]

Damn 24 gb VRAM

[native sparrow]

Greetings, fellow creators. I want to ask if any of you know how to mantain consistency in a face without using ReActor or any other faceswappers out there. Should I just train a checkpoint for my model, try specific prompts (I'd rather not), any other ideas? I appreciate your help.

[marsh vigil]

@native sparrow i've had the most consistent results with masking and LoRAs, especially if there are multiple subjects in the scene

i think it tends to work best in a kind of multiple pass scenario... generate an image with the subject (or subjects) that has the composition that you like but with generic faces, and then repaint the faces with masking and LoRAs

[native sparrow]

@marsh vigil great, I'll try that. Tyvm!

I was wondering, is it possible to train a checkpoint of just one model? I mean, use multiple images of the same character to train a checkpoint that will most often create images of that character (as specified in the prompt) in different scenes and/or situations.

[marsh vigil]

yeah, that's exactly the kind of thing that LoRAs are intended for

[simple heron]

Thanks for the response. I'm open to any tips on making a slow system faster

[marsh vigil]

LoRAs are sort of like overlays for existing checkpoints. they'll often contain just a single character that can be triggered into appearing in a scene by using their name in a prompt

[native sparrow]

oh, that's exactly what I was expecting. Awesome, thank you very much!

[marsh vigil]

nice thing about them is also that they tend to be less than 100mb or so!

[native sparrow]

it will also create unique images with personality, that's what I've been trying to do for days now. Good thing is that I learned a lot in the process. Now it's time to master LoRAs lol.

[static badge]

please what arguments are the best for fast and quality on a 4090?
so far set COMMANDLINE_ARGS=--opt-channelslast --opt-sdp-attention
no half vae thingy as well?

[paper lance]

yo

how can i use stable diffusion to generate a headshot based on a provided reference image?

[warm junco]

Yes its usable for SD

[warm junco]

Make sure your using xformers too

[warm junco]

For best performance you need --xformers --medvram --no-half-vae in your webui-user.bat at the line COMMANDLINE_ARGS=

[warm junco]

Use --xformers instead of --opt-sdp-attention as it uses less vram.
Also --no-half-vae is needed for sdxl models with VAE or anime VAE

[wet cosmos]

Hey im new here

does anyone have tips on generating things on FOOOCUS

this is the model on there, and im running it through AWS so that i have faster speeds

juggernautXL_version6Rundiffusion.safetensors

[hallow plaza]

For comfy UI how do I get mroe samplers? I dont have DPM 2m Karras

[muted steeple]

Hey community, how to use seed id? I want to redo certain images, but i am struggling to find the proper command. Help 🙏🏻

[torn coyote]

I've just started to learn deforum and I'm realizing that you can do so cool animations without having to know after effects or anything like that
but still, deforum is also quite hard. Does there exist a solution for making something like deforum more accessible?

[south basin]

hello everyone 🙂 anyone training diffusion on their own data ever experienced fewer steps leads to better quality? like consistently performance (FID) increases from 1000, 500, 200, 100, 50 to 20 steps (but starts decrease since then) that's so weird

[rich dock]

anyone have any idea what happened to the dynamic thresholding cfg scale fix extension in A1111 -- it hasnt worked for a while now and old gens give an error when i try to regenerate them

is there a better replacement or something?

[raw dagger]

https://civitai.com/articles/554/dwh-1-creating-your-first-animation-with-deforum-in-1-click

[quasi drift]

NeurIPS anyone? last min crashing. hosting an AI3D breakfast tmro!
https://twitter.com/Yosun/status/1735091122202890697

[fallow veldt]

What is people using to create video with Automatic1111?

[oblique meadow]

Hello everyone. I have a question regarding the license when we use Stable Diffusion to build an AI-generated app. Does anyone have any experience with this? Can you discuss it a bit, since I have no idea how to use some models from users on CivitAI or other sources?

[lavish patio]

hey all, any suggestions on a good model/lora for img2img to convert real images to illustrations (non-anime)?

[simple comet]

Could you upload an image and a program tell you what's the prompt ?

[native sundial]

https://replicate.com/methexis-inc/img2prompt?ref=taaft

[simple comet]

HOLY SHIT TY

[keen parcel]

Hey

Is there a Lora for SW podracers ?

What’s the latest best notebook to fine tune SDXL on it otherwise

[dusty sparrow]

😱

[bleak matrix]

Good morning, everyone!

How's everyone today?

[ornate tartan]

what a dogshit horrible idea

$20 a month, what a joke

[wintry stream]

I think all feedback, both negative and positive is very important so thank you for taking the time to give your thoughts, Jumblkayla.
Could you let me know what parts of this membership have you most concerned so I can review with our larger team?

[fervent thunder]

yeah what a stupid idea, withdrawing commercial use of a previously open source and free product

especially when the work is so community driven

without the community you wouldn't have a model to begin with

[flint eagle]

for that membership could blow up like wizard of coast dnd company or unity did

[rough heath]

I really don't like the corporate tone Stability is taking

[flint eagle]

we all remember what happend to them

[visual ledge]

well there goes all cloud AI generators. the first big money grab is taking hold. it only gets worse from here. get your models now while you stil can!

[fervent thunder]

yea time to get a 10TB drive and hoard models

[wintry stream]

Fair enough, is it mostly about the wording itself or is it the messages that are being conveyed that don't feel as genuine?

[sage talon]

My first concern is that the actual terms of the commercial license should be front in center, and I don't see it anywhere.

My second concern is the lack of a specific cut-out for open source projects.

[rough heath]

It feels like I'm being talked to like an investor and not a community member

[wheat anchor]

stability ai is no longer a supporter of open source as it does not adhere to "free-as-in-freedom" principles anymore

[wintry stream]

Thank you - grealty appreciate the thoughts ❤️

[fervent thunder]

exactly

this is against the core idea of open source, this is freemium

[rough heath]

Stablity wants growth

like all companies think

[fervent thunder]

there are other ways than to revoke commercial use from the licences retroactively

and paywall it

[rough heath]

Like Unity

[fervent thunder]

its the mistake unity made and almost tanked their corp into irrelevancy

[drowsy hamlet]

Looks like the models that were ok for commercial use prior to today are not part of the membership list

[wintry stream]

Thank you!

Most of the specifics about the membership and it's terms are on the website here: https://stability.ai/membership

Unfortunately it's just a little much to include in the post for our community here so we didn't want to overload with jargon and make the key points difficult to extract from the post. Apologies if this didn't come across properly

[shut moat]

looks like SD 1.5 and 2.1 are not listed under core models https://stability.ai/core-models

[flint eagle]

just want to remind the stability team what happend at Wizards of the Coast DnD want to change their licance term to wants so much money or Unity wants to change licancing they destroyed in days and made altranative free solutions
I suggest reconsider before doing this.

[latent bay]

Not the same person, but I am a small scale artist sometimes doing commission work. This comes out to maybe 100$ a month, or less. So far the usage of SD and models based on it have been an amazing boon on that front. But between a fee for SD and fees for other art programs/tools needed to actually create fully refined works, there would be nothing left.

I might be in an incredible niche here, as someone who is making money with this, albeit at very small amounts. And yet for me this would bring things to a point where doing commissions would not really be worth it anymore. Which is a pity, as I am passionate about this, but can't afford to basically work for free (or passing all of the small revenue I have onwards) as deserved as it would be.

I do understand the step as it is taken. But it is hurting small scale creators like me tremendously

[sage talon]

Thanks, but I mean an actual license, not marketing. Give us the legalese. Apologies if it is there, and I just don't see it.

[rough heath]

At that point just run an SD locally

[wintry stream]

I think your concerns are fair, and they are coming from an honest place, so thank you for taking the time to make sure they are voiced.

What I will say on a personal level, is that the situations are difficult to compare and us having a way of funding our models to ensure we can still continue to release them is quite important.

That said, I don't mean to discredit your thoughts at all, just sharing my own opinions to add some depth.

[rough heath]

And avoid giving your hard earned dollars to Stability

[wintry stream]

Gotcha! Sorry if I misintepreted.

Is this what you're looking for? https://stability.ai/professional-membership-agreement

[flint eagle]

please don't understand my post wrong I am just saying internet saw and everyone on YT or devloper keep talking non stop and their comminty fall apart in lighting speed. Just be careful on this. This can be create so bad backlash

[humble ravine]

hello

[rough heath]

So in other words you're hurting for money

[wintry stream]

I think you present an interesting situation that raises some valid concerns.

What I will add, is while the standard price for our "Professional" level starts at $20 a month, we are most certainly open to support people who find this a barrier.
If you contact us with your proposal, we are invested in making sure we can find a solution that fits and doesn't take away from the ways people can use this technology to innovate (https://stability.ai/contact)

[latent bay]

I am, but if this is the way models will be licensed in the future, it would still be affected unless I am missing the direction this is going

[rough heath]

Basically a service where you will pay to rent something you will never own?

[rough heath]

20 dollars a month?

[drowsy hamlet]

@wintry stream it's unclear how the professional license fees work. For example if you're a small business running web services that run a stability ai model on the backend. Do you pay just 20$/month? 20$/month per developer? 20$/month per paying user? 20$/month per instance of models you're running?

[rough heath]

That ain't cheap

My locally hosted instance using a 200$ GPU would be priced out in only 10 months

[worn iris]

If $20 a month is too much for your "business" - I guess you should close it.

[wintry stream]

Per month, per company.

(With some possible discretionary credits available depending on the situation presented when someone reaches out to us - https://stability.ai/contact)

[rough heath]

Well if I was running a business I wouldn't be using Stability

I would be running it locally

Less cost in the long term

[ornate tartan]

i just think the monetization of something that was once free is ridiculous. paying for commercial usage?

all this is going to do is encourage people to look for other ways of acquiring this. this will not affect me as I run SD locally, but it's giving people with crappy hardware literally no ability to enjoy stable diffusion's commercial use without paying. poor people often have worse specs, and are now being punished for those worse specs by having to pay for a $20 monthly membership?

extra context: bad specs cant run SD locally, or at least comfortably

take a look at @latent bay's post. the people i mention here is literally existing right this second.

[wintry stream]

We took a lot of feedback when planning this to make sure we kept the cost as low as possible. But, I also want to point out we recognize that a flate rate may introduce some affordability issues based on personal and geographical situations. We have some plans in place to help with credits and grants as well if this cost is a prohibitive factor for some.

[worn iris]

Depends on your energy cost I guess. Anyway, this does not concern 99.99% of poeple 😄

[drowsy hamlet]

Cool thanks! I undertand there will be community backlash. But ultimately research needs to be paid, training models isn't cheap either. Though SD 1.5 and SDXL saw massive improvements built by the community. Hopefully this doesn't make this partnership suffer.

[rough heath]

No because most people don't even know what AI is

[worn iris]

Backlash coming from weirdoes that generate nudes of girls that are too young mostly anyways. Though it doesnt concern them,,,

[ornate tartan]

bro... what?

[rough heath]

Your argument?

[ornate tartan]

actual projection

[latent bay]

As I said before, I am running SD locally by now. Though that wasn't always the case and at points this would have mattered to me.

My worry now is mostly that this direction might affect the licenses under which future models are released. But as I also said, this is an edge case for small scale creators and I do also understand that some form of monetization might be necessary in the future.

[rough heath]

Seems oddly specific

[ornate tartan]

yeah..

[worn iris]

Well, you go to civinity AI and see what models are most popular

[ornate tartan]

lol

[wintry stream]

Thank you, Louis!

To be completely upfront, I also expected some people to not be incredibly happy about this news, but I think as people read into the reasons and try to understand our position it will help keep us all on the same level.
Feedback is always important. I don't think it would be healthy for us if everyone in the community agreed with every step we take. Being able to read people's thoughts and concerns is so incredibly helpful and really lets us understand what is the most important aspects we should always keep close.

For myself, I'm really, really happy that we get to introduce this with a solid current and future commitment to open availability for non-commercial use. There are a lot of people out there who are incredible innovators and can use our models to make amazing things. As you can imagine, this is both really exciting but also something we fundamentally want to support.

It's can also be hard to understand from an outside perspective that models cost money to make, so there needs to be some give on both sides for an honest relationship.

[flint eagle]

@wintry stream
I got to add this I never couldn't earn anything from AI for 1 year and I got very awfull hardware even AMD engineers need to do special torch version because of my backlash for ROCM + Pytroch + AUMATCI1111 + Invoke github
and my country economy crashing me. 20 dolar means to me 20 ABD Doları =
579,28 Türk Lirası for month which is so overkill to me. If its like YT premuim for yearly like 579 tl ( yt make regional prices but steam + EA move tl to $ but still steam saying will regional pricing )
so I am asking after all this awful condions what Stable ai team expect from me depress me more?
Please don't understand me wrong but we aren't my salary monthly 14.914 Türk Lirası =
514,53 ABD Doları. So you can understand my status.

[boreal marten]

Is the new free membership system applicable on dreamstudio?

What's new?

[flint eagle]

and I want also add I need thx to stable ai for 1 year make my life more better but after this bit sad sadly

[wintry stream]

If you are using these models for commercial purposes, please do reach out to us at https://stability.ai/contact

We understand that a flat rate $20 does introduce some concerns with affordability especially for specific regions. We have some opportunities for credits and grants available if you can reach out to the team with your proposal to see how we can assist you.

And, just for clarity, non-commercial use will continue to be free as always with this new membership.

[rough heath]

"(j) No Technical Assistance or Support. No Order includes maintenance, support, installation or training services. Any assistance provided by Stability hereunder will be in its sole discretion and without liability or risk to You, and may be subject to additional fees."

So you may have to pay for something that isn't even backed?

[sage talon]

@boreal marten Doesn't look like membership applies to cloud services.

The Stability AI memberships are specifically designed to give our members commercial rights for self-hosting our Core Models in your Virtual Private Cloud (VPC), locally on-premise, or on your edge devices. It does not apply to usage of our cloud-hosted services or products, which are powered by the Core Models but which are run on the cloud infrastructure of Stability AI.```

[boreal marten]

Mostly local / self hosted then

[rough heath]

So you have to pay to host it locally?

[sage talon]

If you're making money, prolly. And that's where I'm still digging into their info. I'm unsure what exactly constitutes commercial use. Does an open source developer with a patreon owe $20 per month?

This is why specific language for open source projects would have been nice.

[latent bay]

If you are using it commercially. That is my understanding and why I brought up concerns. Especially for small scale creators this could be an issue

[rough heath]

It was free lol

[latent bay]

was, yes. Exactly why this is a concern

[rough heath]

Plus what stops a larger entity from just using the model unethically without paying?

They might get sued?

[sage talon]

If an open source project like The GIMP wanted to integrate SD models, this seems like it would be a no go.

[rough heath]

So it is gimped

[mighty patrol]

Where can we see the new commercial license terms, the one that we're getting the models under with the Pro level membership?

[sage talon]

https://stability.ai/professional-membership-agreement

That link should be on the announcement. Dunno why it isn't.

[latent bay]

Yes, that is usually what happens if you break a licensing agreement. You open yourself up to being sued.

But that is exactly what I meant earlier with my personal example. I mostly do this as a hobby, but there are people who are coming to me for art commissions. Which is something I am happy to do, but not for free. And 20$ off of a 60$ commission is a pretty substantial chunk, making this unfeasible.

It hampers small scale artists who want to work with Stability AIs models.

What I'd like to see is a low barrier under which things can fly by and if it's just 100$ per month or under or something like that.

At the same time @wintry stream already said to reach out to them in such cases. That option should be made more well known/obvious then maybe

[wintry stream]

You got it!

We're very committed to negotiating and speaking with people who have proposals for our models and making sure they are still able to innovate and bring good into the AI space, even if the $20 price is a bit high for their individual situations.

[rough heath]

Especially when it was free

[real acorn]

is there model for wide screen size ?

[hallow plaza]

When creating my own loras I often get random looking eyes. what causes this and how can I fix it?

[drowsy hamlet]

That's my feeling as well. 20$/month for a company that's hosting the model online for others and making revenue that way seems fair. But for a small content creator that might occasionally use one of these models (on their own hardware, usually on open source software besides for the stability model weights), 20$/month seems like a bit much. And will likely push small creators towards cloud services. IMHO this would limit how far these models get pushed. SD 1.5 and SDXL got improved by the community because everyone using a local install can just tweak to their hearts content - not so on cloud services.

[latent bay]

Pretty much, yes. I am all for Stability AI getting money from companies who make substantial amounts based on those models.

The issue I want to make a case for are small scale creators, making less than a hundred bucks or a few hundred at most with this as a side gig or simply to make a bit of revenue of a hobby instead of working for free.

Because 20$ off of a few thousand, who cares, but 20$ off of 100 is a pretty substantial part.
And hell, even creators who make substantial revenue from this. If I'd get in like 500 a month through this, I'd be more than willing to send the 20 onward. At that point it's a small part.

[rough heath]

2 percent versus 20

I think the license fee should be sliding

@wintry stream

[wintry stream]

What way do you think would be best to implement this?

[mighty patrol]

(d) Providing Your Services to Customers and Users. You will be liable for any acts or omissions by Your Customers or any Users in connection with their use of the Software Products, including any acts or omissions that would constitute a breach of this Agreement if committed by You and any such acts or omissions will constitute a breach of this Agreement by You.

So, I have a question -- I'm working on a 2d->3d art pipeline that outputs fully rigged characters. I'm using SDXL (unmodified) in part of that pipeline. This is in the "hobbyist prototype" phase, I don't run a business currently and am not making any money. I could try to monetize this, maybe selling the output, or a "unity/godot game ready" model generation service. Or I could open source it. Or I could sell the source code that I've developed to a 3rd party, and tell them that they have to obtain SDXL via a separate license from stability.ai. I don't have any problem with paying the $20. But if I were to do that, and the 3rd party that I sold my code to failed to properly obtain a license from stability.ai, I am liable? I don't have control over what "hypothetical 3rd party company" does, I would definitely explain the license situation, but if they did the wrong thing internally, I'd be on the hook for damages?

Just a thought experiment

[latent bay]

I dunno if a slide is really feasible. But my prior suggestion was a low monthly amount of revenue under which no licensing fee would apply

[rough heath]

Maybe a percent

Like 5%

[latent bay]

5% would work for small scale creators, but likely cause issues with companies that have far higher revenues

[tall frost]

I have already subscribed as a member!😆👍

[wintry stream]

We do have some considerations in place for this!

If anyone finds the $20 a month cost prohibitive to getting started, we hope they reach out to https://stability.ai/contact as we're prepared to negotiate with credits and grants based on individual situations and proposals to make sure people can still take advantage of our models and technology

[rough heath]

Not my issue

10000 they would only have to pay 500

If someone only made 20 then they only pay 1

500 is nothing

to a company that made 10000

or 1000 for a company that made 20000

[flint eagle]

this is so like unity story

[rough heath]

Well if Stability wants to hurt small creators and large businesses equally then...

[flint eagle]

well I hope they reconsider this at least they need put more limit to cover small content creators which make stablity ai more popular and free training

[sage talon]

This would be like Unity if they were retroactively changing the model licenses. AFAIK, this is not the case. The models have always had a non-commercial use clause. The writing for eventual commercial use cases was always on the wall.

[latent bay]

I have sent off an inquiry. If allowed I would post updates about the process and how things play out here in the future, as this could be of interest to people in similar situations?

[spark vale]

How to get into the professional membership discord?

[lavish lake]

if you have a membership you'll get access

[spark vale]

Just got it. Do I get an email or something? How do I connect it to my Discord?

[wintry stream]

The server has a few things getting set up, so you will receive an invite (likely by email) in the near future!

[storm shard]

There should be a personal (free) plan for people who occasionally sell their art. You can't compare them to businesses.
Midjourney would be cheaper for them, since there's unlimited use of their servers for 30 dollars a month, while the 20 dollars SD charges will only give you the right to use it commercially. That's a big gap for those 'tiny sellers'.

[storm shard]

Could you please elaborate on what's the benefit of the free membership when you're not using it commercially? @wintry stream

[hot haven]

I have a weird feeling that I need to ask my question again: "Can derivatives of the model (as in: images generated by using the model) can be used comercially? I'm writing a small ebook about AI and for the sake of example I want to (try to) generate every single image in that ebook with SDXL. Since I will be paid for writing that e-book, do I need permission to do what I'm planning to?"

The SDXL license seems to be unchanged for 5 months and I downloaded the version with that specific license, so I suppose the old rules still applies?

[lavish lake]

the new membership model is for the new models such as Turbo, Stable Video, 3D, etc.. to my understanding you can use 1.5 to SDXL and derivatives with no problems

[pale latch]

If i'm not mistaken, stability intends to continue releasing base models with open licenses that can be used commercially. It's more specialized models that will be under new licensing terms. Maybe i'm misunderstanding.

[lavish lake]

here you have specific info about what models operate under the new licensing terms and the ones that will stay on their individual license as before https://stability.ai/core-models

[pale latch]

I know that if they suddenly restricted commercial use of sdxl or sd1.5, sites like civit would have to shut down until they worked out a deal

[tulip plover]

Coulr you get me a working download link pls? Been chasing those links too long

[full lark]

Hello! Do you mean a link to install stable diffusion? if you want to try it withouth installing you can use the bots channels for example#1100170312106127410

[sudden ruin]

SDXL is not part of the membership from my understanding

[shrewd otter]

@wintry stream This is the thing with Membership Tiers.

• First of all I'm glad that on the face of it, it (the scope of the announcement) is not as horrible as the recent Unity Pricing fiasco. I see some good attitude in here & not rigidity & threats (like in Unity's case). I must appreciate that.

• Wider Persepective: I'm a bit saddened that a survey was not done before deciding the pricing of the membership. Talking helps companies to avoid upsetting their community by making 'unbalanced' (or in worst case 'outright stupid') moves

• Shorter Persepective: As pretty much clear from some of the responses above, for small Hobbyist, Indies, .. even the 20$ a month (100-120$ a year) is expensive, especially outside of US, say for someone in a small (Per-Capita Income) african/asian country, etc.

Even in US/EU/etc .. a set of 'CREATIVITY' tools like this should NOT be put OUT OF THE HAND of POOREST of the poor.

I don't want to get in the whole 'We need money to survive as well' discussion, because I have been there multiple times, & have seen multiple companies wreck their own product, because they priced their product with good traction into the ground, thanks to their 'rushed' pricing.

Like someone mentioned, I absolutely don't want a derivative site like civit to shut-down.

TL:DR -- Keeping the Models free for those who have a revenue less then 1K $ a month, seems reasonable. Directly going from Free to 1Million $ a year tier is just wrong. There is a huge ground in between these two tiers.

[lavish lake]

Fair point and I'm sure @wintry stream will gladly address your concerns, just hopped real quick to say that this has been a public discussion for the past few weeks. Emad himself explained everything on X/Twitter and asked about pricing, etc. since last month. I think the solution is a fair one and it will hopefully be completely clarified soon : )

[south basin]

hello community :)) anyone knows except for CM/LCM, which one-step(or few step) diffusion method provides training code? thanks in advance

[forest sluice]

@bleak matrix

[full lark]

It happened to me, I trained 2 loras some months ago, the one with like 300 steps was decent, and the other one with 1000 steps (but more images) got completely broken 😵‍💫

[south basin]

mine generates ok results at 1000 steps just not as good as fewer steps...

[calm flax]

Yo

I just downloaded 4x-ultrasharp upscaler

in what folder do I paste the python? anyone