💬｜general-chat

[vestal dew]

Im using openpose with A1111...it isn't doing a good job of sitting poses. The result is people standing in odd poses. Is there something better than openpose,,..DWPose maybe? If so, where do I get the model?

[iron slate]

i know there is, but how exactly do i make ComfyUI output Images as high quality as Auto1111. It always looks a bit funky or cartoony when using ComfyUI

[burnt kite]

Which SDXL? Checking on civitai, there's SDXL 1.0, SDXL 1.0 LCM, SDXL Turbo, SDXL Lightning, and SDXL Hyper.

[warm junco]

all are sdxl as sdxl is a base model

SDXL 1.0 was the first one and the others are based on it

[burnt kite]

Okay

[warm junco]

you can use any model

[burnt kite]

What resolution should I aim for? It's not really worth it if I can only make images up to 250x250px.

[pearl oyster]

Sdxl best resolution varies based on the model, but all are at least 1024x1024

But you need a gpu capable of it

With zluda, a lot of amd gpus are capable of working with sdxl reasonably well

Zluda is a godsend for ai on amd

[iron swallow]

i can do like 1920 x 1080p and even higher res with flux dev

and i only got rtx 3060 12gb

[foggy yacht]

anyone know how to get flux in kohya? like the sd3-flux.1 branch?

[pseudo bough]

Hey does anyone know if there are any open source voice tools that do what RVC does that have been developed recently just seeing what new options are out there

[acoustic dagger]

The goyim will never know

[iron swallow]

correct

there are probably AI tools out there that are decades more advanced than what is open to the public

that the elites can use

[novel vector]

how to geenrate

[low moon]

stare at screen intensely and picture in your mind what you want to see

[lone sky]

Hey everyone! I'm new to Stability and was playing around with the API. I'm curious—what are some good negative prompts for image-to-image generation with Stable Diffusion 3?

When I use image-to-image with a prompt, the avatar ends up looking like a completely different person if the strength is above 0.35. I'm trying to keep the person the same and just change the background. Any tips on how to do that? Thanks!

[low moon]

comfyui - sam select node - profit

[pseudo bough]

Are there any new or relatively new open source video creation out there

Just not something hopefully that takes a huge amount of comfy knowledge

[low moon]

cogvideox

[crimson badger]

Is it possible for me to make stable diffusion use a reference image to generate that same character in different poses or do I need more images on that character at different angles?

[slender panther]

hi

[summer bison]

Hi

[slender panther]

/prompt

[hard swift]

gm all

[errant ocean]

Yes, it is possible to use Stable Diffusion with a reference image to generate different poses of the same character, but there are some limitations.
If the goal is to create more accurate or detailed poses, especially from multiple angles, providing several reference images of the character from different angles or perspectives will help the AI generate more consistent and realistic results.

[crimson badger]

What if I only have one to start with? Can I use that to make mkre angles using it as a reference image? And am I right that the proper reference image control is imgtoimg? Or habe I been doing it completely wrong?

[errant ocean]

If you only have one image to start with, it is possible to use it as a reference to generate more angles, but the results may not be highly accurate or consistent with the original character’s details.
You r correct that the proper method to use a reference image is through img2img, i think

[fervent thunder]

So what is your bet on whats gonna happen, when ai can make humans redundant and it is in the hands of power hungry megalomaniac elites

[fiery wing]

Anyone know if there are any OpenPose techniques or extensions or models or whatever to not only pose people, but place objects? Or any way to best achieve that? I am having an absolute pain of a time trying to specifically get a character with "sheathed sword on hip with character grasping handle" and I cannot get it to work for the life of me

[pearl oyster]

There might be a LoRA for the sword thing. And if not, you can probably train one yourself, if you have enough reference images

[fiery wing]

Yeah I looked around and didn't find a Lora that worked well enough

[sudden oxide]

I came here to ask the same question. Stable Diffusion has "Style", "Character", "Face" and "Object" training but I ideally could use one for Poses.

[vague kayak]

I'm looking to get back into AI gen. anyone got a good youtube video to recommend me for best options and possibly installation isntructions?

[willow hound]

hello everyone，Im a new student。。

[pearl oyster]

Depends on your gpu in part. You have more options with nvidia than with amd. Also depends on your os

[vague kayak]

Win 10/Linux/ 3060 ti

[oblique jay]

Hello, I was wondering what models I should run for SD... For example, on CivitAI, which models would have the most checkpoints? Which would be good for generating images the quickest and which should I use for quality?

Also if I am wanting to run an SDXL checkpoint, do I still need to install SDXL Base and Refiner?

I've noticed this like SD 1.5, 1.4 and 2.1, but I do not know the differences. Same with SDXL Turbo/Lightning/Hyper etc...

I think my main concern is downloading the base/refiners, is this needed if I using something SDnext

[fervent thunder]

SD 1.5 has the most checkpoints

followed by SDXL

followed by Flux Dev

no other model has a ton of checkpoints and loras on the scale of those three

Lightning/Hyper are about acceleration, and they are in either checkpoint or lora form

SDXL refiner is very rarely used these days

[pearl oyster]

Except possibly pony

[fervent thunder]

okay yeah

[pearl oyster]

But with that one you have to be careful with what you ask of it, to be able to post it here

[pearl oyster]

You also don't need to download the base model to use a checkpoint of it. You just need to download the checkpoints you want to use. @oblique jay

[oblique jay]

Thank you!

[oblique jay]

Is this because it is NSFW/

[oblique jay]

Thank you for this information 🙂

[pearl oyster]

It is capable of it, yes. You can get sfw images out of pony, but you have to be careful to be specific

[twin thunder]

hello!

[oblique jay]

What is Flux?

[pearl oyster]

A type of model.

It's significantly different from SD 1.5, SDXL, and Pony

And not everyone can run it, or even a cut down version of it

[oblique jay]

Gotcha, so it's a more intensive model?

How many GBs VRAM?

[pearl oyster]

The full model will max out the vram of most graphics cards, except the highest end cards. Even heavily pruned models are still as fat as sdxl models.

[oblique jay]

I utilize an A6000 48GB for this, is this still not enough?

[pearl oyster]

A 48gb vram card would qualify as highest end. I was thinking more about consumer grade cards, which mostly cap out at 4090/7900xtx atm. Your card is more of a prosumer card, which is not what I was thinking of.

[oblique jay]

ok gotcha that makes sense

But it probably would not work on a 16GB card

[pearl oyster]

A heavily pruned model can be as little as 6gb. But it can easily get over 16gb if you want to try to go for the full model. The 48gb card will be able to handle it, but most consumer cards will want to get a pruned model to be able to work.

An fp16 unpruned flux model can be as large as over 30gb, coming close to giving even your 48gb card a hard time, @oblique jay

[hollow current]

anyone work with the qualcomm ai hub version of SD?

[oblique jay]

Interesting, thank you!

[gaunt panther]

hi

[full lark]

Hello!

[fervent thunder]

Hallo

[open notch]

Hi!

[low moon]

aloha

[hard swift]

hello all

[low moon]

therefore drink your prune juice.

[vocal pivot]

👋🏼

[true turret]

what's a good WebUI and model to locally generate text to speech? is there an equivalent to SDXL I can do locally so I don't have to abide by these stupid limitations websites impose?

dunno what the requirements woud be, I have 16GB RAM, 2070 Super with 8GB VRAM and for CPU Ryzen 9 3900X

if I could get something on the same level as this
https://elevenlabs.io/

[slow quiver]

Hey

Can anyone tell me if the image I generated is good enough for the flux dev fp8 0model? I can send in dms cause I don't know if I'm allowed to send it here

[golden valley]

Who wants to work as a moder or developer in my project?

[oak zodiac]

Hi

[quartz siren]

Lots of good options, I doubt raw output would be good as elevenlabs though. You can try fish speech(extremely fast, good voice cloning, streaming support), VoiceCraft(bad installation process but great voice cloning but slow and no streaming)

CosyVoice supports good Voice cloning, streaming, is fast, and is even promptable(you can make it have different pitch, male-or female, slow-fast)
Parler-tts only supports prompting but is much better at it.

[oblique jay]

What are some of the best Open-Source LLMs?

Looking for ~7B

I use an Arc A770 16GB

[quartz siren]

Probably llama 3.1 for general usage.

[oblique jay]

I was looking to use it for story generation to make a DND campaign. This would still be good?

[quartz siren]

It should still work well, but there are more creative finetunes you can try out.

[oblique jay]

Are finetunes something you can find and implement, or are they something you need to train yourself? Or both

[knotty turtle]

ControlNet Question: I have a word that typed out in Photoshop using a chunky bold font on a transparent background and exported as a PNG. I'm using this image as guidance in ControlNet using SDXL with the intention of prompting to transform the letters of the text into different materials (eg. stone, fire, etc.), with the goal of keeping the background black or white. However, I can't seem to get SDXL to generate the font as a material, and not the background. Instead, the text AND the background become the material that I prompt. Is there a way to seperate the text from the background using a specific ControlNet? I've tried all the common ones (canny, depth, sketch, lineart, soft-edge, and scribble) with various weights, but so far, no dice. Thanks much!

[quartz siren]

Finetuning is basically training the model a bit more to make it perform better in some area. You can make it and other people can make it.

[fervent thunder]

having a transparent background (alpha channel) for the control net image might be an issue

[quartz siren]

You can try layerdiffuse which will only provide transparent images like you are looking for. However not sure if it works with controlnet.
https://github.com/lllyasviel/LayerDiffuse

[knotty turtle]

@quartz siren @fervent thunder Thank you kindly for the suggestions. I should have added, I've tried a non-transparent image as well. Still not getting the results I'm hoping for. I don't need it to be transparent, but just thought that might be the way to go.

[fervent thunder]

IP adapter, attention couple and regional text conditioning, all using the same depth map
would be a next step
but its not really worth the effort

[drowsy sable]

Hello

[crimson badger]

Which model does Automatic1111 use? I am led to belive it's XL but idk

[fervent thunder]

any

unless its a really rare or obscure model

[crimson badger]

oh

[knotty turtle]

TY. Why would it not be worth the effort?

[crimson badger]

ooooooh, is it just determined by the checkpoint model you use?

[fervent thunder]

because even in the best case scenario SDXL will not do this task particularly well

you could do this with Flux with zero effort, and the quality will be higher too

[knotty turtle]

I was unaware that Flux had functioning ControlNets

[fervent thunder]

I reckon it could do this without control net

having said that, some of the control nets are not too bad

[knotty turtle]

Gotcha. I definitely need a ControlNet to keep consistency with my font and layout. But the Flux ControlNets I've tried so far don't work. I'll have to keep looking. Thank you for the help

[fervent thunder]

https://huggingface.co/XLabs-AI/flux-controlnet-depth-v3
https://huggingface.co/jasperai/Flux.1-dev-Controlnet-Depththese two apparently, good luck

[knotty turtle]

Great! I'll give those a try. Thanks again

[rough jacinth]

I want to train a lora for oil paintings but I don't want the colors the artist uses to be the only colors the lora uses when make images. Would captioning the entire image, including the colors used be the best way to address this? I'm basically trying to only train on the art style/textures not the colors.

[fervent thunder]

just a case of avoiding overfitting

and then lowering lora strength during inference if needed

[rough jacinth]

so no need to caption the colors in the painting?

[fervent thunder]

a lot of people are doing no captions

not saying that is better but try it with no captions first

[rough jacinth]

I tried but I didn't see much of the texture pop in until using 1.5 strenght, then it started to look weird. So retraining atm with gpt4 captions.

also training a bit longer.

[vagrant raptor]

Anybody have a good image gallery extension for Forge?

I had one for A1111 awhile back but I lost that. I could sort through old images and sort of have a style gallery with it

[knotty turtle]

Search infinite browser

[vagrant raptor]

Thanks, I will

[vagrant raptor]

That should work just fine, I appreciate it

[knotty turtle]

If you're trying to train a LoRA for oil paintings and you want the model to capture the style (like brushstrokes and textures) without locking it into specific color palettes, captioning the entire image (including colors) might not be the best approach. Mentioning the colors could unintentionally cause the model to associate that style with only those colors, limiting its flexibility.

A better method would be to use minimal captions—or even no captions at all for the first phase—so the model focuses on the style rather than the colors. For example, captioning only “oil painting style” and leaving out any specific color details would guide the model to learn the textures and brush techniques without getting stuck on specific hues.

From there, you may need to train multiple models or stages. Start with a model focused on style and texture, and once that’s working well, you can create new datasets by generating images with different color schemes using simple prompts. With this expanded dataset, you can retrain or fine-tune the model to capture a wider variety of colors. This way, you’ll have control over both the texture and flexibility in color in the final outputs.

Also, as mentioned, if overfitting becomes an issue, try lowering the LoRA strength during inference. Sometimes, running extra training epochs or using more diverse images can help too, but testing with simple prompts at lower epochs can help avoid overfitting early on

[rough jacinth]

@knotty turtle Thanks for the tip! I see what I can make with your tips in mind.

[dusty trellis]

I am running Flux via Stability Matrix on my PC. Can anyone point me in the right direction on how to automate this with Python. I have searched, asked ChatGPT and Google and I did find some example code somewhere but I've lost it now. I just need pointers on how to get started. Thanks

[autumn acorn]

hello

[fervent thunder]

hiiiiiiiiiiii

[hard swift]

gm

[frigid temple]

hello

[frigid temple]

How do you use this software? Is there a tutorial for it?

[alpine stirrup]

hi how to use ai?

[pearl oyster]

Automatic1111 is just an user interface. The actual generating is done by the models you want to use. It doesn't even come with any models when you install it, you have to seek out a checkpoint you want to use with it.

[plain raptor]

ppl be joining server befor they even know wut stable diff even is

<v>

[outer rain]

imagine/ tortoises from behind crawl over white sand towards the ocean, at sunset, photographic realism

[dusty trellis]

If I use a seed, there is no point in doing a batch count/size greater than 1, right? Because the seed/prompt combo will always produce the same image

[timber garnet]

hi 🙂

[fervent thunder]

hello everyone

[full vessel]

hello

[wary belfry]

it will count up, starting from the seed you have set

[quartz siren]

If its the same seed like 42, same prompt, same guidance scale it will be the same img. If its different prompt or different seed or different guidance scale, you will get a different img.

[daring gorge]

Hey there, getting back into AI for fun, and downloaded Stability Matrix, can it be used directly, or is it more like a "portal" to keep things tidy and up to date and better use WebUI from there ?

I saw that Inference is useable only with ComfyUI from Stability Matrix so... i'm a bit lost now, too much have evolved

[ornate flame]

I'm looking for a ComfyUI node that iterates through a list of tags in order and outputs the single tag as a string. For example, the node has a textbox with "1girl, 2girls, 3girls" and it sends 1girl to my string combiner node for the 1st generation, then 2girls for the next generation, then 3girls, then back to 1girl. Thanks for any help!

[sullen turtle]

hey all. I want to make a poster for a friend and want to print it in sizes 70cm[W]x100cm[H]. Thing is I don't know where to generate it where it will do a good job but also allow me to suggest images for a poster of this size. Any ideas?

[warm junco]

Stability Matrix is just an installer for the webuis.
I would recommend installing the webuis the default way

[restive inlet]

Hello,

I've been working with Stable Diffusion for a few days to create variations of an existing drawing. The drawing features two characters in a specific situation, and I'd like to place them in different contexts while keeping exactly the same style and linework (for example, the characters could be laughing, dancing, jumping, etc.).

However, I'm struggling to achieve results that maintain the original style and linework perfectly. Is this possible with Stable Diffusion? If so, could you advise me on how to keep these elements consistent while changing the poses and situations?

Thank you very much for your help!

[dense aspen]

hi

Error code 128 stable diffusion

What is the solution to this error? I have been trying to find a solution for two days.

[warm junco]

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

[dense aspen]

ok

[dusty trellis]

Is Flux 1.1 coming to Schnell?

[quartz siren]

Maybe, but it didn't even come to dev yet. It's only for pro right now.

[hollow current]

sorry, thought I was in #🤝｜tech-support

[terse osprey]

I hate microsoft

I hate how even though I had saved images on an AI generator, they decided to erase the website to build it from the ground up and now a ton of stuff that I had on there is gone permanently because of Microsoft being stupid

guess that should teach me to save every image I like instead of just keeping it saved on Copilot because if Copilot revamps itself again, I might lose everything again

[fervent thunder]

yeah I'm sorry that you lost your image but its worth saving stuff

[main zenith]

yall dont allow flux channels?

[fervent thunder]

basically SD3 is the flux channel

these days

[mellow torrent]

Hello everyone! Could someone help me out with more details on how he’s doing this using Stable Diffusion/ComfyUI? I’m especially curious about this part of the video: https://youtu.be/SHmjC7t3fJA?si=vCqwT3uA-SrUmgSV&t=119. How is he making it happen? Where can I find more instructions, like installation, setup, and a quick guide for 'Blocking to Render'? #1292682865028501606

[brave matrix]

yoyo

i have a question

Are there any startups that are doing basically a MJ competitor using Flux?

Like what Grok is doing

and just buildung a really good wrapper and web ui around flux

[fervent thunder]

replicate

Hy

[brave matrix]

yeah but a full featured web ui?

[fervent thunder]

depends what features you mean

would probably just end up looking like mage.space or rundiffusion

[brave matrix]

like inpainting, style transfer, etc

all in one ui

[fervent thunder]

I think rundiffusion is the best version of this

[foggy yacht]

anyone know how to consume most of the vram while training a flux lora using ai-toolkit?

i did select the 24gb template cause thats the only available there. any suggestion?
i attached a pic, its only consuming 24gb vram out of 48gb machine.
https://discordapp.com/channels/1002292111942635562/1026382406279770152/1292765999644545024

[wintry gale]

hi

[still glacier]

Is it a problem tho ? Just because you have 48gb available does not mean it needs all of that.

Do you notice any loading/unloading into vram slowing things down ?

[foggy yacht]

No it's not a problem. And did not notice any difference. Instead it's very fast. But It would be great to utilise all the vram. like with the AI-toolkit on replicate it's fast.

[brave matrix]

can i run flux 1.1 pro img2img non locally?

can you run it in auto1111?

[warm junco]

Flux works only in forge, sdnext and comfyui locally

[fervent thunder]

Hello

[floral umbra]

wouldn't recommend auto1111, it has yet to be fixed annoyances, and forge is just auto made better.

Doesn't break when modelhopping either lol

[floral umbra]

Hayo

[plain raptor]

ello bri'ish sarah

[humble iris]

yo. I ve been working with SDXL for a few weeks and thinking whether I should try 1.5. Do you think?

[quartz siren]

Depends, sdxl has better prompt following and usually better imgs but sd1.5 is way faster so easier to experiment and has better control nets I believe.

[fervent thunder]

SD 1.5 is an outright better model for some uses

[humble iris]

I have 2.8 it\s on 2080ti using 850*850 resolution

trying to set up tensorrt workflow

[quartz siren]

Sd 1.5 will be way faster, and like NeonNinjaAstro said, it can be better too sometimes.

[humble iris]

ok I ll give it a try thanks!

I just need 3 more 2080ti and I m good

[hollow anvil]

Hey. Which room is best for stable-diffusion-webui ?

[worldly bough]

Given the choice between an RX 6900 XT and a RTX 4070, which would be better for generating images in SD? I heard AMD cards are just way slower.

[thorny rose]

Hey folks, I've been using Automatic1111 w/ ReaActor to try and make a comic strip featuring my son. I haven't touched these things in many, many months. Is there something better I should be using these days? Thank you!

[echo kite]

is it just me or is Dreamshaper a REALLY good generalist model

also, I am trying to condense down my several super specific models into a few good generalist models

any recommendations?

also what model is closest to the quality of modern NAI

[quartz siren]

4070 is probably a better choice.

[solid spindle]

Forge UI - any one know how to disable the "checkpoint merge" Tab? I do not use it at all. Can't find it in Settings. its "Hidden UI" i expected its calles "HIDE from UI" anyways....

[humble iris]

where can I find info as a newbie on comfyui on what commonly used custom node do what, what are the best starter workflows etc?

[floral umbra]

Is there a channel for lora training related? As i'm struggling quite a bit with training SDXL lora's

[icy vigil]

just got into text to video, is stability the best resource right now for generating quality videos

[humble iris]

I heard on reddit this is not so

[quartz siren]

No, CogVideoX is the best open source model for sure now. It's not Kling or Gen3 level but still pretty decent.
Svd was pretty good for it's time but pretty outdated now.

[humble iris]

you can get 3080ti for the same price, it s 30% faster in SD, and I d buy a used one

[dusky halo]

guys i used to work with automatic11111 a long time ago. now trying to start working with SD again. which UI u suggest for win11? kinda dont like to bump into lots of errors when installing and working 😄

[humble iris]

I m starting to learn comfyUI and already feel like it s the best decision

[floral umbra]

either works. Just auto1111 being the shittiest :P

If you prefer it, there's forge webui, basically fork of auto, but fixes most of the negligence auto still suffers

[floral umbra]

I still bounce between comfy and forge as some stuff is better pre-setup for the other :P

[woeful flume]

Is it possible to do after generatione edits, like upload a image, highlight something like a part, a arm, or head, and regenerate that specific part?

[cedar salmon]

ya thats called inpainting

[woeful flume]

How does that work?

[cold estuary]

have a good day

[fervent thunder]

hi

[pearl oyster]

Amd cards are slower because SD is built with CUDA in mind and amd cards have to translate the cuda instructions into something they can work with. Depending on the OS, there are ways of speeding it up significantly(zluda in windows, ROCm in Linux), but they're still just not cuda.

[pearl oyster]

That said, a card like the 6900xt, while slower than nvidia, will still be decently fast. A surprising factor affecting the speed is how much vram the card has, an area where amd has novideo beat. If there's not enough vram to fit the model, it will rely on the much slower system ram or, even worse, the page file. The ideal would be an nvidia card with at least 8gb of vram(if not much more). Second best would be amd with 8-12gb of vram imnsho.

[fervent thunder]

I use a 7900 xtx

[pearl oyster]

And I'm using 6700xt. Mind, I didn't get it specifically for ai, but I figured I'd try it. Has been a surprisingly good experience.

Then again, 12gb of vram is quite helpful

I can run 3 batches of 4 images per batch in 12 minutes, making it about 1 minute per image. Using an xl model at 896x1152/1152x896

An equivalent nvidia gpu can probably do it faster, but this isn't exactly slow.

[worldly bough]

Yeah I heard good things about Zluda but I wasn't sure. I also heard it got taken down lol. But the data hoarder in my downloaded and kept a copy. I am on windows btw.

I've went back and forth on AMD and Nvidia cards. Both are fine and have ups and downs imo. Right now I'm on an RTX 3070 with 8gb VRAM. And 8GB feels so bad. Some games need more. That's why I'm going to upgrade. I just really wish AMD cards could use Ansel for screenshots. In the 5 games that feature is implemented, it is sooo nice.

One thing thats making me lean a little more towards the 4070 is the power efficiency. My computer room is somewhat small and gets hot at times. Less power means less heat, and the 6900 XT scares me about that lol. I think the 4070 uses somewhere like 220w? 6900 XT and RTX 3080 TI both use like 300-350w respectively. From my quick searches anyway. And 4070 is newer so it'll have better resale value if I end up wanting to get an RTX 5070 or 8800 XT or whatever they'll be called lol.

[pearl oyster]

Entirely up to you. There's benefits to both, and downsides to both.

[dusty trellis]

I have questions about the Stability Matrix app and WebUI Forge. Is this a good place to ask?

[hexed minnow]

Is there a way to cartoonize an existing image without completely describing the image in the prompt? I kinda have the problem, that either the image is not cartoonized enough or it ends up being a completely, weird different image.

[cursive birch]

hello chat
would you pick a 6700XT or 3070 Ti for higher definition image?

[pearl oyster]

3070 ti benefits from already being nvidia, but you'll be offloading xl models more often. I can tell you that the 6700xt is decent, able to churn out 12 images at 896x1152/1152x896 with an xl model in 12 minutes. I use it myself. But setting it up requires more work, as the ai tech is made for cuda, an nvidia tech

[prisma solstice]

Hello

[cursive birch]

Could you explain what "Offloading XL models" means 😅
I'm already using 6600 with ZLUDA and the definition + frequent memory errors is kinda annoying

[still glacier]

"offloading model" <=> splitting one model into multiple subparts and then proceeding to load and work with each of those part instead of working with the whole model at once. Needed if you don't have enough vram to load the whole thing but slower as it requires splitting the model, loading some part of it, unloading some part of it then loading another part, then loading another part, then etc

Kinda like baking a cake while still having all your aliments in your car from the store and having to clean your hands in between every steps.... Sure I ll roll with that analogy.

[cursive birch]

Well I guess that explains it

Thanks

[still glacier]

It s much faster to have eveything at hand and ready, without having to have to repeat tedious prep steps all along.

And now I'm hungry, damn it 😄 .

[cursive birch]

The image definition would benefit more from the 6700XT/6750XT's 12GB no?

Or is 3070 Ti enough

8GB sounds a bit on the lower side of things

Speed isn't really an issue here

[still glacier]

With 8gb you should be able to generate (not train tho) anything (for now at least) but it's probably gonna require offloading at some point when generating with heavy models.

It all depends how much you value your time and how much you re willing to bet on AMD catching up Nvidia on the cuda scene.

Personally I value my time more than some others and have enough money to bite the bullet and pay the green tax when it comes to building a pc.
You can check some benchmarks to get some approximations of what to expect : https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html

[pearl oyster]

I mean, under windows there's zluda for amd, which allows me to get 1 min/picture. And on the Linux side there's ROCm. So it's not like there aren't workarounds for the cuda issue.

[still glacier]

Yes there are solutions, not denying that. But they come at the cost of a much harder setup and compatibility issues with some extensions.

And I have no magic crystal orb to foresee if Zluda will be able to keep up with Nvidia in the future. (Sadly) Nvidia did invest a ton more money into their software and have started doing so years ahead of AMD. So team red is playing catch-up right now.

[pearl oyster]

Ugh, I hate how it can't just say ryzen 9 5900x

No, it has to say amd64 family 25 model 33 instead

[still glacier]

To be fair, the cpu doesn't play that much of an important role in SD (unless you re forcing it to execute in "cpu mode" obviously). If there s not offloading of any sore, no "cpu mode", etc... The only significant impact of the CPU will be "how fast can I load the model into the GPU" (and even there, there are some asterisks).

[pearl oyster]

Yeah, but when comparing test systems, even if one part isn't actually doing anything, you'd want it to be human readable. The average consumer doesn't know what this family/model system is. I had to look it up myself

[cursive birch]

Maybe I'll get 6750XT

3070 Ti's VRAM is rather limiting

And 3080 12GB is a tad too out of budget

Oh and used 6700XT is only $200 here holy mama

[pearl oyster]

Yeah, that's the baffling thing about nvidia: they insist on putting just a bit too little vram on their cards.

[cursive birch]

Didn't know Apple made discreet GPU but here we are

[pearl oyster]

Jensen Huang is Tim Cook?!?

Lol

[agile orbit]

pls, i need a 2d animator

i've got the cash

[young willow]

Hi

[spiral oriole]

Hello !

[native root]

Hi everyone - Is there a chat dedicated to people who need help setting up and running Automatic 1111?

[still glacier]

#🤝｜tech-support

[hallow plaza]

What does pony do and how do I use it with flux?

[still glacier]

pony / ponyXL is a model (based on SDXL). Flux is also a model.

"Using a model with a model" makes no sense.

You can use either of those in ComfyUI, automatic1111 stable-diffusion-webui, sdnext, SwarmUI, etc.

[floral umbra]

Is epochs when training just to split load? Or will more steps and less epochs make model better there?

[zenith elm]

Anyone know where to hire people who are really good at inpainting with SDXL models/loras?

[granite magnet]

hello

[frail salmon]

need help, i can't run stable diffusion webui on kabble

well, i can run but when i access in locally, the local rejects the conection

[hard spade]

Hello :D,

[lean mesa]

Hey I need help. I've been trying for like a good month on my own but I just can't manage it. I want to place a specific style on my image. I have this drawing colored and drawn. Then I have a checkpoint and also a lora. I've been looking for ways to transfer just the artsyle to the drawing without messing up the original at all while also carrying over the original colors. I've tried ipadapter, controlnet, t2l adapter and I just can't figure it out. Can anyone please help? I dont want it generating things that aren't there or change the image at all

[oblique elk]

if the image does not contain content that is against the server rules and you could post the drawing/image and the style you want to apply in the channel general with images. So people could try to help

[frail salmon]

i'm new in about using stable but u can try some configurations

but actually can't run the webui

[lean mesa]

Thank you I will try that

[fathom wadi]

Herro

[feral pike]

can someone please explain to me what controlnet models are?

[oblique elk]

https://github.com/lllyasviel/ControlNet
Scroll a bit down, ignore the math and look at the examples

[plain raptor]

mmmmmmm

math

[frail salmon]

i want to use automatic1111 on kaggle but there's are not notebook, someone have one?

[dense shoal]

easy inpaint and outpaint from 6:00 https://youtu.be/cx7L-evqLPo?si=Lzk6QWYGmY2pnzRT

invoks canvas is very handy

[brave matrix]

anyone have tips on how to do flux inpaint non local?

like a website

[loud wedge]

Hi, @zenith elm
I've hands on experience in inpainting.
I often use Flux + Controlnet for the HD image text guided inpainting.
Could you kindly share me with more details?
Thanks.

[ashen sleet]

sup

[celest tide]

hii does anyone know how people make very clean and precise any anime character fan arts using stable diffusion? if you know please dm me. I'm new to ai and have no clue how people make these

[feral pike]

how much vram do i need for fluxx dev and schnell?

[fervent thunder]

I don't know the low end but really low

with the smallest GGUF quants and comfy

[feral pike]

i am using comfyui yeah

where can i find GGUFs?

[fervent thunder]

not sure

usually civit ai, or huggingface

sometimes github

for models

[fervent thunder]

without tiling?
you can but it takes a bit of setup

two passes with res-adapter, deep shrink and PAG, with an upscale in between the passes

[feral pike]

how much time does fluxx dev and schnell takes for u all per image? 20 interations

for me its taking around 25 mins both..

[finite kindle]

hi

[feral pike]

👋

[fervent thunder]

25 mins means VRAM got full

you need a smaller quant

when the time goes really long like that, its mostly vram issue

but flux can fit on very low vram with good quant

[oblique elk]

And schnell should only need 4-5 iterations.

[feral pike]

ah!

[quartz siren]

Yeah dev should not take 25 mins, maybe 1-2 minutes on a not too great gpu but definitely not 25 min.

[remote wraith]

Hi peps

I want to use depth generation for some architectural images.

https://creatixai.com/wp-content/uploads/2023/09/ControlNet-Depth-preprocessor-depth-map-example-1024x576.webp

Im a customer of openart.ai but there semes to be no function for generating images out of depth. Sketch to image works fine but I want to know what else is possible.

Is there a way to generate images out of depth-images online? Or can I do this only on my PC?

Thanks.

[cloud vigil]

Hi! I'm quite new in the IA, I'm working with stable difussion and I have certain questions I am not sure if this is the correct channel, if it isn't please let me know 🙂

I have realizie that usually, the images that stable difussion (Right now I'm using XL) doesn't match proppertly the promt that I write, is this common? are there certain guidelines to follow to make the promts?

I have try using (( )) for add weight to the promt, but it doesn't seem to work really good

[low moon]

Welcome to AI.

[cloud vigil]

Thanks 😄

[low moon]

Don't think of this as a reliable pipeline or process.

It;s more like you're in a casino and you pull the slot machine and sometimes you get something good.

And even fi you do get something good it is not 100% reproducible.

What can I say... early days...

[cloud vigil]

mnmn I understand

But it is hard to get something like the position of the arms or the head with precission

any advice for that?

[feral pike]

i have 16GB vram , rx 7800xt gpu

[low moon]

u cna use openpose control net

[feral pike]

can someone please help , what am i doing wrong , why is it taking so long with fluxx?

i am using comfyui

[low moon]

i take about 1m on 12 gb vram so

its not too bad

[feral pike]

1m per interation?

[low moon]

unless u ad controlnets and ip adapters and stuff

yeah

[feral pike]

oh

what does ip adapters do?

[low moon]

they heklp control output

but they slow it doen a lot

[feral pike]

isnt that what controlnet does?

[low moon]

well its not the same but yeah

they both slow down the generation

[feral pike]

confusing

how many iterations do u usually run with fluxx?

[low moon]

https://www.youtube.com/watch?v=KXxk_bGDCzc&ab_channel=OlivioSarikas

https://www.youtube.com/watch?v=HVYXM9bPFTs&ab_channel=OlivioSarikas

[low moon]

20 for dev and 4 for schnell

[feral pike]

why so different?

isnt dev just a powerful version of schnell?

[low moon]

it does give better results yes but its slower

schnell is like sdxl turbo

[feral pike]

ohhh

[low moon]

faster but a little less quality

[feral pike]

what guidance do u use on fluxx schnell?

[low moon]

CFG? 1

[feral pike]

why so low?

[fervent thunder]

you can go higher if you want

but it takes twice as long

and you have to use extra nodes to fight the CFG burn

[viscid stirrup]

how to use the bot?

to generate images

[quartz siren]

Yeah the same thing Neon Ninja Astro said, Flux dev does not work with normal cfg but distilled cfg and doesn't support negative prompts

I actually recommend trying out Flux.1 de-distilled if you can wait, that supports normal CFG with negative prompts, slower but produces better results imo.

[feral pike]

whats distilled cfg?

[fervent thunder]

its called guidance

its a new fake variable

that the teacher model taught the student model

it doesn't relate to anything in the real world its just a virtual label

they taught the model to imitate it

[feral pike]

oh?

[quartz siren]

Yeah Flux.1 dev and Flux.1 schnell are distillations of their closed source Flux.1 pro. Flux.1 pro is the original model, and does support normal cfg and negative prompts. Dev directly does not, they made it have distilled cfg which is an imitation.

[feral pike]

👀 wow

do u use comfyui?

[frail salmon]

need help i need a working notebook from kaggle about auto1111

[quartz siren]

Not very often, I usually use diffusers.

[feral pike]

never heard about that

[quartz siren]

Thats the main python library to load models, I usually need to use python and diffusers is simpler to run then use comfyui's api.

[feral pike]

i keep getting shape mismatch error 😢

[quartz siren]

Whats the code?

[feral pike]

#🤝｜tech-support message

[pliant bane]

Hello, I haven't used deforum for over a year, since Google colab banned it, have you found a way to use deforum for free? I need to know urgently

[wicked dust]

Hey fellas, anyone here using fooocus with Jupyter Notebook? Stuck at last step :/

Hey there, can you tell me more details?

[wicked dust]

hey you can use it when you rent GPU on sites like https://www.runpod.io/

[pliant bane]

Well, it's not free 😅 I got a time when you could run SB on colab for free for 4 hours

SD*

[wicked dust]

it cost like 0.3$/hour 😄

[pliant bane]

Extremely expensive, a 15-second video can take up to 2 hours to make

I made about 8 videos a day, I spent the whole day on the computer

[wicked dust]

well I didn't try it so I am not gonna argue but highly doubt that when image is for me currently about 15seconds

[pliant bane]

Yes, but imagine 5000 images to make a video

Look

[wicked dust]

btw do you have by any chance experience with fooocus on these sites?

[pliant bane]

I'll send you a video privately

[wicked dust]

ah I guess you are right

[wicked dust]

alright

[sly eagle]

Hi! what tools are new for animation in stable diffusion?

[quartz siren]

Rn, CogVideoX is currently the best open source video gen model.

[sly eagle]

Is difficult to learn? or difficult to install?

[quartz siren]

Well you can install it with comfy ui, or diffusers. It’s pretty popular now as an open source video gen tool.

[sly eagle]

Thank you very much! I'll check!

[trim field]

Hello to everyone... looking forward to an amazing experiene with youguys.... ok then, best wishes!

[unborn hedge]

is there any guide to install flux locally? got a 16gb nvidia card, im looking for a flux model that can use lora's, im not sure what the difference is between pro and schnell?

[unborn hedge]

nvm i think i got it figured out, im using the dev flux model i think

[low prawn]

Hello Managers, my body is ready for Flux Auto canny

[fervent thunder]

not sure how well flux will follow current canny control nets
but good luck 🙂

[dire prairie]

hello

[ashen sleet]

gm

[dark lichen]

I use to create free images before in this server now it seems like it's gone?

[fervent thunder]

yeah

[pastel lynx]

hey, does anyone knows how to setup a grid generation based on flux checkpoint ?

[fervent thunder]

yeah is comfy ui information okay?

[pastel lynx]

yea i'm looking to get a node to use inside Comfy

[fervent thunder]

the way I do it is use KJNodes
he has nodes for adding labels to images, and for joining images

then I copy and paste the Ksampler setup 9 times

[pastel lynx]

thanks i'll look into it

[fervent thunder]

also check out his nodes "widget to string" and "something to string"
to automatically generate the text for the "add label" node

it updates labels based on values in nodes

[scarlet stratus]

Hey everyone, anyone is really experienced in stable and could help me ?

I am trying to place a product on a background. But I don't want to do Image composite by mask, I actually want the AI to recreate the produc. Any idea how to do it to be as acurate as possible.

That would mean >

I load a product image
remove bg
write a prompt
It generates the product into the background

[sudden ruin]

https://civitai.com/models/419539/botos-background-swapper
I made this some time ago, with some adjustments it might be able to serve your use case

I edit the original picture on top, but you can remove that part and have the AI product as you said it

[fervent thunder]

Hi , i like your workflow, would you be open to work with our team?

Please do pm if so. Looking forward to it

[scarlet stratus]

Hey thanks for sharing, how does your workflow handles product recreation ? Is the fidelity good ?

Cant test rn

[loud wedge]

I think you should rely on other projects rather than stable diffusion.

You could change background first with pretrained AI models but not with SD.
And then could harmonize that with another model.
In this case you could customize your background as any image you can.

And no need to worry about fidelity and variety, maybe.

[raven agate]

Something like this? https://civitai.com/models/546180/geeky-remb

I made a custom Remb node for comfyUI that removes the background of one image/images and layers it on another. You can daisy chain them, adjust the x and y position, animate it a bit with simple 2D animations and etc.

[sonic gazelle]

Please I wanna ask if anyone uses MacBook for most of this ai generation and this heavy face swapping software, are they a good options?

Or I should use this heavy omen 16

[dawn mulch]

Whats the difference between 'merging' 2 models together for a generation, and doing 1 pass with 1 model and then another pass with the other model?

[scarlet stratus]

Which project for example ? What's the problem with SD ?

[scarlet stratus]

+1

[fervent thunder]

when you do 2 passes, 1 with 1 model, and the other with the other model

you are choosing a sigma to stop the first model at

and to start the second model at

they are never acting at the same time

if you want the diffusion step at a particular sigma to be a mixture of the two models then

2 regular passes can't achieve that

[sudden ruin]

I use CosXL for it, id say the recreation output is only mediocre, thats why I edit the original picture on top with the workflow

[graceful bramble]

Speaking of merges. Generally speaking will a merger of two models be better at following loras if one has trouble working with them and the other doesnt?

[fervent thunder]

it would probably be in between the two

in that ability

[graceful bramble]

Got it. Thank you.

[fervent thunder]

for the most part merges seem to average almost all aspects

[graceful bramble]

Perfect.

[scarlet stratus]

Ok, and why cosxl ? what's the difference between sdxl ?

[sudden ruin]

Its a great model to edit existing pictures, its based of SDXL

[noble basin]

Hello there, i've already train my own model .. but i'm not satsified of the result because my model generate beautiful landscape but the character are sometimes ugly .. someone can help me to find some tutorial to made a style with more control ???

[pastel lynx]

Hey Neon, the issues with KJnodes is that i can call a base ckpt name but there are no options to call the diffusion model like flux_dev for example to iterate in the grid, that's my main issue..

Perhaps, i found this workflow that allow to call diffusion model/unets etc... with qqnodes https://civitai.com/models/635692/flux-xy-grid-dark-mode but i got a "Grid Annotation" issue at the end of the pipeline that does allow me to make the grid while everything else is working

[fervent thunder]

load diffusion model node, then unet_name widget to string

[pastel lynx]

Oh yea, i see it might works, Thank you !

[scarlet stratus]

Hey guys, I'am looking for an experienced comfyui individual (paid job) ! 🫡

[azure nebula]

https://wccftech.com/amd-tunes-rocm-for-consumers-turns-radeon-systems-into-localized-ai-machines/

Good news for AMD GPU?

[young bronze]

RuntimeError: Your device does not support the current version of Torch/CUDA!
😿

[warm junco]

What's your GPU?

[young bronze]

I'm just venting, I am perfectly aware of my GPU and its lack of CUDA

[warm junco]

Oh okay. If your on AMD, I have guides for that

[young bronze]

Oh, that would be nice
I had Forge working before until something broke it that I can't figure out so I'm just trying to blank slate
which is proving rather challenging what with things like python stubbornly refusing to exit my system so I can put it back in

[warm junco]

Here are all my Guides:
Go for the ZLUDA versions if your GPU has 8gb or more vram.
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[young bronze]

oh cool, thanks
wish me luck

[warm junco]

NP, for any questions feel free to ask me in #🤝｜tech-support

[jovial igloo]

How do you guy remember all the trigger words for your loras? Also is there a way to automatically add the trigger words for a lora?

[minor fjord]

I am new to Stable Diffusion, is it free to use for converting Images to Video on macOS? (I have Intel, not Apple Silicon)

[fallow veldt]

[clear oyster]

if stable diffusion large and medium are out, then what about stable diffusion small?

[fervent thunder]

uh oh

@sudden ruin

[dawn mulch]

'Fight me". JK.

I love that video and quote.

[gritty kiln]

Is it even worth installing S.D with a 3060ti?

[warm junco]

Sure its a good card for SD

[gritty kiln]

Only 8gb of vram though :/ are there any samples of the best quality work that a 3060ti can produce?

[fervent thunder]

yeah, any Flux image

with the quants and tiled upscale you can do it

[gritty kiln]

Is there a preferred gui frontend these days?

[fervent thunder]

yeah comfy

and then diffusers or pytorch for CLI

[gritty kiln]

ok thank you

[warm junco]

Automatic1111 for beginners, forge or comfyui for flux usage

[warm junco]

You can generate large images by upscaling

So no problems with that

[gritty kiln]

OK, I think I will get it installed

[warm junco]

You can follow my install guides.
For Automatic1111 it shows the best config dependending on the vram.
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[gritty kiln]

awesome, thank you

[warm junco]

Np

[cloud sundial]

Hey, i used to run 1111 and switched to comfy, was reading comments on the reddit and saw a lot of people suggesting in particular (ppl with 8gb vram and less) said they got better gen times with forge, is this possible or is it just confirmation bias?

[fervent thunder]

not sure

[scarlet stratus]

Anyone can help me with product photo ?

I am not looking for an already existing workflow as I already tried every single one of them and those are not the results I am looking for, please dm me.

I'd love to talk with someone experienced to get their point of view. 😇

[silk condor]

Is it possible for stable diffusion to make a automatic folder for each prompt? I know it can be annoying when you make 1 by 1 but when i do 10 in batch or more it would be nice.

Does it work with like %folder-name%-%number% or something? 😅

[desert dagger]

Photo, or generated image?

[ashen sleet]

wsg

[warm junco]

Yes its slightly faster and uses less resources

[lofty sphinx]

Wassup

[humble iris]

I m considering modding my 2080ti to 22GB. Or upgrading to 3080ti and modding that.

1. So I ll be able to make 1400-1400 images? But is SDXL trained to do that?
2. I`ve noticed that reducing resolution from 1000-1000 to 800-800 improved prompt following a lot...
   Gemini Pro says that more vram allows potentially better adherence to instructions, is it true? is it significant?
   @warm junco

[scarlet stratus]

Starting from a real picture, to integrate it in a new background. But i don't want to do some basic image compositing, I want the AI to recreate the product so it has a better blend of light and colors

[leaden cargo]

hi

[violet veldt]

Thank you for using comcom analytics.
"comcom analytics" supports all community managers (moderators and server owners) by stats, visualization, and analytics.

If you have any questions, feel free to ask us!
Your dashboard
Help
Support server

Other languages
en: help
ja: help Japanese

[sonic gazelle]

Yo anyone know a live face swapper ?

??

[oblique jay]

The 1400x1400 should be fine for SDXL, but there are so many factors that contribute to how an image is generated, it's hard to say how the image will turn out, no matter the VRAM amount. More VRAM has better adherence since it permits for higher-resolution models.

(By this higher resolution I mean models that are not optimized for lower-end cards)

Reducing pixel resolution generation seems to follow the prompt, though I am not really sure how this infrastructure works.

[sonic gazelle]

Bro please check your dm

[warm junco]

More vram doesn't mean better prompt adherence. Like gwolf, said it allows to use larger models.
It depends on the models.
Sdxl is trained on 1024x1024 but with Upscaling you can go up to 2k, 4k etc

[humble iris]

okay thanks guys

actually when I upscale I lose detail, it becomes worse

tried all the best ones

[warm junco]

If you used hires fix, try lower the denois

Also if you upscale in img2img you can use resize (latent)

[humble iris]

only tried normal upscalers

I ll try hires fix

[warm junco]

Dont use the Extras tab for upscaling it can't generate details

[humble iris]

I m in comfyUI

[warm junco]

Ahh okay

[desert dagger]

Thats.. not going to happen unless you train a lora on your product

[oak minnow]

Hi. When I share a google drive link for someone and I gave him editor role. Could he access the files other then what I shared with him ???

[untold arrow]

some legend wanna run me through some questions i have abt textual inversion?

[fervent thunder]

ok

[untold arrow]

I want to train an embedding on my face.
Im technically fairly competent but this is way above my skills.

1:
when i train an embedding do i have to use the model with which im afterwards gonna generate the images?

2:
Can i use any model? I wanna use dreamshaperXL_v21Turbo for the generation.

3:
how? As far as i can understand a1111 cant train with the sdxl models.
I installed OneTrainer and immediately have no idea what im doing. I cant even find where to specify the location where my training images are and the documentation is virtually non-existing.

It feels a bit like a phd in computer science is basically a requirement for playing with anything deeper than the most basic image generation

[fervent thunder]

embeddings train clip rather than the diffusion model

you do have to use a specific text encoder while training, and its the one you want to use in inference

[untold arrow]

i have no idea what that means =)

[quartz siren]

you don't need to know much but basically stable diffusion has 3 parts

1. text encoder(clip), converts text into embeddings which are basically "meaning"
2. diffusion model(unet), uses the embeddings above, to generate the image as latents
3. vae, to convert the latents above into an actual understandable image

The easiest option is to not train anything, but just use something like IP adapter which a111 supports out of the box I believe. This just needs a single good quality image of your face. Try it first, if it does not give you good quality, you might want lora training.

[untold arrow]

hmmmmmmm okay... Is training a lora less complicated? Do they produce better quality images than embeddings?

[desert stump]

Hey are there any sub servers that are using stable diffusion? How can I find them?

[desert dagger]

explain please?

[low moon]

Slow days.

People are in pain and preoccupied.

"What's next" everyone asks because we all know the way things are can't go on.

[quartz siren]

Lora training trains the diffusion model, embedding only trains the text encoder. Lora training should be better quality.

[untold arrow]

I see

thanks =)

[bright perch]

Quick question for everybody, I remember on automatic111 you could install an extension that would show an image wall of your checkpoints and loras, and then you could auto assign preview images to each checkpoint and lora. I'm using Forge now, and it has the "wall" for checkpoints, but I can't figure out how to give them preview images. Does anyone know how? Thanks

[graceful bramble]

Click the little hammer and wrench icon in the top right corner. Then in the bottom center hit replace preview. But it will only use the last generated image, there is no option to select something else.

[ashen sleet]

gm

[iron ruin]

gm

[scarlet stratus]

I think it is possible with multiple ksampler, and playing with the denois and ending with a restore detail node

Hey, anyone knows how to perform good accurate color match ?

I am placing a subject on a new background but the colors look off.

I have tried using colormatch node but the result is off too

[hard osprey]

Been saving up for a 3090 PC. Any places to get one cheaper? I know there's craigslist but no good deals around locally yet

[scarlet stratus]

guys, anyone knows why inpaint with brushnet and sdxl model changes the background color and makes the mask seem so visible ?

[lavish latch]

hey guys, does anyone know of a comfyui node that gives output in a few separate bursts? like for example, it generates 3 images, but instead of returning it as one bundle of 3 images, it outputs it one at a time?

[young bronze]

@warm junco I'm gonna have to redo everything with 5.7 I think, with 6.1 it ends up crashing while compiling after throwing a bunch of python error

ONNX failed to initialize: module 'optimum.onnxruntime.modeling_diffusion' has no attribute '_ORTDiffusionModelPart' Exception Code: 0xC0000005
then a bunch of rocblas.dll stuff

[craggy heath]

Hi guys. Does anyone use flux with 6Go VRAM ? Searching on internet, I read everywhere some guys who do it and can generate under 5 min... I don't understand. I have 6Go VRAM (a GTX 1060 card) and I can generate between 11 and 16min, approximately... 😞

I use flux1-schnell-bnb-nf4.safetensors and the basic workflow found here : https://turboflip.de/flux-1-nf4-for-lowend-gpu-for-comfyui/

[warm junco]

What was your GPU again?

Normaly its better to use 6.1

[young bronze]

6600 xt

[copper crystal]

Change is always a constant. The way things are can never go on indefinately.

Change is the ONLY constant some have said

[brave vigil]

@warm junco Hey there! Sorry for the ping. I just wanted to let one of the admins know - y'all need a "Programmer" field in the onboarding section where you select your hobbies

I, personally, tackle SD from a software engineering perspective

[hard osprey]

Any recommendations for image to 3d? I want to mainly use it for anime hair

[fervent thunder]

I don't think good image to 3d exists yet

[hard osprey]

Probably not, but I want to get a base and then modify it in a 3d program

[ocean stratus]

Hello!

[rare aspen]

hi

[young bronze]

@warm junco I think I got it running now? It's still iffy and the results is questionable, but at least it's not a grey box, I do still have 6.1 and 5.7 HIP SDK both installed as before so I really don't know what happened before, nothing actually was changed to be honest

It's literally just smudges but still better than grey

Wait, never mind. It's working now somehow

[warm junco]

Okay, good. Let me know in #🤝｜tech-support if you get any error again

[young bronze]

who the f is Jon Snow

[high ruin]

What are the 3 or 4 best Esrgan models for upscaling people??

[safe cradle]

is there ever gonna be a stable diffusion for generating 3d models?

[faint sleet]

I am an experienced AI developer with 2 years of expertise in creating innovative solutions, as well as a fresher web app developer skilled in React and Next.js. Additionally, I specialize in building crypto trading bots and offer my services at affordable rates. Let's collaborate to bring your ideas to life with cutting-edge technology

[grim aspen]

has anyone else issues with forge getting Killed when using --medvram or medvram-sdxl?

[bright cloak]

hey guys, ultimate SD upscale and tiled diffusion, which one is more recommended?

[still glacier]

already are a bunch https://stability.ai/stable-3d

[still glacier]

two different things that can coexist

[bright cloak]

video says that they are two different ways to upscale picture, but which one is better?

[still glacier]

dont know what video you re talking about.
Tiled diffusion is about "splitting the VAE part taking place on the whole "picture" into multiple "VAE operation" producing multiple smaller images and then stitching things back together to get one big picture". Making it much easier on the vram consumption.

and Ultimate sd upscale is about upscaling

[steep copper]

are there any channels for forge webui

[desert dagger]

if it's a technical question, just ask it in #🤝｜tech-support

[stark rapids]

do yall know stability matrix??

[ebon flower]

Hey guys,
I am going to create like an influencer...and I was wondering what kind of settings/extenstions/LoRa's are good to create it, and what do you guys use. I downloaded the pony model already from civitai but I have no idea how I could generate consistent faces. Last year, I managed to generate consistent images, using a faceswap and in the prompt I would input like a celebrity that looks alike and would be generated mostly the same. Is there any better options to go about it? Do you guys use the pony model? if you do, could you share some configurations and prompts. Sorry for this mess that I wrote, and thanks for any comments😀

[brave vigil]

Meshy.ai is the best text to 3D I've found, the models it creates are actually pretty good! Image to 3D is a more difficult problem. Monocular photometry, the most promising method of inferring 3D structure from an image, is still very primitive. Even if it wasn't, there's only so much information that can be gleaned from an image. And if that image is illustration, then there's essentially no point

I've put a lot of thought into it, and the problem boils down to three sub problems: inferring the structure of what is visible from the image, inferring the structure of what isn't visible, and actually translating this structure into a 3D mesh. If you've got the latent then the last part is relatively simple. Multi-view inferrence is the only mostly-acceptable way of solving the second problem, and breakthroughs are required to solve the first

Especially for stylized tasks, the only actual way to do this is diffusion (or similar) in a shared embedding space which can simultaneously decode into both 2D and 3D. If you can do this, then an encoder can reverse engineer a latent from an image, with diffusion (and friends) doing the work to synthesize whatever isn't visible. But the processing power required, not to mention the dataset

shudder

[devout hull]

Hello

[sleek otter]

H a p p y C a n a d i a n T h a n k s g i v i n g!

[devout hull]

wait... no way to use ssl with stable diffusion server?

my reverse proxy setup is ony working (on my https site) usind webserver and stable diffusion server on the same machine... testing from a client webpage i get cors error 😦

[devout hull]

Someone that has familiarity with the SD api and can test the wordpress plugin I made to see if some features that can be implemented are missing?

[gray sun]

even less frequented dc, but coolest of them all, all of dc been pretty quiet tonight, is there some special occasion going on in the whole wide world that I am missing out on lmao

[hasty olive]

I dont know where to ask this question but is there a way I can run SD locally and have it be used in my discord server?

[peak tendon]

best upscaler for CGi graphics? Think of old mario 64 promotional silicon graphics style

[mystic siren]

hello, does anyone know if flux nf4 supports lora??

[young bronze]

canadian thanksgiving?

[devout hull]

Search for ajabot

What can I add to my SD wordpress plugin? https://fidefix.ddns.net/sd.mp4

[stuck sun]

is there a consensus on the best photoreal XL model now? i liked Helloworld, but not been updated in quite a while

[indigo saffron]

hello guys, now I use i2i to do style transfer(with SD1.5), there are some text in my picture(my pictures contain some signs with text). After style transfer, the text became blurred and illegible.
Is there any method that can help me keep the text on road signs clear when performing style transfer? Are there any LoRA weights or workflows available for use?

Thx for your advice!

[indigo saffron]

Maybe I change base model can solve this problem? Do anyone know some base model which do better in text gen?

[brittle brook]

Hi Thomas from LA here, working on https://WandAI.app, a one-stop AI creativity workspace and community especially designed for non-technical creatives .

We inviting 100 creatives to join the internal testing, where you can share the challenges you’ve encountered when using AI tools, brainstorm new ideas, and more, feel free to jump in!

[restive dagger]

hello

[frank halo]

hey guys

Good morning

One thing, to move Automatic1111 and my comfyUI to an external SSD, what do you recommend?

Would it be just a matter of copying and pasting the folder? Or would I have to install the interfaces on the SSD and transfer the checkpoints, loras, preprocessors, etc.?

[hard swift]

yo

[frank halo]

hey

[warm stump]

i am looking for someone as same age as me to be a fellow developer or any other. wanna be a friend?

[devout hull]

This community is dead... I am developing multiple frontends (wordpress web plugins) for textgen and txt2img servers asking people what they like and maybe test functionalities to give feedback, nobody cares, nobody responding... I am asking myself why I am here...

This in multiple ai discord servers.$

[low moon]

They're busy generating cats and ponies.

[frank halo]

I would appreciate it if you could send me more active servers related to comfyUI and A1111, on this server they are totally indifferent; it is a waste of time to ask here.

[main junco]

What is the best to make stylyzed photo of yourself?
Like pin up or something?

[bronze nacelle]

hi hi

[quartz siren]

You can try flux or sd3, both are great at text gen but sd3 is bad at humans. Flux will take much more time but give you better quality.

[quartz siren]

It should technically but it will not be as good as applying Lora at bf16 or fp16(normally)

[quartz siren]

You can use controlnets and ip adapter. You can also use special techniques like https://github.com/songrise/Artist

[gaunt helm]

how can I make a poster on this ?

I need to make a poster on mineral resources for my college can anyone help?

[loud wedge]

Just post here.

[desert dagger]

sd3 is only bad at humans if they are laying down or standing on their head

[quartz siren]

No, it's basically any pose apart from standing. Even with standing, fingers are usually deformed or messed up.

[brave swallow]

Is there anyone that understands llm nature? So I'm having issue where the model will not respect the token limit or the stop commands. Sometimes it will start its response by finishing my question in its own little way. Sometimes it will give me a response and then reply as me and keep going on forever. Other times it will give me a nice response and then start spouting gobbledygook and kanji. In almost all cases it just keeps going forever becoming less and less coherent before eventually cutting itself off mid sentence. I feel like it's something to do with prompt templating but I don't know anything about this. I don't think it has anything to do with the ram because I only experienced that issue in certain models.

[fresh ruin]

anyone with a 3090 can screenshare me their generation speeds?

[lavish ridge]

Is there anyone here who works directly for stable diffusion that I could speak to?

I would Like to ask specific questions about copyright and commercial usage rights for the artworks we create.

[lofty sphinx]

wassup

[young bronze]

why is there a LyCORIS folder if you're forced to move it to LoRA anyways?

[still glacier]

Historical reason. An extension used to be required for LyCORIS. Seperate functionalities, seperate codebase, seperate filetype => seperate folder. Now LyCORIS support is built-in Auto1111 (and most other interfaces).

[gritty kiln]

Guide worked great. Thanks. Any recommendations for models I should get?

[wheat spear]

hey. I want to do a like, me. but a cartoon, in the woods witht he sun shining, maybe holing a lollypop.

is there a good cartoon model?

[hollow orbit]

Hey, I'm not sure if still relevant but yes my Discord bot does it. Contact me if interested!

[graceful bramble]

I have a weird bright blue eyeshadow that has crept into my generations somehow. It is in every image even with no loras or changing models. Is there a cache I can clear without deleting all my settings or something?

Oh! I'm using forge by the way.

[bitter needle]

Hey everyone! I'm working on a large-scale 360 equirectangular video project with over 32,000 frames and need some advice. Specifically, I'm looking for help with:

Controlling latent images and noise for smoother transitions
Mapping latent control to a camera track
Tips or best practices when working with equirectangular projection in animation
I have prior experience with A1111, Stable Diffusion, and ControlNet on a similar project from two years ago, but this one’s a bit more complex. Any insights or techniques would be super helpful! I'm happy working with python, json.

I'd really appreciate some explaining on how to use your own noise layer. and how i'd go about building my own transformation program for equirectangular projections in A1111 or other ui or alternatively how I can use software like blender/TouchDesigner to take my rendered image transform it based on camera tracking data (along with the noise layer) and layer them all with the next frame from my reference video.

[fresh ruin]

is it possible to mirror de UI of FOrge/A1111? So the image preview on the left ?

[fiery badger]

Hello, I am new to stable diffusion, I tried using prompt "a cube on ground with a light source directly above it", The idea is that light source would effect shadow, I don't really want to show light or anything on top of it, could someone suggest what prompt I can use?

[quiet silo]

a doubt at the time of generating is loading the model but it takes some no and some yes and in theory for sdd or hdd disk that is faster to load?

[sleek tundra]

Hello

[azure nebula]

is there a model like chat gpt but offline to make cmments tools tips and etc for your code?

free open source?

nvm found some

[celest tide]

does anyone know which art style model this is

nvm can't upload images

[main junco]

What models are compatible with fooocus? Is it only SDXL that installed with base foocus?

[warm junco]

Fooocus uses only sdxl

But it can work with pony models too

[main junco]

What is the best way to emulate faceswap feature from fooocus in automactic1111 ?

[brisk wasp]

hello:)

[main junco]

What is the best way to change style of the picture

lie make it cartoonish

[sudden grove]

Put the name of an Artist in the Prompt or select a Style Template

[brittle oyster]

All of my attempts to us Flux seem to fail, the preview is a grey pattern shown here. The final image is just all black...what am I doing wrong? I'm using Forgeui and the flux1-dev-bnb-nf4 model. (deleted image since those aren't allow in this channel)

[fervent thunder]

.

Joined since 2022

Still remembered how I'm excited for the launch

[warm junco]

Reactor extension or IP-Adapter face

[hazy copper]

hello which a1111 version is most compatible with extentions?

[pale furnace]

https://lu.ma/ke0rwi8n

[bright ember]

Hi everyone, I would really like to create my own checkpoint from scratch. Could you tell me how to do it? Or even better, a video tutorial? Let me know

[weak smelt]

my images always come out undetailed, even at 30 steps and 2 clip and 4 million loras, anyone know why?

[somber bear]

Hello all- I am looking for help with a project relating to flux and creating seamless images. Right now, it is unable to make images that can be tiled into larger designs. If anyone is up for a project, I am willing to pay for its development.

[desert dagger]

none of the AI's make truely seamless images - that's something you need a different tool for

[still glacier]

the latest

[still glacier]

you would need millions of anotated images and gpu server farms to get it done in a lifetime.

training an already existing model is possible tho

[placid arch]

hey everyone

[still glacier]

Clip 2 won t help in that regard.
It depends of your prompt, your model, your sampler, etc
The more lora you have the more they re likely to clash with each other if you don t do it carefuly.

[unkempt hatch]

https://huggingface.co/ashen0209/Flux-Dev2Pro/tree/main

anyone tried training with this?

[boreal turtle]

Do not use any robotic GPT! Order only human written essays and homeworks -> https://extraessay.com/?key_wpg=5wpgrd

[silk condor]

What Sampling Methods and Upscale methods are you guys using and what do you recommend for Cartoon Style?

[sonic gazelle]

please do anyone know how i can setup a live face swapping on gtx 2050

[misty gulch]

I have been trying to generate Iron Man Prime for years now. Not a single AI image generator can make Iron man prime model 51. Any tips?

[unborn hedge]

could try to make a LoRa off of him yourself using art from the official comics or something? i dont think it would be that hard to do??

[oblique elk]

In total there are not much images from the prime armor model 51. So as mentioned either you train by yourself or use some ipadapter composition / style transfer magic with some of the original images

[weak smelt]

ah ok, thnx for the info. i usually use dpm++2 because euler often has waaaaay too many errors to bother. i usually use pony or juggernaut

[cosmic depot]

hi there, new to stable diffusion, is there a way to use it similar to midjourney where you generate in a private dm?

[terse crescent]

If i have a reference and i only want something inside of the black how would i set that up in flux?

[dusty trellis]

If you have a moderately good PC (memory and VRAM), you can run it locally, with the utmost privacy

[cosmic depot]

hey man thanks for the reply yea been looking into it some more this evening, have comfyui running, just have to learn how to use it now lol

[spare galleon]

hi guys

[fervent thunder]

yess i love getting malware !!😍😍

[young bronze]

who's a mod here

[stuck sun]

sounds like a fantastic scam, count me in, i love scams!

[cursive oasis]

We are a commercial company, and we currently have a paid project requirement. We need assistance with editing some images by replacing the heads of different people into specified positions. Could you please let us know if there are professionals available to help us with this task?

[desert dagger]

there's a #1092446741984444416 forum you should post this in - but this might be better asked in a photoshop discord

[cursive oasis]

😊 thanks

[sudden grove]

Can you be a little more specific with the work involved? Are you looking to adjust facial characteristics? Also feel free to message me about this I might be able to help!

[next osprey]

how do i create an image

[placid hatch]

Anyone know any resources for sdxl-sdxl base-refiner setups?

[hard swift]

gm

[jolly field]

im looking for NSFW enthusiasts to chat with, hmu

[golden valley]

Who wants to work as a moder or developer in my project?

[young bronze]

you lost me at work

[short tinsel]

hello everyone

[novel pasture]

[shy haven]

Hello All

[fervent thunder]

Hi friends, I generated a pic in stable diffusion using epicrealism_naturalS model. I liked the face/character it generated. However stable diffusion isn't generated similar images in the img2img tab when I'm passing in the original image. I didn't even change the prompt, just want different variations of the same face. any help here? Been struggling with this for sometime now...

[peak tendon]

I'm looking for somebody to make a character Lora, since as much as I look into it I seem not capable of doing it

or at least somebody to baby-guide me on it

[low moon]

Try FluxGym, doesnt get easier than that...

[desert dagger]

answered this in the #🤝｜tech-support channel

[vagrant raptor]

So I just discovered something new. This is pretty fantastic for FLUX users. It's a new FLUX LoRA that makes it Turbo. It uses 8 steps for a fantastic image. It's Alpha, but totally worth testing. https://huggingface.co/alimama-creative/FLUX.1-Turbo-Alpha/commit/b2db8dcbd15fb095cffd8ab530499e47883466e7

[warm oxide]

does anyone want to help me with my homework

I need to make a poster about a fictional character running for mayor of my hometown

dm me

[desert dagger]

just go to mage.space, make a free account, generate some images, and make the poster

[cold basin]

i'm new to AI, and I want to incorporate it into my creative medium as one of the process... I'm quite confused with the terms. what's the difference between Model vs Assistant / Artisan? Is it Model being run on my own computer and the other is web hosted? If so, what's the minimum requirement to run a computer-noded Model? I'm only researching at the current stage. What would you recommend for a beginner to start off with? Thank you and look forward to digging more about SD.

[blazing pendant]

I used Midjourney previously.

Can I prompt right here in the chat?

[desert dagger]

an assistant is usually a text LLM (large language model) that you type text to and that talks back to you. a 'model' is an overall term for all AIs. only a few models are small enough to run on your own machine. and you want an nvidia GPU

[desert dagger]

you're welcome to DM me if you want and I can try to answer more specific questions

[desert dagger]

the only prompting channels are the #artisan-1 through #1237460438229450772 channels and you'll need to read the information in the #artisan-faq channel first

[opaque stirrup]

can anyone help me generate a image correct? i have the correct checkpoint and lora but its not coming out like it should.

[desert dagger]

what interface are you running

[opaque stirrup]

Automatic1111 if that's what you're meaning. sorry i'm new

[desert dagger]

that's what I meant. you might find more people that use Auto1111 in the #🤝｜tech-support channel than in this one

[peak tendon]

I don't use flux

i'm in 1.5

[sour mason]

Hello

[sudden ruin]

Do you need anything?

[young bronze]

nah, there was a bot when I posted that

it's gone now

[sudden ruin]

Ah yeah, it can take some time for us to catch everything. Especially with time zones and jobs

[jagged siren]

Hello guys, I am wondering if there is any possibility to create different angle (perspective) of existing photo I have... I need to explain first... I have existing photo of real property, then I have sketch of building, but the sketch doesnt fit the perspective of real photography, is there any chance to transform real photo into sketch of building but retain colours, objects, of course generate missing parts based of the input image ?

[balmy spindle]

this is very hard atm, you need to somehow project the LoRA into 3d and then map it back down using NeRF onto 2d based on camera intrinsic qualities

its theoretically possible but we need someone to do the research on it lol

or u can try publish paper yourself and become famous in CVPR 😄

[fresh ruin]

guys, i got a rtx 3090, however, its doing like 3.3 iterations per second, while my rtx 3060 did like 1.4 iterations per second.

I thought i would get like 5x the speed?

like, its still good tho

[jagged siren]

yes I was searching for something, but not enough information about it, its like changing perspective of existing image, but keeping same forms

[desert dagger]

that's really a photoshop job...

[warm junco]

which webui are you suing?

[fresh ruin]

Forge

[warm junco]

okay make sure its updated, then delete the venv folder.
Also if you upgraded the gpu you may need to reinstall the driver so the webui dont think your still on a 12gb vram card

[fresh ruin]

ill do it now

ty

[warm junco]

np

[fresh ruin]

after deleting, should i run environment or run?

[warm junco]

run

[fresh ruin]

i did, however, it didnt create a new "venv" folder, idk if thats ok

[warm junco]

then something is wrong 😮

[fresh ruin]

like, its working

but didnt create anything

[warm junco]

hmm

[fresh ruin]

ah, maybe i had connected to a1111 root

"@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --opt-sdp-attention

@REM Uncomment following code to reference an existing A1111 checkout.
@REM set A1111_HOME=Your A1111 checkout dir
@REM
@REM set VENV_DIR=%A1111_HOME%/venv
@REM set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
@REM --ckpt-dir %A1111_HOME%/models/Stable-diffusion ^
@REM --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
@REM --embeddings-dir %A1111_HOME%/embeddings ^
@REM --lora-dir %A1111_HOME%/models/Lora

call webui.bat
"

should i remove everytinh g from webui-user?

mm ok its downloading now, ty

[warm junco]

looks okay

[jagged siren]

not really 🙂

[fresh ruin]

im testing again and speeds for sdxl 25 steps with 1 controlnet doing about 3.5it/s

however, i think this is the normal rate for 3090

[desert dagger]

yeah it is - photoshop and perspective warp: 10 seconds. battling with comfyUI: 3 months if not more

[quartz siren]

I think it’s a much more difficult problem since they need to basically generate a different view point of the image with new objects as well. I doubt perspective warp can solve it.

[quartz siren]

What are you using? You can boost speeds up massively by stuff like torch compile, stable-fast(probably best option?), one diff.

[devout tartan]

https://platform.stability.ai/docs/api-reference why cant I access this page?

its stuck on loading

[desert dagger]

he could train a lora on the building - but he'd need more than one shot of it

[fresh ruin]

how do i do that?

[quartz siren]

Yep you are correct, maybe they can use ctrl x as well which allows for a reference img(the building photo) and a control image(the sketch).

It’s not going to be Lora level but doesn’t require training at least.

[quartz siren]

If you are using diffusers, the GitHub has example code.

https://github.com/chengzeyi/stable-fast

[fresh ruin]

i dont understand sht about code

[quartz siren]

Ok I would just recommend sdxl lightning models if possible, they should give similarish quality while only requiring 4-8 steps. They will have slightly less varied images tho.

[primal plume]

when I add in multiple loras to my prompts I always get this error: Lora not found: WesternCartoonClassicDisney100, andav, shan

anyone ever deal with or fix this?

[jagged siren]

i actually have multiple drone shots of the area

problem is I dont have the right one, and clients are not able to shot them again 😄 😄

[desert dagger]

generative AI is deliberately coded to not make identical images of things it learned about. so first it would need to learn about that specific building, and that requires training a lora, and then you'd need to figure out how to keep it from being random. you might want to try what @quartz siren suggested

[devout tartan]

is the api page down? https://platform.stability.ai/docs/api-reference can someone please tell?

[desert dagger]

we don't support the website or API - you'll need to contact customer service

[jagged siren]

I see...thanks anyway

[desert dagger]

if you want to DM me, i can try to assist you

[jagged siren]

yea sure why not 🙂

[unborn hedge]

making art to put art into the loRA to make art easier / faster, never ending struggle lol

[high ruin]

What are the 3 best ESRGAN models for upscaling people and faces??

[low moon]

All those upscaling models ar emore less the same.. :/

I havent found one that blows all the others away

[deep osprey]

帮我画一副吸血鬼图

[near seal]

Totally new to the server. Where do I create images? What models are available? And do you offer loras? Or is this a support discord for people creating art on their home systems? Please tag me with any responses.

[desert dagger]

you need to read the information here #artisan-faq as far as creating images here goes. other than that, each channel is fairly well named. and if you have technical questions, ask in #🤝｜tech-support

[void fossil]

hello

[primal plume]

does anyone have expertise that can help me recreate a style? it's giving me a hard time

i have an example prompt to follow but my results are not good at all

[desert dagger]

what sort of style?

[cold basin]

Would be great. Thanks. I will dm you.

[primal plume]

someone made custom portraits for randomized characters in a game and I would like to recreate it, but my results are not very good even when following his prompting

i'll link to you in the images chat

[dusty trellis]

Newbie here: when I use inpaint to mask something out of an image, I get it to work, but there is always a faint color difference, that I can never fix. For instance, I masked a person out of a desert scene, and the mountains, cacti and sand behind the person came through, the person is gone, but there is a faint ghosting around where I had the mask. What (if anything) can I do to fix that?

[fervent thunder]

yeah a color match node

they usually use tonemapping

there are many versions

[keen ember]

hi guys and girls, how can i make a stable diffusion site that generate images with a specific LoRa. im trying to make a SaaS app that sells halloween and christmas decoration on images. Im having trouble with thinking out how to host a stable diffusion in my virtual private server.

[fervent thunder]

python django or flask

and then something like diffusers, comfyscript, stable diffusion C++ lib or a pytorch script for the stable diffusion part

[keen ember]

yeah ok, i am finding a developer i can ask questions to on upworks. but they all don't have the knowledge of how to do it

is there a stable diffusion service on the web that can host the gbs that is stable diffusion and i can just using those scripts to use it on my server?

sorry im 10 years experience in laravel and react when it comes to Ai Im a noob

[fervent thunder]

Fal api is good

https://fal.ai/

[keen ember]

ah thanks

[fickle drum]

hey

[ashen sleet]

gm

[worthy bone]

https://x.com/StabilityAI/status/1848720074057859268 lol

[trim nymph]

Sd 3.5 out!

[white pollen]

Hello, joined at the right time 🙂

[gusty mist]

sd3.5 but prolly still not better than flux or WHAT? 😄

[coral lark]

woman lying in grass on announcement post is a nice touch

[austere sky]

they could have selected a better gen and not the one with crooked hand

[gusty mist]

and eyebrow... and more...

[austere sky]

people see hands first

but yeah those too

[fervent thunder]

how many legs are in a woman laying on the grass

[coral lark]

as many as she can fit

[visual lagoon]

The images on the site at least have less of a plastic skin look than flux default outputs

[gusty mist]

Thats what ive looked for. I dont see any texture on the womans skin?

[visual lagoon]

https://images.squarespace-cdn.com/content/v1/6213c340453c3f502425776e/3c9a05b4-aa10-4a46-882e-1c32ac7bf18e/diverse_sd3.5.jpg?format=1500w

[gusty mist]

except for the neck maybe

[visual lagoon]

These to me look a little less plastic than the flux outputs

[versed yoke]

i cinda like that they didt tese the model. i can just downlode it now

[chilly minnow]

pretty bold to use a woman lying on grass as an image for their blogpost

[visual lagoon]

we'll see once people start generating more images

[gusty mist]

thats somehow true

[versed yoke]

i wonder why this plastic look even exists

[gusty mist]

but from what i see, the licensing looks manageable

[versed yoke]

ye

[cerulean kraken]

Let's see if they fixed it, I honestly was not expecting another launch for Stable diffusion

[visual lagoon]

dunno, dall-e 3 has it too

[versed yoke]

i dont raly care about the licence

[versed yoke]

yes it has it very strong

[austere sky]

holy its a 16 gig model

[cyan temple]

where to try SD 3D guys?

[austere sky]

ill wait for the prunes

[vernal ore]

Ah yes,
Sd 3d

[gusty mist]

hf spaces

[austere sky]

https://huggingface.co/collections/stabilityai/stable-diffusion-35-671785cca799084f71fa2838

[chilly minnow]

https://replicate.com/stability-ai/stable-diffusion-3.5-large

[cyan temple]

thanks!

[gusty mist]

thats 2 links to sd 3.5. not sd 3d?

[versed yoke]

where is the comfy workflow for it?

[chilly minnow]

ah yes, thought he was talking about SD 3.5

dunno where the SD 3D is

[cyan temple]

haha SD 3D

[gusty mist]

.

[fervent thunder]

better prompt following, lower aesthetics

[cyan temple]

found it on hugging face but dunno is this correct or not

[fervent thunder]

aesthetics can be trained by preference tuning

prompt following on the other hand...

we'll see

[gusty mist]

better prompt following would be crazy. Flux mastered it imo

[grizzled oxide]

if any staff is here, $0.04 per image w/ 3.5 turbo vs $0.003 per image on schnell is kind of brutal :c

[fervent thunder]

Did they remove the NSFW and why???

[grizzled oxide]

cant integrate based on that

[visual lagoon]

flux is too slow so i hope sd 3.5 actually lives up

[grizzled oxide]

anyone else a bit taken back by the $0.04/img on the smallest model?

x.x

[vernal ore]

It's probably about as fast tbh

[fervent thunder]

not mastered but a massive step forward

[vernal ore]

Flux is insane

[visual lagoon]

4B less parameters, wonder what level of diff that'll make to speed

[vernal ore]

They're like 3 steps ahead

[gusty mist]

flux quantized version / hyper models are pretty fast

with good quality

[fervent thunder]

17 seconds per image on default settings on a 4090 with a long prompt

[visual lagoon]

those are the ones i use but still too slow for me

the default release versions are almost unsable even on 3090

they work

[gusty mist]

unfortunate. They are consuming lots of computing

[visual lagoon]

but your pc has to be totally idle when generating

[sleek osprey]

why is there no comparison with flux 1.1 pro

[gusty mist]

im running on 3060 with 40 secs per image

[visual lagoon]

if you start doing things, the generation just stalls forever

[fervent thunder]

too scary

[vernal ore]

Cause SD wants to look good, not bad

[visual lagoon]

what size images

[gusty mist]

but not the base model *

[vernal ore]

Flux 1.1 pro is a huge leap up from dev

[gusty mist]

it was 1024x1024

[gusty mist]

oh yeah thats true. absolutely forgot about that one

[vernal ore]

Anyway I'll.be enjoying flux dev, as its almost the same as sd3.5

[visual lagoon]

i wonder if a 5090 would be able to run that locally

[gusty mist]

one funny thing in your bio " has overpowered pc" 😄

[visual lagoon]

i hope so and they decide to release the weights

[vernal ore]

Well
I lost my pcpartpicker

[soft sparrow]

so

is 3.5 good?

[vernal ore]

But it's 64gb of ram, an MSI 4090 and an i5 13600kf

[soft sparrow]

you need that much to run sd 3.5!? it's just 8b params

[gusty mist]

nice 👍

[soft sparrow]

or will community just optimize the shit out of it like is always done?

[vernal ore]

No one said you needed that much

[gusty mist]

does the 4090 comes with 24gb vram?

[vernal ore]

Yes

[iron willow]

yes

[fluid stone]

hello guys

[gusty mist]

Man im sitting on my lil 3060. However, if ai will continue like this we will need 8090 soon

[fluid stone]

wasup

im new here

[versed yoke]

maby you will be bale to run the smaller model thatreleses soon

[versed yoke]

hi

[fluid stone]

im an artist im 78.6 years young and i like turtles

[vernal ore]

I wonder why tf they made stable diffusion 3.5 medium when there's sd 3.5 large turbo

They both perform literally as good

[gusty mist]

oh man ive given up. Its cloud computing for me from now on

except for highly quantized models

[noble field]

So how anti NSFW is the new model?

[fluid stone]

can anyone rate my picture that i justed drew

[gusty mist]

most of the people here dont respect "real" art. LOL

[vernal ore]

Also, how is it possible that sd 3.0 is on par with flux schnell?

According to their measurements

Cause that's gotta be skewed in a way

[gusty mist]

[vernal ore]

Look at their announcement page and scroll down lol

[lyric snow]

Is SD 3/3.5 available in auto 1111 yet. Had a break from ai so I'm not up to date again.

[gusty mist]

i believe you.

i just say its cap

[vernal ore]

... It released like an hour ago

[gusty mist]

sd3 == flux schnell. like on what earth

[versed yoke]

i think similar to flux

[noble field]

Stuck with SDXL forever then?

[vernal ore]

Just train a new model

[versed yoke]

sdxl also is like flux maby even stricter

[gusty mist]

or nsfw unlock loras

[vernal ore]

Lmao

[gusty mist]

id say its stricter

[versed yoke]

yes

[low moon]

that omnigen is good, if it does all those things it promises its a game changer

[versed yoke]

for me the model was not working and i got errrors and i thogh wtf but then i notivced i had another ai program also running lol

[versed yoke]

lol funny that you mension it that was the ai that was open

[low moon]

lol

[noble field]

But it's been decensored pretty harshly, while there's nothing really conclusive for flux as far as I can see But obviously I don't know anything about it, I'm just amateur users

[versed yoke]

omnigen needs like 24gb vram

[vernal ore]

So does sd 3.5 doe

[versed yoke]

but they can still improve it a lot by reducing acuracy

[vernal ore]

I think

[versed yoke]

3.5 currently needs 17gb but maby its ofloding something to the system ram

[vernal ore]

Yeah, no 17gb cards

[versed yoke]

24gb cards are there

[gusty mist]

I dont really understand how they even managed to create GGUF of flux. Would this be possible for omingen too? I havent seen a single tranformer model which has been quantized to gguf except for flux

[versed yoke]

also i am running it in the hardest mode

[orchid wyvern]

What is Stable Diffusion guys?

[versed yoke]

yes it would be even easyer for omnigen becasue its based on the phy llm

omnigen is a llm that is finetuned so it can las make image tokens

[gusty mist]

Ohh

isee

[versed yoke]

so it works very difriently

but it has cool abiletys like context and beeing able to see multible images.

and adding new cababiletys is easy you just have to make training data. everything else is just lerned

[gusty mist]

That as a single checkpoint? No agentic process?

[shadow tinsel]

Hey, can someone tell me what encoder 3.5 needs?

[versed yoke]

but the qualaty is like sdxl or sometimes worse

[lyric snow]

Do you mean it released on auto 1111 an hour ago? Last time I used auto sd3 wasn't available yet.

[gusty mist]

ill check it out

[versed yoke]

here is everything you need https://comfyanonymous.github.io/ComfyUI_examples/sd3/

[ebon fiber]

so flux dev makes better images but prompt coherency is worse then sd 3.5?

[versed yoke]

i have to test to make sure but there data says that the astetic cuallaty of flux dev is a bit better.

but they say finetuing might be easyer

for sd3

[ebon fiber]

and model is smaller i assume so faster?

[chilly minnow]

flux is hard to finetune, so if the SD 3.5 is easier to finetune, the community models should be on par with flux or even better

[ebon fiber]

should take a few days still I guess xD in any case i see from the cover she can lay ont he grass so thats a big step up from 3.0 😄

[halcyon granite]

is SD still releasing models?

[peak orchid]

people forget that base SDXL and base SD 1.5 suck compared to the popular community finetunes

[pale juniper]

brother, did you read the announcement

[chilly minnow]

let's not forget SD 2.0 which was so bad that everybody forgot that it even existed

[halcyon granite]

yeah I'm just confused, I thought it was over for SD or something

[pale juniper]

no, they hired james cameron, took some time off to perfect the model so they didn't release another crappy one like 3.0, and just released a new one

[shrewd nebula]

hi if i am using a amd gpu for stable diffusion is it not possible to also make use of my cpu? also amd

[halcyon granite]

film maker right? interesting

[long talon]

Hey folks. Is there any talks going on about training a diffusion model in a similar manner as INTELLECT-1 ? I have been raving about the need for decentralized training for well over 3 years now and it finally seems (?) like it is happening

[forest trout]

What's the 3.5 license like? If it's overly restrictive I might just stick with flux.

[errant yacht]

The Stability AI Community license at a glance

We are pleased to release this model under our permissive community license. Here are the key components of the license:

Free for non-commercial use: Individuals and organizations can use the model free of charge for non-commercial use, including scientific research.

Free for commercial use (up to $1M in annual revenue): Startups, small to medium-sized businesses, and creators can use the model for commercial purposes at no cost, as long as their total annual revenue is less than $1M.

Ownership of outputs: Retain ownership of the media generated without restrictive licensing implications.

https://stability.ai/community-license-agreement

idk if that's the same or not

the link says last updated in july, the post says they're releasing their new license