💬｜general-chat

[tall bolt]

it does a really good job but the edges are blatant

and its like lighter

[glass gulch]

"Does anyone know models for image generation, SD (not SDXL), with a nice or permissive license that allows creating content and publishing it without any other third-party models or licenses involved?"

[faint timber]

@glass gulch I think PhotonLCM might be a good choice for SD 1.5, but you might as well go up to SDXL to double your resolution. The Proteus model is trained on open source images, so it's a good choice if your making images for commercial production.

[hallow plaza]

Are there any ai models good at making sprite sheets?

[glass gulch]

@faint timber I'm mainly interested in the license, because I want to create different types of content without legal consequences, especially if I publish them but do not profit from them.
I've checked the model "Photon - LCM", and it appears to be released under the CreativeML Open RAIL-M Addendum license. I prefer the base Stable Diffusion models because they are easier to use and require fewer add-ons or extra steps.
As for the Proteus model — its license is even worse for me and too restrictive. That's why I’m happy to use any model that isn’t SDXL, Flux, Pony, etc., but is a regular base SD model — as long as it has a good license.

[glass gulch]

I checked these models on the Civitai website, so if anyone has a model with a cool SD 1.5 license to recommend, I’d appreciate it in advance, as well as any help.

[vale ocean]

Sup all!, can i also ask a question about AmuseAI?

[verbal dome]

i see no such "outpainting" in webui forge neo version. How do people get it installed?

[fervent thunder]

Hey guys, I’ve been working with flux-fill-pro and flux-fill-dev but I keep running into an issue. Even when I give a very clear prompt for the background (and no figures), the model often hallucinates people, animals, or objects in the masked region. This happens no matter how big I make the mask.

Has anyone figured out a way to make Flux reliably just extend existing background textures without introducing new figures?

[warm junco]

Its at img2img Inpainting and then at the bottom under scripts there should be outpainting.
Or you install the openoutpaint extension. (But idk if it works)

[fervent thunder]

I've switched to using google/nano-banana instead of any inpainting models.

[verbal dome]

i found a helpful tip from my experience: how do people change checkpoints? I found my pc freeze or at best - take around 10min if i switch one checkpoint into another in forge neo, but if i close it and in 30sec open then changing takes no time and is superbly fast. why is it? i dont know.

[verbal dome]

is it on offline computer? because that would be amazing

[primal pike]

Hello friends

[primal current]

Does anyone have a SUPIR workflow with WAN 2.2 t2i?

[vital hearth]

hi

[runic patrol]

So i am wanting to move away from illustrious. What do you all suggest, what makes amazing images these days, i don't mind downloading the latest stuff

[fervent thunder]

no I am using apis

[karmic glacier]

.

Friendly PSA

Mod 4bus3 in other severs and victim harassment in shared spaces

[still glacier]

Not sure how this has anything to do with SD (or any of its staff and people in here) ... Nor why you jump in this server to copy paste stuff like that without any context and tried to @ everyone for this....

Removing your copy pasta because this has nothing to do with SD or any current discussion and is promoting some dodgy sectarian looking discord community. (rule 5)

Any person that is harassed here (or any other discord for that matter) is free to report it to any mod / community guide / etc and discord itself.

[karmic glacier]

It wasnt about harrassment here. It was drawing attention to harrasment happening in other groups by a mod who happens to like to use multiple alts to harrass and stalk people.

But funny how the various discord communities don't like to talk about abuse

Even though we know discord is famous for not doing jack about reports

But way to assume and make me highly suspicious of this server

Later gators

Catch you on the flip side

[still glacier]

Still at work so I ll make it short. Harassment bad. Ok that was short enough.
I ll repeat it to be sure, if anyone here feel harassed here we (mods, community guides, stability staff, etc) will examine and timeout/ban accordingly. It has happened in the past but fortunately is not that big of an issue here.

If harassment comes from a mod, go to another mod, go to stability.ai staff directly, go to discord (yes they could handle things better). I believe there is enough adequate people to talk about those things in here and many of us are not even related, barely even acquintances. So that should make things easier and minimize conflicts of interest.

Now with that said, you can't enter a bakery and start yelling "Hey do you know that soooooome bakeries piss in their floor to save some water" and expect the cashier to not give you the stink eye. Jumping on random discord and dissing vague accusations about harassment with no evidences to back anything is bound to put you in the spotlight.
Moreover this kind of discussion would be better suited for #🌶｜off-topic .

Also. promoting other discord servers without approval is not a thing here. Regardless of the astral projection, esotheric science, conspiracy, etc themes that I saw in that server. From the quick look I took at it, that discord raised more red flags to me than anything else. It felt like a discord looking out for vulnerable people to pray upon rather than one that could help them. I might be wrong, I only took a brief look at it. But in the end it doesn't matter because rule 5.

There are many discords, websites (7cups, blahtherapy, etc), local associations and phone lines qualified to talk about subjects such as harassment and able to help anyone with it. I've even been part of some of those in the past. Personally I would always recommend getting help from some close friend / family member if possible before reaching out to online strangers.

I think that's all I have to say and I will probably not answer anymore about it. Especially not on #💬｜general-chat .

[alpine bear]

Guiz banned me over a small disagreement. Fruit taunted the entire anime AI community with the Miku role. Caz unjustly deleted the miku role. Aether smelled the malpractice from a mile away and dipped. Then we lost the otter. Still haven't recovered from that one Truly, nothing but devils on this island.

https://tenor.com/view/reiner-braun-braun-reiner-reiner-wake-reiner-tired-gif-19555964

[small kelp]

hi

[primal current]

My friends, do you know of any AI (preferably open source) that can dub from English to Spanish or another language? I use Eleven Labs, but it's very expensive and not that good.

[floral umbra]

So damn exited for the script i'm having a.i slop hallucinate forth!

It'll be a python venv hub. And you don't have to do squat other than python hub_cli.py install https://github.com/lllyasviel/stable-diffusion-webui-forge, and it will do the rest for you Missing a python version? No problem. It will install the latest one compatible for you. And install all dependencies in one go for you too in that same command. And deduplicate all files if you got other projects like comfyui by making a full information database of all the files it fetches and installs, and if you remove say forge, but comfyui had symlinks forge held? Then it'll auto move those original files over to comfy and delete the symlinks, as well as re-assign symlink with new paths for other projects that happens to also need those files Saving space and time.

[worn pawn]

Hey, I’m looking for someone who can train LoRA/SDXL models (influencer / selfie / NSFW style) with true photorealism – consistent identity across a series, no AI tells. I’ve seen WAN 2.2 making progress in video realism, but for my use case I need LoRAs that hit that same level of believability in still images. Anyone here got concrete samples to share?

[verbal dome]

training lora aint that hard, just takes a ton of time

but im pretty new myself, trained one lora successfully so thats a nice milestone

[brave matrix]

Hey everyone - I've been vibe coding this project using gpt5 for a few months and am getting ready to go public with it. I would love if anyone would be down to test and give feedback: https://www.sloptok.ai/

[frosty hedge]

awesome

[trim cliff]

bjbfkdjs

[fiery ore]

Yeah, with a clean dataset and careful training, SDXL LoRAs can get really close to WAN-level realism. The identity stays solid across selfies and NSFW shots if the setup’s done right, and you can avoid most of the usual AI tells.
Can I dm to show you some of my recent work?

[worn pawn]

thanks for the response, yes of course!

[orchid cloak]

Hello guys,
I have to find a guy with NSFW for my character and it will be a long tern partnership with payments but i need some samples and i will provide images.

Somebody interested?

Add me if it intersted

[regal fulcrum]

Got a question for any folks who mostly dabble in 1.5 or XL, and also use hypernetworks.

Which web UI do you prefer? A1111, Forge, ReForge, Comfy, Swarm, something else?

Forge Classic/Neo looked interesting but it removed hypernetwork functionality, which unfortunately is a dealbreaker for me.

[winged rapids]

hello

[serene shoal]

I wanna do a project with pros.

[abstract quarry]

who needs hypernetworks?

loras are more stable and better

[sterile sparrow]

Hello

[serene shoal]

Who want to work with me?

[unborn hedge]

Sora 2 anime stuff seems interesting, wonder if it'll actually be viable one day

[inland vapor]

What text to video platform is the best

[primal current]

Does anyone else experience ComfyUI's venv crashing and not recognizing dependencies when trying to open a new Runpod?
I've been working with storage (1 TB) for a month now, without any issues, but since Runpod crashed 2 days ago, this garbage is useless.

[fervent thunder]

hello

[urban notch]

Hey everyone, new here, what is the best ComyUI Template to use? Thanks

[limber steppe]

Hello

[bronze merlin]

Hi

[tepid stump]

Hi everyone, I’m Virgilio (Vili).
I’ve been working with AI, visual effects, and creative coding, and I’ve used Stable Diffusion since its early releases.

I’m interested in exploring hybrid workflows (AI + VFX + procedural tools) and experimenting with open-source pipelines like ComfyUI. Excited to connect, share experiments, and learn from the community here.

[rough bone]

Would anyone be able to help me shoot setting up a workflow for comfy ui in stable diffusion? I keep running into missing node errors and I'm not sure how to resolve

[proven badge]

best song... Moon Walker - Monopoly Money

[prisma yew]

guys, which model is good to use locally in phone?

[safe elbow]

Hey got question, what model of AMD graphic card do you recommend for AI?

[warm junco]

Hey, would recommend the either the 7900xtx for its 24GB vram or the 9070XT for its newer features and support

[safe elbow]

for qwen model suitable can be 9070XT ?

[warm junco]

Idk, I think with smaller versions it can. I tried qwen on my 7900xtx and it was pretty slow

[safe elbow]

how slow?

[warm junco]

Would need to test again

The problem with qwen, chroma, wan, flux is that they are huge and not that performance friendly.

[safe elbow]

Ohh

[warm junco]

They still work but in comparison to sdxl based stuff its slow

Can generate an upscaled sdxl/illustrious image with multiple loras in 16 seconds.
For flux it takes like 60 seconds for a normal image and upscaling is not possible in the same task without getting oom error.

[safe elbow]

that is fasten then what I cna do now XD

[warm junco]

What's your GPU right now? xD

[safe elbow]

32 GB Ram

vram 8 gb

and I am building now PC since one i have has GTX 1080 Ti

[warm junco]

Ah okay my old PC has a 1080 and then I switched to the 7900xtx

[safe elbow]

well i hoped to find one wiht 7900xtx but it seems they are lackingnow and then i thought of 9080 XT that may come next year too

but idk if it is true it will havve 32 GB vram

[warm junco]

Yea that would be the best card of them If its true xD

[safe elbow]

To be honest idk if 16 gb is good too Xd

truth is I am newbie when it comes to generating images

[warm junco]

16gb is good for the start but if your focus is on ai stuff then 24 or more would be indeed better.
But best amd price/performance GPU is still the 9070xt

Hard decision xD

Boath are also perfect for WQHD Gaming and FSR4 works now on 7900xtx too

[safe elbow]

tbh I have option C which is wait for the 9080 Xt tooo

[safe elbow]

damn

7900 is like 1080

[warm junco]

Yea if you can wait, AMD is presenting new stuff on January 06.2026

[safe elbow]

I hope it will be as good and cheap as other rx models

and compatible not like rtx ....

[warm junco]

Yea true, but AMD is improving compatibilities with ai stuff currently. They probably present that too on that day.
For the price of a 32gb rx 9080xt im not so sure. Maybe 800?

[safe elbow]

where i live Rtx is overprices as hell

5080 cost here around 3000 $

idk whatthey put in there gold or hell knows what

[warm junco]

Wtf 3000?! crazy

[safe elbow]

it is due totthe taxes and other bs stuff goverment can do to get cash

since the "corruption" must go on

[warm junco]

damn :/

[safe elbow]

Btw what processor you can recommend?

Amd or intel?

[abstract quarry]

Amd processors are currently better than intel

[rose eagle]

hi

[warm junco]

AMD for sure! Currently the best to go for is the 9800X3D or the 7800X3D that i have, perfect match for a 7900xtx or 9070xt, but a 9700X would be okay too.

For Gaming AMD is currently the best. Go with an AM5 board as its the latest Gen for the new CPUs, and pair it with 2x32GB DDR5 6000MHz RAM for max performance.
I only have 32gb RAM and when using large AI models it gets fully used, so an upgrade is planned xD

[safe elbow]

Ram can always be updated that is easy to do

Worse for changing the graphic card xD

[fallow wraith]

Hi
anyone got a sora 2 invitation?

[eternal arch]

this is incredible man!!
watch with sound
https://youtu.be/ReBBflR0oi8

[austere relic]

Did SD release a roadmap yet?

Or any announcements regarding better image or video models? Sora2 equivalence is probably 2-3 years ahead though

[regal fulcrum]

How do I privately report an unsolcited DM request from another user in this server?

[still glacier]

You can contact anyone with the CommunityGuide role. As example.... me.

[regal fulcrum]

DM sent

[rain token]

Hi everyone
Is RTX 4070 with 8GB VRAM good enough to run SD3.5 Medium? Also, am I missing out on something if I am not using the TensorRT version (apart from performance boost)?

[warm junco]

Yes medium should work easily but its not good. Your better of with using illustrious based stuff currently.

Wouldn't setup tensorRT as its not compatible with everything

For the best performance I would recommend using Forge Neo

[spring spear]

hello there, i need some help, im using ComfyUI and want to use one of those Slider LORAs but they dont work for me, i´m either to dumb or i missing something if someone knows how to run those properly please tell me how (im using the normal LORA loader of ComfyUI)

[atomic lake]

Has anyone here using Pathfinder KM LORA? I can't make it work no matter what I try, results are super distorted

[fiery ore]

Slider LoRAs don’t work with ComfyUI’s normal LoRA loader because they’re designed for special nodes that let you adjust parameters like style or expression strength. You need to use nodes such as LoRA Block Weight or Impact Pack’s LoRA Detail, depending on the LoRA type. After installing the right custom nodes through ComfyUI’s manager and restarting, load your LoRA with one of those special nodes, adjust the slider value to control intensity, and make sure the file is placed in your models/loras folder.

[spring spear]

thanks i will look at it

[abstract quarry]

this is not true. Slider loras are normal loras. The "slider" is the strength you use

the special thing on sliders is the way they are trained. Besides that they act as normal loras. You might have to use higher strength, some sliders start showing an effect with strength>2, but that depends on the lora @spring spear

[flat ibex]

#💬｜general-chat hello

[quartz pelican]

Hey what did I miss?

I don't feel motivation to learn about generative AI anymore, but I don't want to give up the idea either

The whole setup is just so hard to do and I don't know where to start

Is THIS the difficulty of generating pictures you guys keep talking about? Just fucking with installing the program and not working with it?

[abstract quarry]

depending on the tool

comfyui has most experimental features and is easiest to extend, but it's complicated to use and has a skew learning curve

invokeai on the other hand has an easy installer and usually works out of the box. It's not more complicated than any other graphics program

[quartz pelican]

Of course it isn't. None of it can stand close to Photoshop. Also it's name makes me vomit

Invoke
Tell me that you outsource your thinking to ChatGPT without telling it directly

[quartz pelican]

Whatever. If it's so easy to start then I'll take it

Where do I get it. This garbage better be for free

[abstract quarry]

dude, developers of such tools might also vomit about this mind set of some users

google is your friend

[quartz pelican]

Of fucking course

[quartz pelican]

Also thank you for help

[unreal raven]

hello all nice to joing this community

[neat turtle]

NIGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGG

A

[proven badge]

there was a person who uploaded videos of wwe game videos for 7 years getting like 10 to 60 veiws but as soon as they made ai goon videos of wwe and stuff they got up too 2,000 veiws to 1million views instead 🙁 https://youtube.com/clip/Ugkxv8_YDCbnQiFd5MXmno5fDZd9E2nwNyZi?si=NAohFkdm3dj__nwW

[abstract quarry]

who knows if all those views are not ai bots themselves xD

I think that's what we are moving forward to: a world where ais that pretend to be artists are generating content and ais pretending to be users are consuming them. Huge waste of energy.

[unborn hedge]

locally training loras seems so hard man, i installed kohya and its like looking at alien technology, d*mn you civitai for not working

[still glacier]

Not the place for that

[rich arch]

hi, friends. nice to meet you
I have a new idea to collaborate with you that could be profitable.
how about it?

[abstract quarry]

if you don't like command line and config files, just use the many ui options

for kohya there is a webui

[white shore]

hi guys good morning, afternoon, or evening.
i have a question. i was on civitai and i saw a catagory called pose. and i want to know how to use it. any tutorial vids or guide?

[frosty hedge]

use the "Pose" category by loading pose images into ControlNet in your Stable Diffusion UI, then enable the OpenPose model and apply it to your generation

[white shore]

ah ok.

where is the controlnet?

i dont seem to find it,

[frosty hedge]

in automatic1111, you can install it by going to Extensions

install from URL, then entering the repo
https://github.com/Mikubill/sd-webui-controlnet

after installation andd restart, ControlNet panel will appear below your main image generation

[white shore]

Alr thx!

[frosty hedge]

you're welcome

[sleek forge]

Does anyone could send me a Sora 2 invite?

[urban notch]

did you get one yet?

[earnest vine]

Guys i need suggestions for downloadable models that will give me mudjourney esq style (not flux)

[fierce wyvern]

Having trouble getting a sora 2 invite if anyone has one... thanks in advance!

[still glacier]

ok bye then

[orchid current]

Hi I’m a Senior developer with experience in AI/ML, Blockchain, Web, and Mobile development. My expertise covers:

• AI/ML → LLMs, RAG, Computer Vision, NLP, Speech AI, TensorFlow, PyTorch.
• Blockchain → Smart contracts, DeFi, DEX, NFTs, Solana, Ethereum, Polygon.
• Web & Backend → React, Next.js, Node.js, NestJS, FastAPI, Django, GraphQL.
• Mobile → React Native, Expo, Flutter.
• Databases & Cloud → MongoDB, PostgreSQL, Firebase, Supabase, AWS, Azure, Docker, Kubernetes.

I’ve built scalable systems combining these technologies — from AI agents and automation tools to blockchain dApps and modern SaaS platforms.
You can check my portfolio here: https://chee-portfolio.onrender.com
If you need developer for a current or new project(business), please DM. 🚀
Thanks

[midnight violet]

hey everyone. I'm trying to do inpaint with an sdxl model + a lora of a character into a specific background image. Now I cant seem to achieve that. I use swarmui. Do I have to get better at my control of specs such as denoise and mask blur etc or is there a better way to do it ? I usually do a remove background of the character and then paste it on said new background that I want but that has problems when I want to animate, as the I2V video gen AI will see the subjects body is not blending well (in the scale of small pixels), for example on a chair its sitting. It will see it as not sitting and the subject my start to fly away as the AI will see it floating, even if by a few pixels, etc. I have discovered that it matters to do a good mask too, not just but a rectangle box where you want the person, and actually try to give it think the mask legs, arms, a head. But I still cant get a good result and am a bit lost. Should I up my prompt game ? Should I mention the background as well ? What to do ? Any helps and tips will be gladly appreciated ! Thanks everyone !

[untold vault]

[sharp mortar]

hey everyone . i am working on real time avatar generation of face image into 3-D disney style pixar baby style avatar of any person uploaded photo or image .i am using dreamsher-8 model with the baby face lora's weight and ip adapter for face identity preservation but i am failing to achieve the 3 d disney pixar style avatar of older person above age of 15years tried using stablity sdxl latest model but still failing to achieve that is there any one who has worked on this use case . what do i do ? any help will be gladly appreciated!

i have also tried using inpainting pipeline to achieve the same however still not able to achieve the target output

[glossy imp]

hey is there a discord server that assists with people who are trying to do nsfw art? I want to get help with my renderings but I dont think this is the right discord haha

[fervent thunder]

😏

Show me hihi

[abstract quarry]

edit models are really good for such kind of things. There are also loras for edit models that are finetuned on object insertion tasks (you copypasted a image into another image and want it to blend in and add shadow and so on)

[honest sun]

I'm totally new here, but have been creating Ai images on MJ, Sora, and a couple of other models for a while now. I've been considering getting a new computer system and installing SD locally. Looking around here I see lots of good images, but I haven't yet seen anything that looks like high quality photographic image. Is this because of the creators' intent, or because of the limitation of SD?

[atomic mortar]

Mostly because most people here are using stable diffusion finetunes, try looking on civitAI images and the highest rated ones

Youll find photorealistic models of all companies there

[honest sun]

thank you for the information.

[undone garden]

I'm looking for a comfy node that randomly picks images from a folder as input, does anyone know of one? I had one that worked just right but it broke after I updated UI and dependencies

[faint timber]

The way I do it is point a LoadImages (Path) node at a folder and get the count. Then use the count as a max value on a random node that you drive from an incrementing seed. Feed that value into the offset of a second LoadImage (Path) node pointing to the same image folder that only loads one image.

[spark sage]

hello everyone

[spark sage]

yes?

[untold vault]

[halcyon granite]

I'm using get actors > set array > set view target

I have an issue where sometimes when setting the view target to the player pawn, instead of it's camera it switches to 0,0,0 in the pawn clipping inside the body.

I have isValid nodes setup but it's still happening somehow

[rotund dune]

👋 Hello! I'm a Senior Full-Stack Developer & AIOps/Agent AI Specialist with 9+ Years of Experience
I architect and deliver powerful, AI-enabled digital systems—from scalable full-stack apps to Agentic AI workflows and automation pipelines. With a deep foundation in Python engineering, AIOps, and low-code/no-code agent frameworks, I help startups and enterprises streamline operations, enhance UX, and build future-ready platforms.
If anyone looking for a developer, please let me know.

[warm junco]

Hey, delete your messages and read the rules

[bright ember]

Hi everyone, I want to create a new story about a favorite anime girl. Who would you like?

[atomic mortar]

Rule 5 bozo

Unless its another bot

[novel ridge]

Where would one get a Sora 2 Invite Code around here? I see people asking about it in various channels, but I don't wanna spam the place up too much myself.

[warm junco]

hi, delete your message and read the rules

[deep echo]

Socialise? Anyone?

[oak dock]

does anyone have some tips on how to use a pic and only change the clothes on the legs for example? it always looks like shit what can i adjust more in inpainter? i watched so many videos

[atomic mortar]

Rule 5, man whats with the bots lately

[craggy rune]

qwen image edit can change text in an image as well as clothing 📝

[brittle rain]

hello

[craggy rune]

hi

[lean timber]

ComfyUI should be renamed to BrittleUI. I spend wayyy more time trying to fix others workflows, or my workflows that break in some update for some obscure reason, than I spend actually generating images

[tiny shale]

hey guyssss

[coral sigil]

so is comfyui a buggy broken mess for everyone or just me

i find it insane that anyone is actually using it for anything

[atomic mortar]

Never had this problem with comfy, though if you use custom nodes you shouldn't update

[timber flame]

Good morning everyone, I'm pretty new to all those things related to stable diffusion AI, with your help and experience could it be possible restore wall painting renaissance periods with shadows and harsh shadows damage? like in this image?
https://www.wga.hu/art/m/michelan/3sistina/5spandre/10_4pe4.jpg

thanks to eveyone in advance for right workflow clearly if it's possible to do.

Target will be: restore this old wall painting renaissance era without changing anything about the image, just repair it, clean it, remove harsh shadow area.

I've a 4090 RTX and stable-diffusion-webui correctly installed and set up

[sly fern]

;)Love & Unity To All! It is a pleasure to be here! Ready for this journey of enlightenment

[floral umbra]

Hashing: 1753697074.safetensors
File already in database: 1753697074.safetensors (292MB)
Hashing: 1753787382.safetensors
Hashing: 1753787993.safetensors

My dedupe and lora sorter script working beautifully

Scans model hash, as well as notes name, path, any safetensor metadata it may have to a sql library and for every model it has scanned, it compares those models to the sql database it made to find dupes no matter if they have a different name or anything :P. So i expect my vast farming of models from huggingface to drop significantly as i had a bunch of dupes lol.

Actually, one addition to the script i just realized, i need the deduper to also take deblating into account. Like those 4GB older models that got pruned to 1.9GB for sd 1.5 for instance

[deft pecan]

How to generate an image here?

[atomic mortar]

Hi, check out #artisan-faq

[deft pecan]

Thanks

I have not found any free plan at artisan but there are only paid subscription plans

[atomic mortar]

There are a couple of free credits on civitAI but unfortunately everything thats free has a credit limit

Or you could look into running it locally if you have a powerful pc

[cinder elm]

I tried to change the background of my image but using inpainting give me bad results (bad composition). I use flux fill workflow from the photo. Any tips how to get better blending?

[steady atlas]

this sounds great

[rugged sundial]

Hi everyone, I need your help! 🙏
I’m trying to restore the quality of a hand-painted artwork. But after countless tests I can’t get any satisfying result. The output always looks almost identical to the original. I’m still a beginner, but I’ve seen some incredible AI restorations online, so I know it’s possible. I just can’t figure out what’s going wrong 😅

My original image is a hand-painted artwork, very blurry and noisy. I’m trying to get a cleaner, sharper version, keeping the original hand-painted texture and colors. I tried both cleaning and upscaling with AI to recover details at 1:1 scale. I can upscale it ×4 successfully, but when I bring it back to the original size, it’s still as blurry and noisy as before. It’s driving me crazy 😅
I’m looking for a true restoration, not just pixel upscaling.
I have a NVIDIA RTX 5090, so compute power is not an issue.

My tests so far:
-Automatic1111 + SDXL + Refiner + RealESRGAN 4x → Tried different denoise/CFG/steps, still blurry and noisy.
-ComfyUI “Flux Kontext Dev Basic” workflow + LoRA “Flux Kontext Restore Painting” (from CivitAI) → almost no improvement, still soft and messy.
-ComfyUI RealESRGAN_x4plus → good upscale, but the same blur/noise remains.

I’d love your advice 🙏
If anyone knows a ComfyUI workflow or model specialized in restoring painted artworks, I’d be incredibly grateful!

Thank you and looking forward to read you!

[timber flame]

I've asked something similar here waiting an answer too

#💬｜general-chat message

[oblique elk]

Sorry mate but / sharpening upscaling is kind of easier compared to your question. Color, structure etc. are completely shredded in your example shadow image.
So most of the models which extract colors, line work, depth, … would simply be irritated by the amount of “noise”
Not sure if there are special models out there for reconstruction….

[worn burrow]

Anyone have experience downloading Stability Matrix to Linux?

[true shuttle]

Hi bro

Have you used i2i denoise?

[true shuttle]

I've been feeling the same issue and fixed that with i2i denoise model
Let's chat on dm

[worldly basalt]

How can I generate fine tuned ui for mobile apps? I have extensively searched loras too but couldnt find anything with a wow factor. How troublesome would it be to train a lora on 3-4 mobile apps ui?

[timber flame]

No mate

https://postimg.cc/V0Q8gG82

AI is perfectly able to do something like that, I was searching something to use offline. I was thinking someone here better than me could help or if experienced suggest right stable diffusion generator with img2img ability capable to do that

[rugged sundial]

Yes check on #🏞｜general-with-images I add some good answer! 😍 thanks @oblique elk

[vestal dew]

Is there a local version of somerthing like Nana Bonerana? I tried it...it made a tinker's bollux of the photo I tried to edit

QWEN perhaps?

[undone solstice]

the user said "not u" which is a bit confusing. They might be responding to something I said or pointing out that it's not me. I should acknowledge their message in a friendly way without overcomplicating it.

got it. The user said "not me" and wants a friendly response without overcomplicating. I need to make sure the reply is simple and acknowledges their message.

[worldly basalt]

How can I generate fine tuned ui for mobile apps? I have extensively searched loras too but couldnt find anything with a wow factor. How troublesome would it be to train a lora on 3-4 mobile apps ui?

[still glacier]

@worldly basalt please don't do whatever that bot scammer asked you to do.

[worldly basalt]

Of course not 🙂 thank you!

[pine path]

Interesting paper dropped. Makes use of DDT, looks promising. This paper does use DINO which I think someone asked about when I posted the DDT paper here several months ago.
https://arxiv.org/abs/2510.11690

Diffusion Transformers with Representation Autoencoders

[unborn hedge]

does comfyUI natively support Wan 2.2 Animate yet? i heard there was suppose to be a like template workflow but i updated comfyUI and checked and i cant seem to find it?

[oblique elk]

Since comfyUI separated the frontend and the backend you need to upgrade front end also.

[unborn hedge]

im just doing comfyui update from the manager, isn't that doing everything?

[cinder elm]

Hi,
I’m trying to place a glass bottle in a new background, but the original reflections from the surrounding lights stay the same.
Is there any way to adjust or regenerate these reflections without distorting the bottle itself?

[unborn hedge]

update i had to reinstall everything from scratch but now i got the wan 2.2 animate template, i ran the workflow to get the first frame of the input video but now idk if its doing anything or not

nvm i think it is? i see the percentage bar going up like its doing something so idk

my favorite part is when i tried to update a node and then everything broke and i had to proceed to reinstall comfyUI from scratch because of course i would have to do that

[floral oxide]

what are some tools/tips you use to help you character and image you want ?

[atomic mortar]

Comfy manager is known to break comfy a lot

[unborn hedge]

yeah well thankfully i got it to work

[atomic mortar]

Flat 120€/hr rate but otherwise not interested sorry

[rugged vortex]

I've been finding reversing the frame order through seedvr2 seems to preserve details better and avoid the detail ramp up I was seeing when feeding forward, ? anyone else deal much with upscaling

[atomic mortar]

@vapid dove spam/scam bot

[still glacier]

lol yeah they were already suspicious this morning

Let me get my hammer

Cleaned.

[atomic mortar]

Nice

I baited them in dm's but seems like their bot isnt too advanced

[vale path]

Hi everyone,
Looking to chat with and learn from people building AI for GTM (sales agents, AI lead gen and outbound, call intake, customer success, RevOps, and GTM automations).
I’m keen to hear about your development processes and any other insights.

Please comment or DM me! Really appreciate it!

[sweet rover]

Guys, is there a tool like blender for 3d images and videos generation?

[dawn mulch]

anyone used krita ai before? its not detecting my controlnet models. i need halp

[winter shoal]

https://docs.interstice.cloud/models/#downloading-models check that all the models are as the plugin requires

in the right folders and so.

if using comfyui, that is.

at least the clip_vision model needs to be renamed also.

[viral kiln]

is stable diffusion still working on audio models?

is Stable Diffusion Audio 1.0 the only open source audio model by stable diffusion?

[vestal dew]

I am so disappointed in local , open source music generators like YUE.

[atomic mortar]

Im disappointed in big corpo keeping the better models closed source and happy that theres companies releasing models open source***

[vestal dew]

yes, but yue is useless. It often just ignores the lyrics, spits out nonsense lyrics.

Ive tried others like diffrhythm...they garble the lyrics.

[atomic mortar]

Have you looked in comfy? Theres a few different ones but to be honest im not a big fan yet

The ones that follow lyrics sound like dog water or it sounds ok but the lyrics are horrible

[still glacier]

No ? They launched Stable Audio 2.5 like 1 month ago https://stability.ai/news/stability-ai-introduces-stable-audio-25-the-first-audio-model-built-for-enterprise-sound-production-at-scale

[viral kiln]

this is not open sourced.

[still glacier]

True, my bad.

[fleet moon]

Since comfyUI separated the frontend and the backend you need to upgrade front end also.

[rich galleon]

Guys, is there a tool like blender for 3d images and videos generation?

[proper tartan]

A website that hosts SDXL Turbo models and other things are being removed.
Is Stable Diffusion is making these models and any variations or LoRAs for these models ILEGAL on the whole internet
or is it just due to the specific websites income and what it offers as services?

Also, if someone wants to host a SDXL Turbo model or the others in the list,
and all they do is host model files,
do they need to get a license to host those files,
when the only purpose is to provide downloads?

I'm trying to understand:

• if Stability AI is trying to ban all these models
• if you are required to have an active license to share these model files, no exceptions
• or if it is simply the certain website that has to remove the models/loras/etc due to stuff like being "commercial" or "using for profit" or whatever

oh i see FAQ at the bottom

It sounds like the removal of the model files, is specific to that one website, due to not just being a website that hosts model files and lora files.
That Stability AI has no cares about people sharing the model files and lora files for free

So, i'm 95% confident that Stability AI is not trying to shut down and ban these models and loras.

If they were trying to purge all copies of the models on the internet, I'd be very angry

[rain jacinth]

[obtuse isle]

Could someone knowledgeable about LoRA training please DM ME.

[azure chasm]

Need someone to train a character lora for chroma , paid work dm me your portfolio

[untold vault]

[plush sparrow]

I need some explanation when Im using comfyui

Which template do I use when I want to use image to video

I used one but its crashing my pc with default settings🤣

[atomic mortar]

Wan image to video but what gpu do you have?

[plush sparrow]

4080super

[atomic mortar]

You might want to use a gguf (quantized or smaller version)

[plush sparrow]

I downloaded it from the official website

Or am I doing something wrong

[atomic mortar]

The 480p model is the best one as it can also do 720p generation without issue

Hmm personally i use swarmUI so im not familiar with all the noodles

[plush sparrow]

I just need some help setting up my comfyui

I downloaded it from the website and installed the template do I need to do anything else?

[atomic mortar]

Hmm i recommend posting the workflow & logs in #🤝｜tech-support

Its 2am for me now but if its not solved when i wake up ill take a peek

[plush sparrow]

Thanks

[urban rover]

Hello all, I am building pc for image and image to video generation, will 5070ti gb is enough or should i go for 5080 16 gb(max i can afford)

[sturdy jay]

We benchmarked 10 top models, both closed and open, and the results are sobering – the best results are from o3 which has a precision of 6% and a recall of 21%. All other models score below 4% precision and 10% recall. We also look at what happens when you run a model multiple times: across 8 trials models rarely discover the same errors and generally assign a near-zero confidence to their claims. The appendix contains breakdowns by field, error type, ablations for when figures are omitted, and more.

[plush sparrow]

My shit is pending forever in Comfyui

I activated the node load image

Why does it take ages

[karmic hawk]

Are you guys getting a new audio error in Stable Audio???

[gritty lava]

oi mate

i got a question aye

how is it that i switched from cpu, to apu, but the images generated on both are absoulty the exact same image xdd, down to the pixels, of course, yes, i used the same params.txt accross both, but due to the diffrenece in floating point, and environments, how is that both cpu and apu produce the exact same image, down to the teeny tinest details

[atomic mortar]

Seed determines end result, some samplers make tiny differences based on hardware though iirc

If you used the same seed and prompt/params, you get the same image

[gritty lava]

hm, interesting

so, it also has something to do with samplers as well, and ig DPM++ 2M happens to be one that seems to create the same regardless of hardware

[atomic mortar]

Also depends on scheduler

But most people like "euler a " but thats a little little different

[gritty lava]

yas

i been using dpm bc i heard they just look better

and i can kinda confirm, tho its a bit slower, seems to make sum pretty stunning tings

[plush sparrow]

anyone got a good wan nsfw workflow

[atomic mortar]

just like any other wan workflow really, just change the prompt

and theres a ton of wan lora's nowadays to increase performance

[floral umbra]

Processed 491900 media files... Fuckin hell, my script to sort, link files to models and identify up the butt is gonna take a few weeks before all models and it's preview images/metadata text files is gonna take a hot minute month lol

[sonic ivy]

since business offers and stuff liek that are not allowed here, where are this type of stuff hapening? curious to get an idea of the pricing

[atomic mortar]

Like what kinda requests? Lora making? Workflows for businesses?

Mostly spambots doing it here tbh

[sonic ivy]

just generating images following some general guidelines, like, generate 50 picture of a shortstack goblin girl with an axe, then 50 with a sword, 50 with thios position, thjat position etc

[oblique elk]

On other discords you can find opportunities or job offer channels.

[sonic ivy]

i can do it myself tbh but i'm looking to save time and i wana check whats the pricing would be like

[atomic mortar]

Oh just image generation hmm api costs are pretty low and image generation cloud based too

Iirc seeing someone doing like 5 cents per image (quality not guaranteed, upscaled, wildcard for poses/scene)

But prices vary wildly for that kinda stuff

[sonic ivy]

i see, i'm gonna try to check some servers then

thanks guys

[atomic mortar]

Its honestly worth the effort to do it your self

Its easy to set up and learn

You can even use reference images to have exact poses

[sonic ivy]

oh yeah i know but i cant both generate work on my pc with 1 gpu

[atomic mortar]

Ah true, your using them are reference images or something?

[sonic ivy]

(i work with 3d mdoels, so i need the gpu for both

[atomic mortar]

Ahh

[oblique elk]

Some discord got a request Channel where people generate for free …

[atomic mortar]

I'm bored anyways, dm me a reference image and ill crank some out

Walking the dog first

[oblique elk]

The biggest Problem with hired / Payed Image Generation is the Quality. Easy to type a prompt and generate 50 images all with 11 Fingers missing Elements etc.

[atomic mortar]

Yup, tbh if he wants like 200 images i can do it in 10min with random poses, weapons and scenes

Image to 3d model is also possible

Needs lots of cleaning tho

[sonic ivy]

is it that commun? lol a while back i was looking for a way to automaticaly change little things in the prompt and generate things constantly, but couldnt find it. so basicaly i'm stuff with generating a bunch of image with a prompt, then switching a little thing anf gening again. cant really focus on anythnig else if you're always interupting youself with changing the prompt

[atomic mortar]

Try wildcards

[sonic ivy]

thanks, and yeah thats what i mainly do, but i've got pretty good with modeling nowaday so i can maybe a decent 3d gen looks good

[atomic mortar]

Hmm in comfyUI theres a option to generate while you type but eh

Unless you got a high end card its not worth

Or you use sd1.5 lol

[sonic ivy]

6800xt, its a bit slow

haha

i'm not even sure how to check wich SD i'm using

i know i reinstalled it like 6 months ago and it was already lmuch better

[atomic mortar]

Hmm mostly by its name but if its anime based its probably sdxl illustrious

[atomic mortar]

Turbo models can archive this too iirc
https://www.youtube.com/watch?v=0HxEPcS93KY

[sonic ivy]

@atomic mortar damn, was about to dm you and turn out i already messaged you a while back and you hezlped me lol

[atomic mortar]

I somewhat remember lmao

I was quite busy at that time

[gritty lava]

sum1 turn on heater aye

it cold in here

[gentle scaffold]

Hey guys I got a few complimentary Perplexity Pro invites (it includes access to ChatGPT 5, Gemini 2.5 Pro, Claude, Grok, and early access to all the new models for free).
If anyone actually wants to try it out, just DM me and I’ll send over the official invitation. I just don’t want the invites to go to waste since they’re limited so please don't ask if your not going to use it.

[gritty lava]

ive been asking gemini 2.5 flash for everything

and its just as good as pro

i dont see a reason why to use pro

or even their new nano banana

flash does it all anyway

least it meets wut i wan

[atomic mortar]

All llm's so far are horrible for any specialized work

[gritty lava]

maybe specialized yas

but to be fair

whod wanna trust one for that

risky

[atomic mortar]

But for like a kids homework its fine but they set themselves up for failure

[gritty lava]

i asked one one time, for a horse's internal anatomy randomly, and it seems to have put all its parts into its left leg

must have been a mutant

[atomic mortar]

ah no atonomy is a big no no

[gritty lava]

xdd

apprenlty also bananas

as ive seen on YT '

[atomic mortar]

ask gpt for the seahorse emoji and see what happens

[gritty lava]

xD

dang

u wasnt kidding

gpt flipped out xd

[bright ember]

Hi, could you help me? A new AI program called Grok has come out, it's amazing, etc., but it can't animate NSFW photos. Is there a free AI program like Grok that does it? I don't have a very powerful PC, unfortunately. 🙁

[atomic mortar]

grok also actively censors nsfw but the filters are poor

hmm

since video gen consumes a lot more then image gen i dont know any other then sites doing one free video or something

[gritty lava]

i always find it a lil funi that sum ppl are just so open about their need for nsfw, may as well have asked, "is there a free porn generator?"

xD

[atomic mortar]

well based on the requests i get 3/4th of the stuff i make is nsfw

nothing shamefull about that kinda stuff nowadays

[gritty lava]

yas i suppose thats tru

idk, its just a lil funi, so openly degenerate xdd

[bright ember]

🙁

[gritty lava]

sorreh mate, i wasnt dissing u or anyting

i just meant like in general xd

[bright ember]

Hi, How do I get Animatediff to work in Forge Neo?

[abstract quarry]

if you have good hardware, use Wan

[bright ember]

My PC takes almost 30 minutes to generate a 5 second video 🙁

[elder crow]

hello

Hey, I am fullstack and AI developer

Is there anyone looking for dev ?

[atomic mortar]

the original animeatediff hasnt gotten a update for over a year so im not sure if it would work in forge neo

theres a comfy version but remembering the body horror from animatediff im not sure youd want to use that

theres a wan 1.3b if you really want to use video gen or LTX (LTX worked on my 3070 in the past)

[meager sage]

is there any discord i can reach people who knows about controlnet or t2iadapters?

[atomic mortar]

i recommend posting your actual question instead :^)

like i KNOW about it

but what do you want to know

[meager sage]

i posted already at tech support nobody said anything

[atomic mortar]

you posted it 2 minutes ago

[meager sage]

nah i posted it 5-6 hours ago

[atomic mortar]

oh the other one, had to scroll

hmm first id never use chatgpt for any of this stuff

do you know the banodoco discord?

theres more comfy users there

[meager sage]

well this my second day with comfy and i dont know anybody to help me so best i can do is chatgpt and guide videos i guess

[meager sage]

no

[atomic mortar]

if you dont mind i can redirect you to there since the people there. its also a bit more active

[meager sage]

that would be aweswome

[bright ember]

WAN 1.3b? Can you give me the link?

[atomic mortar]

Rule 5

[orchid tundra]

Guys, is there a faceswap workflow?

[harsh vale]

Hello all I'm new here and I'm in the midst of an installation process of Forge Neo. Can anyone help me?

[warm junco]

Hey, post the question with GPU model and CMD log in #🤝｜tech-support channel

[harsh vale]

where I can find those?

[warm junco]

If you get an error just copy the cmd text of forge neo

GPU model can be found in task manager under performance, GPU

[hexed minnow]

Hi there, what's the currelty best way to use SD on Windowns + AMD, still ZLUDA and Automatic1111?

[warm junco]

Hey it depends on your AMD GPU. For the newer cards the new rocm/therock support is even better than Zluda but wip

[hexed minnow]

I'm running a RX 6700 XT

[warm junco]

Forge neo with TheRock nightly is currently the fastest for my 7900xtx

[warm junco]

Ah okay then zluda is needed until TheRock supports rdna2 gpus

[hexed minnow]

Ty 🙂 and then ZLUDA plus automatic web ui? In my command shell it get a message that ZLUDA works best with SD.next?

[warm junco]

Can't recommend sdnext. Its overcomplicated and the ui is not good.
Auto1111 or Forge with Zluda are boath fine to use

[hexed minnow]

I already have Auto1111 but I haven't used it for quite some time and it's outdated now, can I just do a gitpull again and override or better a fresh install?

[warm junco]

You can do a git pull and then delete the venv and .zluda folder for a fresh setup of those

Also you should make sure to upgrade to python 3.11.9 and hip SDK 6.2 or 6.4

[hexed minnow]

Ok, I'm still on 3.10.11, I'll do that, thank you for helping. You also helped me a few years back to install it the first time, very much appreciate it 🙂

[warm junco]

No problem 🙂

[desert stirrup]

is there any good prompt extensions for Forge Neo? the few I installed just gave the most random keywords

[atomic mortar]

I seriously thought my account got hacked but no its a few bots repeating the convo from yesterday @still glacier

@serene urchin @karmic hawk @sturdy jay @urban rover @dense gust
Bots in question

[still glacier]

I saw those in artisan and was monitoring them to see where it would go :p

Well the jigg is up

time to get the mop

[atomic mortar]

Hell yeah

Just waking up and seeing fjve bots repeating a convo i had before going to bed really messes with your head lmao

[gritty lava]

and in 2 hours u could have a 4.5 min vid or so

:3

[south seal]

Hi! I have been testing the text-to-image generation models provided by Stability AI and they work great. Now, I am looking to use an image-to-image generation model where I can input both an image and some text as a prompt to generate a new image. Is this possible with any of the models provided by Stability AI?

[still glacier]

yes. you can look for img2img kind of workflows or controlnet ones

(please dont spam every channel)

[ocean ferry]

Hi, how can I make anime AI videos? What software is the best for that?

[south seal]

Thanks. Could you please elaborate a bit more on what you mean by img2img workflows or ControlNet ones? I am still trying to understand how they work.

[proven badge]

im going to war with mojang soon over censorship and if they don't like the freedom of speech without filters 100% then i guess if i win the court case then EULA dead. Every server = free speech paradise.... Worst case: You lose but go LEGEND inspire 1M gamers to rebel.... Likely case: $1M settlement + "No more chat filters" clause.

[runic patrol]

General question. How does everyone set up their neo forge. By that i mean. Best prompts, best settings, ebst extensions, best layouts of those extensions etc.

[runic patrol]

What a bargain! :p

[still glacier]

still not cool .... spamming shady under the table job ads in every channel is not the way to go.

[atomic mortar]

Hmm there isnt a best prompt so to say but theres a recommended prompt per model

And settings

In the ui i use i have presets per model since it changes a lot per model

Like samplers, negative prompt etc

Positive? Whatever i feel like making

[floral umbra]

:page_facing_up: Detected base model from associated files: Flux.1 D
:mag: Analyzing associated files for meg-ryan...
   Analyzing: meg-ryan.metadata.json
   Analyzing: meg-ryan.metadata.json
   Analyzing: README.md
   Checked 3 associated files
   Found 65 clues, analyzing...
   :dart: Detected base model: Flux.1 D (confidence: 18.70)

Clues being amount of relevant data it has found about it. Confidence score is how sure the script is that assumed base model is to being correct.

This is gonna take a looong ass time 87000 models, and it's currently scanning everything relevant to each model marked as unknown/other, included clues in the path itself for the model's base model type lol.
A script i'm working on that eventually will intelligently scan one's collection for models, build it's own database of what is of everything it can fetch of metadata, hashes and model infos, and sort them properly by model_type/base_model.model_name.extension".

And as it uses civitai.sqlite with everything so far accumulated of model info from civitai, i'm also gonna have the script allow users to put them in sub dirs based on whether it's a person, character, clothing etc, and also to sort them by SFW status lol

[south seal]

I have been testing the Stability AI Control types specifically Style and Structure for generating a children’s book cover. I used the following prompt:
"A children's book cover illustration featuring the boy from the input image playing joyfully on a swing in a vibrant playground. The background includes green grass, colorful slides, and blue skies. Add the book title 'The Joyful Swing' in large, playful font at the top and the author name 'By A. Sharma' at the bottom. The boy should retain the same facial features, hairstyle, and outfit as in the input image. The illustration should be bright, detailed, and have a cheerful, storybook style."
While the character and the scene are generated nicely, none of the outputs include the title nor the author name appear in the image. I have tried variations and different control strengths but no success so far.

Is there any recommended way to reliably include real, readable text in the generated image?

[bright ember]

Hi, do you know how to use Sora 2 via browser? I live in Italy, and it's still not available. 🙁

[atomic mortar]

Vpn to usa

[worn turtle]

Hi i want to merge 2 voices with to create 1 existing voice, with only voice samples, does anyone have any clue where I could do that?

[cerulean folio]

hellooo all:)

[vestal dew]

I've done it with f5 tts, but it didn't come out as expected

[wide cosmos]

Hello - is this a good place for image requests?

[rich arch]

hi, Aryetis. nice to meet u, i'd like to chat with u

[frank brook]

So Im new and gonna start SD in ComfyUI soon...

But I already did mess with dalle3 firefly mj etc before

[worn turtle]

Thanks man ill try that

[worn turtle]

Oh man it's so complicated

@vestal dew Do you know anything thats more simple and quick

[summer geyser]

Hey guys

[vestal dew]

Install it in Pinokio. It works with one click to install

[worn turtle]

e2 f5 tts?

@vestal dew

[vestal dew]

Yes

you can install it with pinokio

many thiongs dont work there, but f5 one click installs

[worn turtle]

Oh the Multi-Speech

Option

[lime lily]

hey peeps what are you guys using for FLUX lora training these days? I used Fluxgym ages ago and it worked well but now when i try installing it i get a ton of issues which no one nor AI can seem to solve. Is there a better option for lora training now?

[atomic mortar]

rule 5

though not sure if that matters anymore

[stiff compass]

helps a lot in research/study with its agentic capabilities You can use this link: https://pplx.ai/ss2084

[sturdy beacon]

Hello, I am looking for help, I got Stable Diffusion on my computer, not worried about it being able to handle it. It can, but I am looking for help on how to do several things. The foundation is off of huggingface. I am not a fan of it, but it is what it is for now until I can learn how to use the UI more effectively.

I am looking to upgrade the Stable Diffusion 3.5 to Large, and import models and references to learn off of.

Where do I go to start to do this? Or can anyone help get me on the right foot?

I am a neophite in this, so my knowledge is about as much as a child's ability to comprehend quantum phyisics.

[fiery ore]

No problem, you’re off to a great start. Basically, Stable Diffusion 3.5 Large is just the higher-quality version of what you already have, and you can upgrade pretty easily depending on how you’re running things. If you’re using Hugging Face directly, you’ll grab it through the diffusers library; if you’re using a UI like ComfyUI or AUTOMATIC1111, you just download the new model file and drop it into your models folder. From there, you can start adding LoRAs, custom models, or image references to teach it styles or faces. Once I know which setup you’re using, I can show you exactly how to plug everything in and get it running smoothly.

[sturdy beacon]

I think it's Gradio?

[sturdy beacon]

ugh, no pictures here... hold on... where can I show pictures of what I am looking at?

[fiery ore]

#🏞｜general-with-images

[sturdy beacon]

Just posted something for you to see.

[icy harbor]

Hello! Where can I find documentation about clip_l, clip_g and t5

[icy harbor]

I suggest to check out comfyUI and to follow a few tutorials

[frank brook]

https://generativeai.pub/the-math-art-of-artist-0thernes-not-the-typical-96e009060bc1

So I got published. Take a read. Tell me what you think?

[velvet slate]

Hello, new here. How are you.

[serene zephyr]

Welcome 😊

[velvet slate]

Thank you.

[abstract quarry]

For newcomers I always recommend InvokeAI. It has an installer that also automatically download the models for you and it has the most intuitive UI of all tools. In contrast to other tools the UI adapts to the selected models, such that functions are automatically disabled or enabled based on the chosen model. So you don't fall into the trap that you accidentally use incompatible models and adapters.
It's not as efficient and powerful as Comfyui, but it's a good start. I know most people recommend Webui because it's the oldest tool around, but it's a mess and its forked so many times that most tutorials are outdated anyways.

[cerulean isle]

Hey guys is this discord channel the right place for questions around WAN 2.2? Merci <3

[abstract quarry]

I would ask questions about wan in the discord of the used tool. When you use Comfyui to generate Wan videos, ask in the comfyui channel

[cerulean isle]

Thank youuu

[gritty lava]

boop

hope chu all having an Ai'tastic day

[frail osprey]

Please tell me, I’m downloading a Lora model that requires Illustrious, but how can I find the right base model among hundreds of them?

[atomic mortar]

where are you getting the lora from without the model?
on civitAI theres a simple filter button

[frail osprey]

@atomic mortar
https://ibb.co/jSqC38Q
https://ibb.co/39qzfpLq

[atomic mortar]

never heard of those sites

oh screenshots

yeah

illustrious, pick something you fancy and use that

but click

models

then

filter > illustrious and checkpoints

sorty by top of month and pick a nice one

[frail osprey]

You mean that any Illustrious version will work for the Lora model?

[atomic mortar]

most of the times yes

otherwise, use the lora and see the attached images for model information

[frail osprey]

thanks

[brisk plinth]

@foggy vessel hi

I am froststack9

[brazen shard]

@foggy vessel how about?

[foggy vessel]

lol why I am getting randomly pinged here

[atomic mortar]

bots lol

[magic ridge]

Hello everyone! 👋

I’m a Senior Blockchain Developer and Full-Stack Engineer with 8+ years of experience building cutting-edge solutions that merge blockchain, artificial intelligence, and scalable web technologies. I’m passionate about creating secure, high-performance systems that deliver real business value.

✨ Why work with me?

Blockchain & Web3:
Ethereum (Solidity, Hardhat), Solana (Rust, Anchor), smart contracts, DApps, DeFi platforms, NFT systems, audits, and launchpads

AI/ML Development:
GPT models, RAG systems (FAISS, Pinecone), model fine-tuning, LangChain, PyTorch, and Hugging Face integrations

Full-Stack Engineering:
React, Next.js, Node.js, NestJS, FastAPI, TypeScript, and GraphQL APIs

Cloud & DevOps:
AWS, Docker, Kubernetes, CI/CD pipelines, and serverless architectures

I’ve successfully delivered numerous blockchain, AI, and SaaS projects — from decentralized finance platforms to intelligent chatbots and automation tools — all built with clean, maintainable code and robust architecture

[spare gyro]

Hey guys, I’m new here. I wanted to learn about running AI locally. I have set up a pc and I run inference here and there. Issue is that I don’t code, but I want to get better with ComfyUI.

I currently ssh into my server to try to run it. What are rudimentary things I need to learn to run basic workflows?

Will appreciate any help

[foggy vessel]

hiya, babs. what are your goals for what you generate? are you looking to do still images, animations?

[spare gyro]

Thanks for replying. I definitely will want to start with still images, then videos (I hope).

Eventually maybe short “complex” clips

[foggy vessel]

what GPU do you have?

[spare gyro]

I have dual 5090s

[foggy vessel]

this is going to sound silly, but.. why?

[spare gyro]

Haha when I started by speaking to folks I knew personally. What would you suggest instead?

I know it’s water under the bridge, but just for your perspective

[foggy vessel]

nah, it's irrelevant if you already have it. I just wanted to know if you already had a use case.

[spare gyro]

Just really wanted to immerse myself into the space and have adequate resources to do what I’ll need

[foggy vessel]

it requires a very specific setup to use multiple GPUs; most folks do it to offload parts of the process (clip, VAE, if you're running an LLM)

I suppose there are some use cases for video as well.

[spare gyro]

Gotcha! So in hindsight, what would you recommend? In the event I can pull off some consolidation down the line

[foggy vessel]

again, I'm not going to comment.

[spare gyro]

Haha okay

[foggy vessel]

because consolidation implies that your needs or personal situation changes.

[fiery ore]

Yeah
Something like AnimateDiff

[foggy vessel]

I can't predict that, don't know you or your circumstances.

[spare gyro]

This is true. My priority is getting beyond cosmetic knowledge outside of 3rd party tools

[foggy vessel]

so, if you're starting with still images you should start with one instance of comfy running on one GPU, while you get familiar with workflows. 5090 has plenty of vram.

[fiery ore]

Have you ever generate something on ComfyUI or you just want to get started?

[spare gyro]

I generated something via vibe coding on terminal, but it was a hallucinated image lol

[foggy vessel]

what does that mean

[fiery ore]

What model did you use?

[foggy vessel]

it's totally fine if you haven't generated using comfyui - everyone starts from the same point.

but talking about vibe coding and terminals when we're discussing comfyui is.. confusing.

[spare gyro]

Can I share a pic of my screen here or it’s not permitted

[fiery ore]

Exactly

[fiery ore]

#🏞｜general-with-images

[spare gyro]

So basically a lot of the installations I did via terminal using cursor. That’s what I mean

[foggy vessel]

looks like you ran an inference using SD 1.5. It's an older model, you generated a pretty cool image if that was the model

'cause that model is ancient at this point and not coherent/cohesive compared to newer models.

[fiery ore]

You can you epicrealism as your model for more realistic images

[spare gyro]

Yeah I think that was the model. AFK now. Yeah it’s why my noob self was wondering if there’s any foundational material to learn a bit more

[spare gyro]

I’m don’t get this

[foggy vessel]

foundational for comfy >> https://docs.comfy.org/development/core-concepts/workflow

[spare gyro]

Use epic realism*?

[foggy vessel]

it's the name of a model.

[fiery ore]

Epicurealism is a model just like SD 1.5

[spare gyro]

Okay, thank you so much

[fiery ore]

Yeah and some other models like wan

[spare gyro]

Wait, I think I downloaded a version of Wan. Will confirm tomorrow and respond. Will also go through the documentation and ensure I’m using just one GPU.

Do you guys need to do any coding?

[fiery ore]

Nope

[spare gyro]

Okay. So I have my server (Linux), but I ssh into it via a mac - hence the terminal interface prior to assigning a port.

[foggy vessel]

right, that's fine. plenty of folks do similar set ups.

[spare gyro]

Will dig into the documentation for starters. I appreciate it. And I’ll revert. Thanks a lot for the help.

[atomic mortar]

i do not recommend using a sketchy download

especcially since base comfy has most that you need already

[candid mauve]

Understood. I'm just here to share a specialized tool and learn from the community. If it's not for you, that's fine. I'll be here if anyone finds the furniture workflow useful and wants to share their thoughts.

[atomic mortar]

the workflow in its self is fine but a portable kit that runs out of the box (while comfy already does that) makes it sketchy

and the fact its a mega link instead of lets say a github or website is also 👀

yikers and its a new account aswell

[warm junco]

will remove that sketchy mega upload link.
Maybe just provide some custom nodes or workflows of your stuff for users to test and not a manipulated comfyui install that cant be trusted.

And Business advertising is not allowed see Rule 5

[dusky orbit]

hello

[atomic mortar]

Hi

[hot vortex]

Hello. I wish I saw this the first time I started with ComfyUI. Now I'm only praying the newest Krita with AI generative plugin will work on my main PC because the live generation stopped working with the new ComfyUI, I can't preview what will be generated.

[oblique elk]

Did you tried the preview setting in the manager menue? There you can choose the desired preview method for the ksampler node...

[ionic zenith]

Hola

[tropic bane]

would anyone be willing to spare a sora a code for a poor android peasant? 🙏

[primal current]

Does anyone have GPT Pro chat? I have the Plus option and can only make videos up to 10 seconds long and 720p with Sora 2.

I want to know if GPT Pro can make videos longer than that, 1080p, and with Sora 2 Pro.

I also want to know what the daily limit is for creating videos.

[hoary vortex]

Guys

Did someone failed to download dreambooth?

I was download its like tutorial but it was not appear dreambooth tab

[hot vortex]

I still remember a few things from there. I got it working on my main PC strangely, but I chose a different model. Idk if there is any good model for applying parches skin but I opted for photoshop-like techniques to make skin look pale instead and a bit hollow.

As for the python I had to redo he environment and install requirements for CUDA 11.8

[hot vortex]

You could make videos linger as long as you have enough VRAM if you used cloud solution / local with enough power

[frigid tusk]

hello

[atomic mortar]

Dreambooth is outdated and is absolutely not recommend

One trainer or koyhaSS for local training

[broken bloom]

Does this thing still exist

[atomic mortar]

What thing

Stable diffusion? Absolutely

[hoary vortex]

can you recommend me some kinds of extension nowadays

[atomic mortar]

Onetrainer or koyhaSS, its a standalone program for lora training

Onetrainer is the more user friendly one iirc

[balmy marsh]

guys one question ,asking here since other ch seem less active. Im going to train my first lora with 2 characters ( actually the same character but transformed) in one lora. I have both datasets ready with same number of images , im using the colab google form for training. How many repeats do I use ? Im expected to run the same number of steps per character as normally ? (I usually go for 500 steps so since I have 23 images that would mean 2 repeats for 460 steps each character , but 920 total) , or im expected to have the step number with 2 folders be the same as I usually do ? (so 1 repeat for 460 steps in total)

[atomic mortar]

i do recommend hopping into a lora training server (most probably for the program you use)

[honest timber]

hi everyone

im a beginner

i wanted to do something like that
https://openart.ai/workflows/grouse_artistic_16/line-art-coloring/hGB6m9XaS8T7bS7LU81U

[atomic mortar]

so your own line art to image?

[honest timber]

my friend's

my friend is an artist

makes comics

[atomic mortar]

important question:
do you have a decent/good gpu?

[honest timber]

but she is a solo artist so it takes up to a half year for one chapter

[honest timber]

no but it can handle this checkpoint

8gb vram

[atomic mortar]

also coloring/details can be inconistent

[honest timber]

i plan to train lora for that

[atomic mortar]

hmm yeah 8gb vram is decent

you basically want a illustrious model with lineart controlnet

[honest timber]

i want it to only color and not change the drawing

[atomic mortar]

though it works best on one main subject per image instead of full pages at once

[honest timber]

i tried something

[atomic mortar]

the better models are probably something you cant run sadly

[honest timber]

https://openart.ai/workflows/o00oililil/manga2animebased-on-sdxl/pg5CswTCQHPDrUyGjP8T

[honest timber]

yes

[honest timber]

i copied this exactly

can i not upload images to there?

[atomic mortar]

to openart.ai? never heard of that site

[honest timber]

haha

but results were great

can we swtich to general with images chat

[atomic mortar]

hmm i dont know what you need help with since you already got a workflow and something with great results?

[honest timber]

can you go to the other chat so i can show you

[atomic mortar]

just post in #🏞｜general-with-images

[honest timber]

ok

[sturdy beacon]

Question, how do I train my AI if I am using ComfyUI?

[thorny coral]

hey everyone
A comfy newbie here trying to learn
i saw this tweet https://x.com/ingi_erlingsson/status/1977339952460632145?t=-iijBoxET7S-KSeL8AD_XA&s=19 and i wanted to like recreat/upscale an ad
so i use vast ai , i import work flow when i hit run there is some erorr about my wanvideo encode and sampler (workflow :https://github.com/cseti007/ComfyUI-Workflows/tree/ccf656c3cfc4fc09d8dc365e32dbb25eb80c5a6c/upscaling)

which i can understand any of u can help me?

[fiery ore]

The workflow error happens because it uses custom nodes like WAN Video Encode and Sampler that aren’t installed in your ComfyUI setup. Once those missing nodes are added and ComfyUI is restarted, the workflow should run correctly.

[leaden smelt]

/generate

/ generate promt: dark contrast noir photo realism with detective and ufo

[hot vortex]

Hi guys, is there a website for the samplers to use with SDXL checkpoints_ I had to do it with trial and error and the Automatic111 website that has an overview of samplers wasnt helpful. e.g. dpmpp_2m is used for SDXL models but dpmpp_2m_sde_gpu gives bad colours (something like when you are looking at a broken TV)

[atomic mortar]

nice offer but rule 5

dramatically overpaying and you can probably do 20 outfits for the same price

[cobalt vector]

1

[primal current]

G, do you know of a video upscaler that improves the realism of people? I'm trying SeedVR, but it takes a trillion years. Also Topaz, but it smooths out the upscaling.

[normal finch]

hi;]

[hallow nacelle]

anyone got ipadapter working for wan2.2 in comfyui?

[urban gust]

Hey download, login, and ask perplexity a question with my referral h ttps://pplx.ai/smanasrine85925 thanks (remove the space beetwen the h and ttps) only on computer

[atomic mortar]

For anyone needing basic internet security tips: do not follow random instructions to download, login and enter links from random people

[hot vortex]

Guys, do you have any tips for a better looking video? like 5s clip or so. I tried the official workflow that's Image to Video in video templates and at least I got result, I'm just not too sure what else to tweak. I've got banding in my image, I think that's due to the quality of the the model (took the lowest GGUF 2-bit Wan 2.2 14B) the link for it: https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF. But I'm severely limited to 6GB VRAM for now. (GTX 1060). SHould I try lowering the resolution from:
640x640 to 400x400 or something similar?
and increase frames from 16FPS to 24FPS? (roughly 23FPS is for anime, which should be smooth enough), also I use:
sampler: euler
steps: 4
CFG: 1

[atomic mortar]

I'm not sure if you can archive better quality with the 2bit model, shrinking the size would help with speed but not quality

Could try it

Iirc wan expects an output of 81 total frames

[hot vortex]

hmm, I'll try, it's just that it takes a bit too long, images are way quicker since I went from notebook to PC, but video is a whole new category for me, so I went with the default settings to try it out and making sure it works

Are there any guides for the parameters to use in order to get a better result?

[atomic mortar]

Yes, im on a 5080 and wan image to video q4 m takes me 3 to 5 min

Quality is varied

[hot vortex]

So you use GGUF 4-bit version, right?

I won't probably fit into that, that'd be more than 6GB VRAM, but I'll gladly try it out and lower the qualtiy to something over 400 pixels

[atomic mortar]

Yes but i have 16gb of vram and using 25gb of normal ram

Again, not a big fan of the output quality so i dont bother with it

[hot vortex]

I just don't want the banding to exist xD

[atomic mortar]

Trying a 14b video with 6gb vram is ambiguous

Trail and error i suppose, if you use comfy i recommend joining the comfy discord since most folks here use forge or something else

[hot vortex]

I got used to ComfyUI. I first tried Automatic111, it worked for a while, then I encountered a problem with it and since then I never came back.

Recently I got back into AI generation since I saw a few AI videos on the internet.

[hot vortex]

guys, I have a question, in the old ComfyUI, I remember, was a node that stitched images into a video and from video to image, but I can't fidn it in the newest ComfyUI version

(I talk about old ComfyUI like an year ago)

[hidden oyster]

This is my impression of brock Obama if his car broke down: My fellow Americans does anyone have jumper cables

[meager swallow]

Hi.

[flint wharf]

how might I recolor a character (pony) without inpainting

Ive been using qdiffusion which might not have enough features

[atomic mortar]

Manually photoshopping or change the color in the prompt and hope it's mostly the same

For smaller stuff a adetailer (automatic inpainting) can do stuff like hair color or eyes

[atomic mortar]

Rule 5

[hearty tendon]

gotcha

[hot vortex]

so I found out Wan 2.1 Fun Camera takes around 18 min less time to generate, better quality, but the first 2s were static images (I'll try fun control too, just don't know the difference betweeen them)

anyway hello

I used 1.3B parameters variant

[hot vortex]

And I also used uni_pc sampler instead of euler I usually for images... I wish there was a sheet of samplers, the one on stable Diffusion website isn't up to date anymore

[jovial spruce]

hello

[still glacier]

#✍🏼｜rules-and-tos message , rule 5 please

[past kindle]

Does anyone know where I can download the sd3.5 model?

[fervent thunder]

Internet

[past kindle]

Very funny. The model's creators recently updated their policy. So I wondered where I could get a local version legally.

[atomic mortar]

civitAI hosts it too, otherwise huggingface iirc

[hot vortex]

hugging face 🤷‍♂️ just use google and you'll get it

I thought SD 3.0+ models were only premium or something like that, I first saw them on paid services

or through paid API

and SD 3.5 model seems extremely large, what image resolution was 3.5 / 3.0 model based on and what are the differences?

[atomic mortar]

Iirc 1MP

But yes its large, 3.0 is not recommended

3.5 medium is okay for what it fan do but i prefer large

[hot vortex]

I'm thinking of the off-load to RAM node, but I htink it might not be worth at all and I think ComfyUI has some kind of off-loading to RAM when I monitored my RAM and GPU VRAM usage

The video nodes look really complex, trying to figure out the best combo, not sure why some people say 4 steps is alright when it clearly isn't, 20 steps for Wan 2.1 Camera control gave me a great result, Wan 2.2 14B 2-bit (the lowest) gave me horrible banding and with 4 steps (basically the default settings and left it as it is)

[vital prism]

You're also going to need a 3 or 3.5 workflow

As well as clip_l and clip_g models

[sage wave]

hi, nice to meet you!

[jade barn]

Im looking at creating ai for onlyfans is this the best place to be for help

[fiery ore]

What kind of help do you need?

[still glacier]

no nsfw stuff here.

#✍🏼｜rules-and-tos rule 2

[plush skiff]

are there bots here in this discord server for image editing? if yes then waht are the limits? does it have option to remove watermark of stablediffusion from the image?

[naive ocean]

hey

[plush skiff]

pls halp

[still glacier]

no, to my knowledge you can t use artisan bots to do img2img / edit already existing images. (but you can edit outputs of txt2img)

[plush skiff]

oh

[still glacier]

You d have to use local tools, dream studio or third party tools

[plush skiff]

i hav potato pc :p

i only have one problem

how do i remove stablediffsuion watermark from the images in down right side

[still glacier]

not sure what watermark you re talking about.

[plush skiff]

@still glacier the watermark which comes with output image when u either generate or edit the imgaes from the stablediffsuion.com web

[still glacier]

stablediffsuion.com (or stablediffusion.com without the typo I guess) is not an official website, this website is owned by "black technology" / "darkmagic.cc" and if you ask me... They re just taking advantage of their url to hijack potential clients and confuse people....

They probably want you to give them money to unlock """"""the full potential""""" of their AI tool or whatever (including the non watermarked output)

And I would NOT recommend giving them any money as :
1/ there are actual proper non sketchy website third party websites out there
2/ we saw a lot of people coming here asking for support after being declined any kind of support through their "official" mail support
3/ their website seem to be malfunctioning, people "often" come here to complain that they can t generate stuff because of error x y z

Same for stablediffusionweb.com

(same peeps behind it)

For anything "stablediffusion" official related, start from the official website https://stability.ai/

the official web tool for generation, edit, etc can be found linked there and is dreamstudio

For third party ones that aren't sketchy, civitai.com and leonardo.ai come first to my mind.

good luck and don t give any money to crooks

[plush skiff]

so, do they dont have watermarks?

[still glacier]

None.

They both have a premium pricing model where they give you some "tokens" per day/week/whatever and if you want more you have to pay for it.

[plush skiff]

thats good, atleast i wont have to tackle this watermark issue using free tiers

[plush skiff]

still wants validation by writing xd

[hot vortex]

Guys, do you know workflows I could take an inspiration from that use text to image in sequence? Such as a human eating. I created a basic workflow for 2 images that should go in sequence, same seed and Ksampler setting for both. I'm not really versed in the ComfyUI nodes, but I use CR Text -> String Join Multi node (which has 4 string input currently for:

1. the description of the character
2. position
3. action of the character
4. composition (angles, vivid, masterpiece and whatever trigger words)
   it all goes into clip along with the LoRA stack with LoRA loader

it works but the problem is I want to do for example 5 or 6 such images and the workflow would get huge

thank you for reading👍

[primal current]

My people, how do I make the WAN output videos have the same name as my video input (load video) in ComfyUI?

[viral pagoda]

Hello everyone! I’m new here. I have a question about Stable Diffusion — how can I improve or enhance the quality of my text-to-image generations?

[atomic mortar]

Mostly by improving the prompt, using a better finetuned model etc, any specifics?

Like you want better faces, or the style not to your liking etc

[viral pagoda]

I'm using the stable diffusion xl v1 but now it's already deprecated.

In hugging face how do i used the models? such some stable diffusion image generation to integrate in my project?

[turbid sand]

hi

https://www.tiktok.com/@mrcomfort22/video/7566282769509223693

Does anyone know how to create this style of drawing? This is a moment from a movie, and the author seems to have run it through a neural network.

[safe plover]

I need help, I've made and edited my videos, i need a tool to convert them to different anime styles or other styles

[hot vortex]

It's still a good model, check some fine-tuned models for you case scenario

[viral pagoda]

where i can find some model that are already matured and free ?

[hot vortex]

Look on the internet, there is also civitai, and many other websites with models
Fine-tuned = base model (you use) + additional dataset to make it richer

What I know from my search there is currently SD1.5 (512x512), SDXL (higher res 1024x1024), Flux
Video: Wan...
Asssets generation: Hunyuan (I think, haven't checked yet),

And SD3 and SD3.5

[viral pagoda]

it is free?

have you every try some models from AWS?

[honest sun]

what Linux Distro is recommended for using SD XL ? and also an LLM for having chat buddy

[plush skiff]

guys is there any AI service that allows u to use LoRa on their website for free?

[hybrid geode]

i only have a 3050 4gb vram and 16gb ram graphics card. does anyone have any suggestions on what model might run well?

[atomic mortar]

Your asking the wrong question sorta, try looking at different UI's like koboldccp, ollama or an other service for llms and like swarm, comfy or forge for linux and see whats compatible

[atomic mortar]

Sd1.5 will definitely work well, sdxl will be slower but could work

[atomic mortar]

civitAI has a token system (buzz) and you get some free but you can make a few free sfw images

[atomic mortar]

civitAI gives some buzz to add loras to a image gen, iirc its 3 buzz per lora on a image so with sdxl its 6buzz per image

[plush skiff]

can i use that to change character expression slightly or gestures?

also does this daily 100 buzz gets reset every day or is it one time use?

[atomic mortar]

One time use and 25 free a day

[dull cipher]

Hello and namaste from India 👋

[atomic mortar]

You can save up however

[plush skiff]

how?

[atomic mortar]

By going to the generate tab and hit claim, you also get little amounts for posting the image and getting likes or liking other peoples stuff

[plush skiff]

thats cool

[atomic mortar]

I currently have 1k saved and was gonna use it to make a lora

[atomic mortar]

Also, every generation costs buzz, even minor prompt tweaks

[plush skiff]

so overall, is freemium version of civitai dope?

[atomic mortar]

I would say its one of the better ones for sure

No free nsfw gens tho (but paid is possible), and theres a buzz beggars page so if you get lucky you could get some but i wouldn't bet on it

[plush skiff]

well i just neeed to change character expressions a bit so no nsfw...

how much buzz do they take for image editing btw?

[atomic mortar]

No idea, i just generate local

I just know those rates from announcements

[plush skiff]

ok nvm guess ill check it out, thanks anyways 👋

[hot vortex]

limit around HD resolution, so you could technically run anything as long as it's not a big video models (usually 14B parameters, I used GTX 1650 4GB and GTX 1060 6GB), you won't be able to load the model unless you use GGUF variant

[hot vortex]

doesn't matter, you can also use Krita with an AI plugin, at least that's what I used for hands instead of workflows in ComfyUI

but inpaiting could work well, I'm just used to Krita a bit (for hands edit)

[plush skiff]

welp the problem is i have potato pc so idk using krita,

yeah inpainting is fine for minor adjustments

[hot vortex]

I used it on my GTX 1650, 16GB DDR4

[digital sigil]

https://youtu.be/EQy5TJxejEs

[plush skiff]

i have beast GPU: GT 710 xd

[hot vortex]

welp, that's really slow

[atomic mortar]

Rule 5, who needs the kind of developer who doesn't bother reading the rules

[plush skiff]

dude where should i navigate to save up buzz?

i cant find that generate button...

[atomic mortar]

Create, generate

The painbrush icon

Hit claim on the image generation tab

And you get 25 buzz per day, if you claim it daily

[plush skiff]

thnx

[summer geyser]

Namaste 🙏

[warm monolith]

Hello, I'm new to AI, and tried adding auto captions to images via comfyui, and I get this error when clip interrogator executes, anyone know what this could mean?: CLIP_Interrogator
<urlopen error [Errno 11001] getaddrinfo failed>

[hot vortex]

show the worflow, what are you trying to do?

[wanton pier]

Hi everyone,

I've been trying to generate stuff for couple months, mostly interested in imagetoimage and realistic pictures. I mostly have pics taken from vrchat or rpg games and input them in to my comfyui to generate realistic images of my characters. I had the most success with qwen image edit. tried flux context but i might have done wrong since its giving unstable results. Is there another model i can use or which models you guys recommend for this kind of task. or maybe im missing a crucial knowledge

[glad torrent]

Just started using ComfyUI, do you need to install xformers yourself? Is it recommended? On newer Nvidia graphic cards.

[restive bobcat]

hello

[hexed trail]

Fuck

Fuck

[warm monolith]

thanks, no longer help needed, got some sleep woke up and magically it started working...

but I get confused with comfyUI, I generated captions for each image in dataset, but how do I apply those captions for each image in training?

[hot vortex]

I only tried audio training, but I imagine image training would be far harder, I don't know the specifics.

[still glacier]

no you don't have to install xformers yourself

[grave stag]

What is it about sora 2 that has a hard time not getting rid of human ears on cat girls and making her chest and swimsuit larger every time I generate a new remix. I finally got it to fix the ears in a pose I like by styling the hair but then her chest increased 3 times and her swimsuit looks baggy sigh.

How do I stop or fix this without making it say content policy does not allow also?

[atomic mortar]

@still glacier odd behavior in chat

[honest sun]

I will be putting my system together in the next week. Where are some of the better places to learn how to write image prompts for SD? I've used Sora, Flux, MidJourney, and whatever it is that Kindroid uses; but I've not yet used SD

[atomic mortar]

Prompting per model can vary different, SD3.5 large uses very different prompting then lets say illustrious finetunes

Hmm civitAI definitely has some examples of this (image + prompt & settings)

[honest sun]

and that's the other caveat; I need to do some investigating to determine which model I want to start out with.

[honest sun]

• Just found a good source for comparing models.

[lone rampart]

https://tenor.com/view/halloween-happy-halloween-trick-of-treat-pumpkin-spooky-gif-13662735232838864690

[elder tapir]

anyone knows how to bypass chatgpt prompt for NSFW?

https://github.com/Physton/sd-webui-prompt-all-in-one

[hot vortex]

you don't

[wild forge]

Hello everyone, how can I fix this issue in comfyui?

"GrowMask
can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first."

It occurs whenver I use the "GrowMask" node.

[sick fractal]

Meh these companies today are so greedy and fvcked up, every ai service for generating images/videos/characters costs 20-40$ monlty or even that have limit like 50 images insane few years ago I used stable diffusion for 100% free and generated for free like over 50000 images

[last parcel]

How do I download an install this from github? https://github.com/Stability-AI/generative-models

[atomic mortar]

It costs energy and if you use an Api its not that expensive

Sd 3.5 large costs like 8 credits and 1 credit = $0.01

[sick fractal]

0.08$ per image so 50 000 images = 4000$.
And I do it for less than 10$ my electricity bill

[atomic mortar]

Meh, they also pay for the hardware

So take model development, hardware, and electricity costs in consideration and its suddenly not too expensive

Especially for the average joe that just wants to mess around sometimes

For power users its always cheaper to run it locally or rent a gpu

[atomic mortar]

Rule 5

[torpid sorrel]

Amazing

[atomic mortar]

Its not, its most likely a bot

Since i see these 5 a day sometimes

And what good is a dev if he doesnt even read the rules :^)

[last parcel]

How do I download and install this from github? https://github.com/Stability-AI/generative-models

[fervent thunder]

Merry Scary everyone!

[summer geyser]

Hey guys how yall doing today

[floral umbra]

Working on a script to sort my near 20TB collection of models and dedupe them

� Processing model batch 405: models 40401-40500 of 81946
📊 Found 75 potential associations for 100 models
💾 Committed 75 associations for 100 models (Total: 5248)
� Processing model batch 406: models 40501-40600 of 81946
📊 Found 101 potential associations for 100 models
💾 Committed 101 associations for 100 models (Total: 5349)
� Processing model batch 407: models 40601-40700 of 81946
📊 Found 101 potential associations for 100 models
💾 Committed 101 associations for 100 models (Total: 5450)
� Processing model batch 408: models 40701-40800 of 81946

So far i'm making sure the part where it'll scan all the random orphaned media files to link to their rightful models if i have any models for them, then once that's done, the actual move and sort can begin to also dedupe them to save a goof few TB there..

[low moon]

So things have cooled down a bit since 2023 yes? No new game changing models dropping every 2 months.

[atomic mortar]

Maybe not every 2 months but new models are being dropped constantly and video gen is making big strides

Oh and audio, llm etc

[low moon]

Yeah I noticed no one gets excited about image gen now. It's all about video!

[atomic mortar]

Well qwen edit has a really good one, iirc monthly or bi monthly release

[hot vortex]

would be nice to have some kind of a summary video

[atomic mortar]

Damn no blockchain this time? Rule 5 😴

[junior urchin]

It's easy enough to set-up automod for those types of spammy ones, due to the many star tags

[atomic mortar]

To be real though, the activity has been dropping like crazy

[junior urchin]

Indeed

[weak raptor]

Is there anything rewarding here ?

[atomic mortar]

If your beyond the "how do i generate a image" skill level
And you don't feel too interested in getting others started or sharing
Idk

[weak raptor]

I get what you mean. I’m more focused on improving my workflow and exploring advanced techniques, but I don’t mind helping when I can.

[tall wave]

hi everyone

[hot vortex]

hi

[tall wave]

im new to all ai generation stuff

i got comfyui but its so confusing and mind boggiling to use, any other ones?

i used to have stable diff 1111 a while ago it was a simple ui in my web browser, but can it run img2vid?

[hot vortex]

it's easy nce you understand basic nodes, Automatic111 is too old and idk about forge, but it uses the same things as Automatic111, I think it's almost 1:1 copy, but ComfyUI is defiinitely more versatile

[tall wave]

yeah but its hell to use

i wish i had a template ready where i could just paste my Lora, Vae, and checkpoint

[hot vortex]

and the basic workflow is prepared right from the beginning + you can choose from a side menu whatever you want (this side menu didn't exist a year ago)

[tall wave]

and boom easy as that

[hot vortex]

you might still encounter some issues, but it's mostly the requirements and what GPU you use

[elfin tapir]

just get used to comfyui, it's frequently updated and very popular and flexible

[elfin tapir]

i would assume somewhere there's a template similar to that.
A protip: There's screenshots that can be loaded as templates in comfy, took me a while to figure that out

[tall wave]

I have 2 loras, no vae and 1 checkpoint, (base model i think) how do i slap them together? does or can someone send a file of a template or a quick guide on what to do so i can just paste my safe tensors and get to work

[tall wave]

screenshot upload what?

[elfin tapir]

If you search around on the web you'll find screenshots of comfyui layouts. Some of these can be dropped into comfyui directly and load as a template/workflow save file, no need to manually wire them based on a picture

[tall wave]

Im on the getting started tab on the templates, i picked the basic image generation one, loaded my checkpoint custom base model, wrote my prompt and it spat it out good, just wanna know how to add a VAE and lora

[tall wave]

what kinda file should i look out for if im planning on dropping it in comfy

[hot vortex]

Is there anything specific I have to do in order to enable drag & drop? It didn't work for me at all, so I had to manually put the json files into user/worfklow folder in ComfyUI

[elfin tapir]

https://comfyui-wiki.com/en/workflows/lora

the images in this wiki, drag drop them into comfy or save them and drag the image into comfy work area. it should be able to load the png files into actual workflows

[elfin tapir]

idk, I think some are just images, they need to be properly exported as workflows to load, or something like that

[hot vortex]

no, I meant literally workflow json files. I tried to transfer from old ComfyUI to a newer one.

Also, is there a way to align the nodes? I haven't found any way so far, but I see workflows of others neatly aligned.

[tall wave]

after a tough 30 minutes, i give up

[mental perch]

is here anyone who is renting the GPUs? Im not asking about vast AI and etc., more for private rentals

[atomic mortar]

private as in?

renting someone elses gpu/comfy instance from a distance or

[hot vortex]

I didn't know Microsoft Store has Python. Who even uses Microsoft Store.

[atomic mortar]

For work? Me its amazing

For private? God dam i hate it

[fiery ore]

Runpod

[timber raft]

Bruh, did you got lay off?

[atomic mortar]

its a bot

[mental perch]

I just wondering what is the price for renting for example 2x3070 (or more) for 1 day/week/month and if its in interest in community or bigger corpo just took the market and ppl are renting from them

[atomic mortar]

im still gonna say renting by the hour is still gonna be cheaper since you can turn it off or run on demand

according to salad:
they rent a 3070 with high constant use a 10 cent a hr, should be around 80 a month or something

[mental perch]

Many breaks during the month they have. I understand it's becouse of actual demands

[wooden grotto]

Hey folks 👋

We at Alchemyst AI just launched something pretty exciting — we’ve been building the context layer for AI agents to finally solve the data and memory problem in AI.

And now… our Chrome Extension is live! 🎉
It lets you carry your context across GPT, Claude etc — basically all your favorite AI tools — so your research, notes, and chats stay connected wherever you go.

🎯 Try the Chrome Extension → https://dub.sh/alchemyst-chromeext

🌐 Or explore the platform → https://dub.sh/getalchemystai

Would love to hear your thoughts or feedback — we’re building the missing context layer for AI agents 🚀

[atomic mortar]

for aanyone needing basic internet security tips. do not add random extentions to your browser. expecially spammed ones like these

[tame glen]

Hey fams
Is there anyone looking for a skilled and reliable developer?
If you contact me, I will show you my project experience and work.
Thanks

[atomic mortar]

O gee another one

[wooden grotto]

Well its not a spam actually if you do feel like that very sorry ,
You can head over to AlchemystAI on linkedin [https://www.linkedin.com/company/alchemystai/] and look us up,
just trying to get through our product and trying to get some feedback on how we can improve more on this, also we did go through a lot of verifications on the chrome web store.

Not a fraud one for sure.

Shared it here as a lot of people here are tinkering with AI and i thought might be a good place to get some people to test it out , and give some reviews as to weather they did find it useful or not.

[atomic mortar]

well if you see the amount of blockchain/code/developer posts get posted here youd understand the reaction

regardless rule 5

[wooden grotto]

I'm new here , so yeah i guess don't know much about that.

Still sorry if i did break a rule.

[jolly mason]

THAT WAS SO LABUBU...

[solid ferry]

its a bot

[wicked rampart]

一只猫

[worn spear]

Has anyone here joined the Supercent AI Challenge?
I participated using Stable Diffusion

[torpid radish]

never heard of it?

[proven badge]

how do you get a job at censoring people for big tech platforms so i can hmm censor all there user base XD the biggest trolling ever more power then being a discord admin for a poor server XD i would be able to shutdown all discord servers hahahaha

[atomic mortar]

Well its not an easy job and your actions would be monitored so big chance youd be fired in no time

Oh no, do not link your discord account to random sites peeps

[still glacier]

please. rule 5

[hot vortex]

Feels like rile no. 5 is like...

Moths flying into fire

[atomic mortar]

I mean thats bots

Cant we add "Full stack developer" and "blockchain" to the automod or something

[junior urchin]

Vercel apps are the worst links to click

[still glacier]

I m gonna keep nuking every copy pasta "I m a dev full stack nfp/ai/crypto" portfolio. Especially when those come with suspicious links. Same goes for "Hey we re hiring, please contact us, totally not a job scam, please join us in [insert country with huge criminal organisation presence]". It is getting tiresome to see them poping up daily. Therefore I admit my "delete message's/ban user" trigger is getting more and more sensitive. If anyone feels like they ve been legitimately wronged, please contact me or someone else from the mod team in DM.

For the rest, it is best for people to reach and ask mods before posting. Or at the very least use appropriate channels such as #🌶｜off-topic or #1092446741984444416 .

Overall I'm not in charge of this discord's organisation so I cannot promise that any channel such as #Self-Advertisement will or will not be added in the future. Personally I would tend to be against it as it would allow malicious third party to easily pray on others through scams/malware. And it would take a lot of times for us to check everything. Not to mention (not a lawyer but) some people could try to claim liability in case something does go through the filter.

Also it goes without saying but please keep in mind that moderating this server is not my job. I'm not part of stability.ai crew.

[atomic mortar]

Proof its an actual bot lmfao @still glacier, is it possible to put in a yell for some automated moderation, since just two hopped in since you posted it

[still glacier]

Yeah it s getting worse, we tamed most of the mr beast scams. So I m confident we can do something about those, we don't want to hurt the real ones tho.

[atomic mortar]

also 100% yeah, im not pointing the fingers at you since moderating is a thankless job (thanks tho)

[vocal ferry]

so ive been doing image generation for SillyTavern with a 1+ year old copy of a1111 using Pony + loooots of LORAs. what is the current best setup for NSFW generation? Flux? Chroma? something else?

[atomic mortar]

illustrious for anime

considering its still a sdxl twist so it uses less vram, good quality and less cursed prompts then pony does

[vocal ferry]

what about realistic? I use a 4060ti

[atomic mortar]

if you lack vram, illustrious again (realism tune) if you have plenty Qwen Image (20b model) or sd 3.5 (also heavy)

so if its 16gb, i recommend illustrious if you run it thru sillytavern

or a normal sdxl finetune for realism since that uses natural text more

[rugged knot]

Hello, I'm pretty new to stable-diffusion. I have a image that i have generated in 512x512 that i would like to scale up. But when i try to use the upscaler in "extras" it gets very blurry. What can i do to make it clear and "crisp"? I'm using webui forge 🙂

[warm junco]

hey, best is to load the image back into PNG-Info tab, then click on send2 txt2img.
Then regenerate it but this time with hires fix enabled.
Hires steps on 10, denois on 0.5 if its anime or semi real. 0.35 if its realism.
Upscaler: Esrgan 4x anime6b for anime

[rugged knot]

Hmm png info are not giving me the original prompt :/ so i doubt it will be able to give me the same image back

[proven badge]

LETSS GOO ALL... Figure 03 neo robot is like irobot but for 20k you can pick your own up

it will not be like irobot hopefully btw the robots but however it will control your whole life but

[atomic mortar]

For complex tasks it will be controlled by a human for "learning"

[dry sinew]

给我一只鸭

[ionic dock]

。。

[glad torrent]

Adding control net for open pose actually degrades the image quality?

[bleak robin]

Oh no, do not link your discord account to random sites peeps

[still glacier]

it shouldn't

[exotic sleet]

I want to build an Invoice Extractor, in which if I can upload a invoice pdfs(different structure,different contents), in which I have to extract specific fields into json format, get the correct json output(with accurate fields), (The issue I am facing while using OCR is , it is not detecting what's what, like what's the company name , invoice number, date , invoice table, tax summary table ,etc.) and also that I can use for free

[frozen sorrel]

Hello everyone

[hot vortex]

These services are paid in my opinion, at least from what I know from my colleague

You'd have to train a model for the specific types of invoices

[sage wave]

I can build this invoice extractor by combining PDF text parsing, OCR, and intelligent layout analysis. Using tools like pdf2image, EasyOCR, and LayoutParser, I’ll detect document structures (headers, tables, totals) and extract key fields through a mix of regex-based entity extraction and machine learning models (e.g., Donut or LayoutLM) for more complex cases. The result will be a robust, layout-agnostic system that outputs clean, structured JSON data from any invoice format—completely using open-source and free components.

[exotic sleet]

How man?

[fair ocean]

Hey everyone, what's the best text-to-image AI that runs smoothly on an RTX 4050 6GB VRAM and produces realistic, high-resolution images (1024x1024 or better) like SDXL or even higher quality? Rn i use juggernautXL_ragnarok.safetensors.

[livid knoll]

hi

[still glacier]

please, rule 5 #✍🏼｜rules-and-tos message

[atomic mortar]

with 6gb your currently running it

tbh

[fair ocean]

@atomic mortar So there is no better model I could run? I mean I could do it with offloading.

[atomic mortar]

the next level would be flux or chroma

sd 3.5l

with 16gb the normal sd 3.5l takes a minute btw

and then i refine it with sdxl

[proven badge]

eveyrone the most fire song ever... FJ OUTLAW- "Everybody's Yag" ft Forgiato Blow

[unborn hedge]

man i wish one day sora could have a local version cause its peak (when it gets it right)

im addicted to it now..

[dawn mulch]

Hot damn IP Adapter Face ID or basically insightface is not working again. insightface breaks so much. What alternative is there to insightface inside of comfy UI for stable diffusion to either detect or use face references?

[hot vortex]

It works, you must've connected wrong nodes or used a wrong model.

[dawn mulch]

Well I hope that's true then it might still work , I'll give it another try I guess. I'm thinking it has something to do with a Python library mismatch or different Python versions I'm not very knowledgeable but I see if I can make it work

[proven badge]

im building the most free speech,anti censorship website soon we will have a app... no one will be able to shutdown... don't worry all screw the government and all these's companies.. we are not going to be following any tos and we will be the most free app ever and if google takes it down then we will go after them hard for not letting anyone be truely free and if the fcc comes after us then welp i guess screw them too we will have everything from memes to culture to ai to anything i mean anything it will more then likely break a few laws but i don't care we will be so so free and no rules no tos no nothing where you can post what you want and how you like

[junior urchin]

[atomic mortar]

Sounds like a recipe for disaster, if your not based in a country that doesn't enforce those laws i think youll change your mind pretty soon

But seeing you mention the fcc, american

And based on your post history, still in k12 probably

[ivory cipher]

Hello

[tall wave]

hi

[summer geyser]

This is hella sick

[nova lily]

with A1111 wats the best way to go about image expansion? (expanding a portrait image into a landscape/wallpaper)
tyvm or do places beyond a1111 have better capability of that

[stone tinsel]

Hi team,
I'm Gagan Ryait, Partnerships Manager of BlockseBlock.
I wanted to discuss a partnership proposal regarding developer ecosystem expansion. Could you please guide me with whom can i talk?

[marsh wren]

Hey, I have this app called Image MetaHub... might be helpful. You can search by prompt, neg prompt, seed, cfg scale or filter by checkpoint, lora, sampler, etc...

https://github.com/LuqP2/Image-MetaHub if you wanna take a look... free, open-source, 100% local...
Works with A1111, Fooocus, SwarmUI, Midjourney, InvokeAI, Draw Things, Easy Diffusion, Forge, sdnext and parcially for ComfyUI (and growing).

I mean, won't help specifically with what you're doing but still :P

[honest sun]

I may get slapped on the wrist for asking; but here goes. I'm a retired photographer and I've done a lot of nude fine-art photography. I'm interested in recreating some of my images and modifying and improving them with SDXL versions. I've noticed that most of the image models I've used briefly have difficulty with female nipples. I'm looking for a LoRA, for SDXL specializing in anatomically correct female nipples. ChatGPT provided me with links to LoRAs in Civitai, but they all turn out to be broken links, or the LoRA is no longer available. Thanks for any information I receive.

[atomic mortar]

Hi, gpt is horrible for any ai related information ironically
Not home rn but its most likely a combination of checkpoint, maybe a lora or workflow

Once im home (3hrs+) i can look thru my workflow and help

Or atleast see what you need

[honest sun]

Thank you very much for taking the time. Greatly appreciated.

[floral umbra]

Oh neat!

Gave me an idea actually to see if i some day can make a hybrid mashup of both lora manager and metahub, to essentially have 2 above tabs for the gradio/webui, to where lora manager handles the loras and media sorting, metahub for managing/searching for images with metadata, as well as if i manage to, have a third thab that is for my model sorter and deduper, as it is quite a bit more thorough with seaching for file correlation than lora manager is :P

[atomic mortar]

Scams

[glad torrent]

What is a hyper Lora? Maybe give me some links to those models.

[hot vortex]

Hi. You probably forgot, but the Discord on the webpage and your current discord don't align.

as you wrote, I quote: ""
My word

It's the age of AI. Please don't fall behind the times and update all your sites."

so time to update 👍

[pseudo silo]

Does hugging face have models for stable diffusion? can I use all?

I dont see it in apps and I couldnt find a download link for a model after logging in

I filter safetensors libraries and I can use any?

[hot vortex]

you don't need an account to download the models

it's just like github repo

[pseudo silo]

is there no download for some that are web based or do I have to look for a git download

I saw files but no download, im prob blind

[hot vortex]

try civitai, it's better