💬｜general-chat

[upper plinth]

I still use A1111 for simple stuff since it is a finished product, it is glorious, reliable, and easy to use. Troubleshooting comfy can be a pain

[peak cypress]

Hi, there.
I am an AI/ML developer with over 8 years of experience solving real-world challenges across multiple industries like healthcare, law, education and so on.

I specialize in developing AI-powered solutions such as chatbots, AI agents (MCP/Agentic/Voice), Prompt engineering or implementing RAG system, and LLM models (training, deployment, and fine-tuning).

With a deep understanding of both AI/ML and web technologies, I provide end-to-end solutions from conceptualizing AI models to integrating them into practical, scalable applications.
My track record of successful projects across multiple industries ensures that I can deliver high-quality, tailored solutions that meet your specific needs.

Services I Offer:

Automation: I specialize in automating tasks using tools like n8n, Zapier, and Make.com.
NLP: I handle advanced NLP tasks with models such as GPT-4.5, GPT-4o, Claude 3-7 Sonnet, Llama-4, Gemini2.5, Mistral, and Mixtral.
Model Deployment: I assist with the seamless deployment of machine learning models across various platforms.
TTS / STT: I implement both TTS and STT solutions for interactive and conversational AI experiences.
AI Agents & Chatbots: I develop custom AI agents, Agentic AI, chatbots, and VoiceFlow applications for diverse business needs.

** Check out my portfolio in my discord profile**

I always try to learn new and cutting-edge technologies, and I place great importance on collaboration with team members in development.
If you have an innovative project idea, feel free to reach out. Let’s bring your vision to life!
Thanks.

[gritty lava]

one two, three and to the foe, snoop doggy doggy and dr dre is at the doe, ready to make and entrance so back on up, cuz you know we bouta rip shi' up

xd

[white fog]

gm

[jovial lotus]

Hello

[dusky root]

[lethal dew]

Hello

[onyx ridge]

Helloooo

[idle pilot]

hi. whats up?

[gritty lava]

[whole valve]

[latent adder]

Does anyone of you have the suno ia pro???

[visual mauve]

[loud cipher]

Gentlemans, does anyone have a 5060ti (the 16G version)? How stable diffusion goes on that card? I would like to play a littlebit more with stable diffusion but i am not player and don't want to invest that much into hardware.

[knotty rain]

How big of a file should my sdxl lora be?

[mild sedge]

yo, that 5060ti 16G should handle stable diffusion just fine for casual stuff

[loud cipher]

Thx! And what about speed? I am currently melting my poor laptop CPU and it is painfully slow. What should i expect? Approximately ofc....

[upper plinth]

Anyone know of any good sora jailbreaks pleaaaaaaase dm

[silk furnace]

hi

[tepid grove]

ryzen ai max 395 is available locally now

im trying to use wan 2.1 t2i and i2v and learning how to use comfyui in the process, how compatible should i expect things to be with that chip

i've learned zluda is a common tool for using the cuda specific stuff on amd which im not too keen about, but beyond that are there any known incompatibilities i should expect with that chip

[warm junco]

Zluda works but you should also look into TheRock. Its AMDs project on bringing native pytorch support to AMD GPUs on windows.
Its in wip but it works for ryzen ai max rn and its damn fast.

[fiery zephyr]

i am using adetailer and it keeps resizing the images now somehow and in cmd it says ADetailer: inpaint dimensions optimized -- 512x512 -> 1152x896. anyone know how to fix this?

[dawn mulch]

Has anyone used AI to create nodes and workflows?

[idle pilot]

mee, bro. why

i used cpmfyui, pykaso, and savro

[mild sedge]

hows savro? is it good in terms of generating image??

[hard condor]

Question: Does anyone know a good Discord community focused on coding LLMs?

[dawn mulch]

What have you built with it?

[valid aurora]

hello, guys, i have a question, does monitor matter at all in any way forr stable diffusion ? (would appreciate if anyone knows if it matters on blender, hunyuan, photoshop, runway too !) Resolution: FHD (1920×1080)

Refresh Rate: 180Hz

[tepid grove]

has anyone tried it and documented the compatibility, specifically for strix halo?

[nova lily]

Need advice on preserving face.

On base generation ran with Adetailer I already got a good face, when I go to upscale with USDU+Adetailer even with low denoise or empty prompt it keeps changing it too much. what could be my problem, or is this an inevitability? ty

EDIT:
I dont think its possible, best you can do is put img2img denouse at 0.5 and adetailer denoise 15, but adetailer denoise being so low you wont get any face detail improvement sadly

So I think base gen with Adetailer is an extreme double edge blade, You should do it so you have a good non-broken face to work with when you upscale but if you get an amazing face you'll probably run into the above situation. Only other solution i see is Hirez fixing during original gen.

[warm junco]

Yes here are prebuild wheels for pytorch for strix halo:
https://github.com/scottt/rocm-TheRock/releases/
And here was a comfyui test:
https://x.com/adyaman/status/1926368074866757857
And here:
https://x.com/adyaman/status/1927019695904850338

[tepid grove]

I was more looking for the wan ones

[warm junco]

It should work with wan I think. But idk how fast

[tepid grove]

Ill just wait for gb10 then ty

[celest bridge]

i am working on a face detection model that outputs a rect box of size 1080 x 1920 which contains a face from an input of a big movie/video frame. any recommendations on models and stuff?

[wicked wing]

Hey is someone having problems with ReActor in forge ui ?if i install it forge ui crashes and wont start again, couple of days ago everything worked perfectly

[tepid grove]

https://www.reddit.com/r/StableDiffusion/comments/1lrciak/is_ryzen_ai_max_395_any_good_for_stable_diffusion/ this is the only instance of someone using wan t2v on strix halo i found, and that speed seems suboptimal

i was hoping someone around here can suggest otherwise, strix halo is quite a bit cheaper

truly unfortunate

no fp8 & fp4 so guffs are gone

[warm junco]

Oh okay, yea not that great performance

[errant bronze]

Hello everyone.
I am an experienced software developer with a passion for creating visually stunning and highly functional websites and web applications.

I am used to delivering critical features on tight deadlines and solving emergencies in complex code bases.

Proficient in several technologies, [UI/UX, React, Next.js, NodeJS, NestJS, Python/Django/FastAPI, AI agent/Voiceflow, AI contents(audio, image...) creation], automation and workflow apps.

If you are gonna build website or applications, I am available to work on project and ready to discuss further.
Thanks.

[dawn mulch]

That's very kind of you to offer your service for free. But I can do all of that and more with AI - also largely or completely for free.

[tribal crown]

Is there any chance I can put lightx2v into this workflow?

I found this really good workflow on civitai ((https://civitai.com/models/1297230/wan-video-i2v-bullshit-free-upscaling-and-60-fps) that just works for me and has really good quality. Unfortunately generations time is like 4-5 minutes which is a bit long and I would like to reduce it. I found that ligthx2v lora does this for me, however the settings that in need to edit (cfg, shift, etc) i am not finding as recommended in this thread: https://www.reddit.com/r/StableDiffusion/comments/1lcz7ij/wan_14b_self_forcing_t2v_lora_by_kijai/. So basically it asks me to edit the WanVideo Sampler settings to make it work properly. I am not finding the node however in my workflow (seems like only Ksampler available).
I am using safetensors for this workflow and i don't want to mess around with any other gguf files (my internet is quite slow so I am quite tired of downloading 32gb files......).
Any help pls?

thanks!

[gritty lava]

🥜 🧈

[young orchid]

Hey, I am interested in generating sound samples from text. Any good opensource model for it.

[calm hearth]

A web dev looking for work in the Stable Diffusion discord is like a cattle rancher pitching steaks to Beyond Meat

[rare holly]

have a few questions ... if i wanna animate a image i made in forge ui i heard comfy ui can do it are there anything else i can use that is more simple to use for a simple giff or something ,, thanks

[valid aurora]

@rare holly am no expert but try runway or just google "ai video software"

its under a paywall tho i think

[rare holly]

what if i want to use it like forge and prompt what i want it to do .. or is runaway like a auto thing ?

i have seen videos on comfy ui i spoke to some people and its hard to use

[abstract quarry]

check out Wan 2.1

there are workflows in comfyui available

you should have a good gpu, though

video generation is much more demanding than single image generation

[clear steppe]

hey guys

[undone laurel]

Is there any model similiar to Vxp Illustrious ?
i got this error when using Vxp model
RuntimeError: CUDA error: CUBLAS_STATUS_ALLOC_FAILED when calling cublasCreate(handle)

[potent ocean]

Hello

[dawn mulch]

how do you deblur or de 'bokeh' something Whether the whole image or just part of the image through something like inpainting

Is there a note or model that focuses on doing that?

[opal hedge]

Stable Diffusion altered their license that you can't use their models for anime boobies

Might be the wrong place to ask

[warm junco]

come to #🤝｜tech-support

[glass gulch]

ok

[trim ledge]

I'm looking for a local LLM which can quickly and accurately convert text into json, specifically for saving sports results into a structured format

[tropic frost]

anyone here that made lora's before?

[sharp oak]

Hello everyone, check out my detailed tutorial for camera tracking in syntheyes. I am producing tutorials for everything from Modelling to houdini fx. So subscribe to my channel if you want to learn a new skill.✌️
https://youtu.be/NSJvVEJ2LHw?si=41vI3YN79O63GgEX

[cursive harbor]

Where is NSFW content posted?

[still glacier]

elsewhere

no nsfw in this discord.

[cursive harbor]

Do you have any NSFW discord with StableDiffusion/ComfyUI? If so, send it to my inbox.

[upper anchor]

hello where can i generate images? there used to be a channel here where we could generate

go to hell scammer

[near silo]

Seeing bghira/pseudo finally get what he deserves after abusing my friends and the people around me (myself Included) is so cathartic

Anyways, what are you all up to? Lol

[sweet finch]

https://civitai.com/models/153568/real-dream anyone got this model they wanna share?

[untold depot]

👉 AI Engineer | 9 Years Experience
Specialized in building, training, and deploying real-world AI systems—from autonomous agents to deep learning models.

What I Build:

• Autonomous research & data-gathering bots
• Multi-agent systems (delegation, memory, planning)
• AI assistants, IVR agents, trading bots, support agents
• End-to-end ML/DL pipelines with TensorFlow, PyTorch, Keras

Tools & Tech:

• Agent Frameworks: LangChain, LangGraph, AutoGen, CrewAI, ReAct
• Models & APIs: GPT-4o, Claude, Hugging Face, OpenAI, DeepSeek
• Stack: Python, Docker, Git, Jupyter, MLflow, Streamlit
• Domains: NLP (chatbots, classification), CV (OCR, detection)

🤝 Open to startups, AI products, or ambitious projects.
DM me if you’re hiring or building something smart.

[marble sandal]

Hello beautiful people! 👋

I’m the founder of Mirage, a new platform for AI creators to share work, sell prompts, grow in the community, and actually earn from what we love doing: AI Art.

We’re still early, and I’m inviting fellow AI artists and prompt engineers to help shape Mirage so it truly works for our community.

✅ Fill out our quick feedback form in our website and unlock early-member perks:
• 90 days Premium (free)
• 0% commission for your first month of sales
• Permanent Founder Badge
• Early access to the platform

Let’s make this our space. 🚀

[formal torrent]

HELLO

[peak cypress]

Hi, there.
I am an AI/ML developer with over 8 years of experience solving real-world challenges across multiple industries like healthcare, law, education and so on.

I specialize in developing AI-powered solutions such as chatbots, AI agents (MCP/Agentic/Voice), Prompt engineering or implementing RAG system, and LLM models (training, deployment, and fine-tuning).

With a deep understanding of both AI/ML and web technologies, I provide end-to-end solutions from conceptualizing AI models to integrating them into practical, scalable applications.
My track record of successful projects across multiple industries ensures that I can deliver high-quality, tailored solutions that meet your specific needs.

Services I Offer:

Automation: I specialize in automating tasks using tools like n8n, Zapier, and Make.com.
NLP: I handle advanced NLP tasks with models such as GPT-4.5, GPT-4o, Claude 3-7 Sonnet, Llama-4, Gemini2.5, Mistral, and Mixtral.
Model Deployment: I assist with the seamless deployment of machine learning models across various platforms.
TTS / STT: I implement both TTS and STT solutions for interactive and conversational AI experiences.
AI Agents & Chatbots: I develop custom AI agents, Agentic AI, chatbots, and VoiceFlow applications for diverse business needs.

** Check out my portfolio in my discord profile**

I always try to learn new and cutting-edge technologies, and I place great importance on collaboration with team members in development.
If you have an innovative project idea, feel free to reach out. Let’s bring your vision to life!
Thanks.

[valid trellis]

hi

[glass spruce]

You can just use LM Studio, Ollama, or GPT4All to run local LLMs like Mistral or Gemma , they're fast and good for structured tasks like converting text to JSON.
If you're okay with cloud, Grok (xAI) is also free and works well for this.

[trim ledge]

I have ollama, I mean specific models that can do it quickly and reliably (eg llama 3.1 8b cannot). I usually use claude haiku but it's a pain.

[glass spruce]

if so how about Mistral 7B instruct?

[trim ledge]

Thanks, I'll test it out

The problem is usually losing context and getting incorrect results

[hasty orbit]

Hey, any tips on keeping character context between generations and add slight varations so I can create a comic?

[carmine dawn]

So I've just installed Stable Diffusion using Forge for the first time and I'm trying to install ReActor through the Extensions window. Is it normal for it to take a long time to download and install? Would it be best to install it manually?

Maybe I just need to try install and let it run

[dawn mulch]

There's a bunch of nsfw on civtai last I checked 30 years ago or so.

[sweet finch]

anyone know any models I cna download that can accurately convert my regular prompts into SD prompts?

[hushed radish]

Greetings!

[sweet rover]

Guys
Is there a tool that helps you create an interactive slide deck?

Like timelines, you click and get to that part of the deck

[still glacier]

Oh reaallluyyyyyy

[valid aurora]

Hello, I have a very noob question. I keep hearing about SDXL, Pony, Illustrous, Flux, but I have a hard time differentiating them. From my understanding, Flux is an SDXL checkpoint, no different from ones you can find on Civitai, like Illustrious XL or Animagine XL. But why does it have its own setting in WebUI? Or am I wrong to assume Illustrious, Animagine, and Flux are in the same group (SDXL checkpoints)?

[still glacier]

pony, illustrous are based on SDXL, Flux is it's own thing.

[valid aurora]

Sorry, could you explain what exactly Flux is to me? if its not a checkpoint, I'm unsure what to classify it as

[still glacier]

it is but it's not a stable diffusion checkpoint.

it's not the same architecture under the hood.

just like sd1.5 is different from sd2, sdxl, sd3, etc

it's also not from stability.ai

[valid aurora]

ohhh ok that comparison with other sd models made it easier to understand thanks !

[abstract quarry]

Flux is a much larger model than sdxl. It's more comparable with SD 3 Large , just better

[unique ravine]

inpainting, no, try regenerating the image with bokeh in negatives

[unique ravine]

civitai discord

[unique ravine]

what gpu do u have

[dawn mulch]

I might have tried that before but I'll try again when I get the chance.

[abstract quarry]

depending on the model you use, there are often loras or slider loras that can remove the blurriness. It also helps if you add details about the background into your prompt

[feral lark]

Quick question, because i am new here, but can i download stable diff locally or I have to pay for the online version?

[warm junco]

if you have a good gpu you can sue it locally for free

[zinc island]

Hello

[dawn mulch]

You need about 6GB VRAM (gpu ram) to run it locally though lower is possible. It probably has to be an nvidia gpu also. Additionally there are a lot of online services that provide stable diffusion for free.

[zinc island]

Anyone with extensive knowledge of stable diffusion and Ia creations? I'd like help.

[zinc island]

It's not help with the software, it's something else.

[feral lark]

I meet the requerments, can you recommend me source for download or its the same from 2 years ago?

[warm junco]

First link of the pinned messages in the #🤝｜tech-support channel

[floral umbra]

Hope gpu is a good talker to win the case /s

Also hella annoying training hunyuan is. Especially since onetrainer can't do samples for videos, so no idea how long i need to run the damn thing for

[pale wedge]

hmm, Flux, Illustrious XL, and Animagine XL are all SDXL-based checkpoints, just different fine-tunes trained to excel at certain styles. Flux has its own setting in some WebUI setups is that it sometimes comes bundled with special configs, VAE, or metadata that certain frontends recognize and auto-configure

[pale wedge]

Prezi. it let you build interactive slides with clickable sections and smooth navigation

[valid aurora]

thank you 🙂

[pale wedge]

welcome! got you always! lots of people help here

[fervent thunder]

It's difficult to tell that DeviantArt has a bad AI feature and not getting things right. I mean, they might have a stable diffusion or Dall-E feature.

[abstract quarry]

Flux is not an SDXL finetune

it is its complete own thing

[dawn mulch]

civitai is the main place, there is huggingface which has more unusual models, but its hard finding models on huggingface. civitai does hold the most models and resouce, but they don't hold all such as the more unusual or so called unethical rsource that huggingface and other places might have.

Anyone remember the crazy times when crazy people were making it an issue about training AI on existing materials especially on artists and other content. Is it still a thing. Are they still trying to cause problems or have they all largely come to their senses.

It was less about, what art meant, and more about them losing potential muhnee and having to go work at mcdonalds. But they made it seems like it was about art or purpose. When it was mainly just capitalism.

Im def goona get some hate saying all that. But maybe not here.

'Real artists' would say hell yeah train AI on all my content I wouldn't mind seeing how good it is. But 'captilaists' would be screaming NO!!!

[dawn mulch]

It makes plastic or 'latex' people. Isnt that one of its main specialty. Maybe im confusing it with another model.

[abstract quarry]

what are you talking about 😂

[feral lark]

any idea where i can find NSFW female model ? 😄

[true shuttle]

Hello

[cerulean merlin]

What tools do you guys for making WAN loras?

[wanton garnet]

Hello

[languid axle]

#💬｜general-chat hello

[placid drum]

Is there a channel somewhere where you can hire people?

[woeful cliff]

建筑

[signal dome]

I forgot this server existed

[fossil smelt]

Hiiiiii, Latam?

[stuck beacon]

https://youtu.be/Yn8GlIq45H0

☝️free energy, this man needs your prayers/support/likes

[floral umbra]

Same. And judging by the spammers above, it has become a random ditch for spam lol

[fossil smelt]

Any person from Latam to talk about interesting topics?

[wanton garnet]

When using a merged checkpoint, do you just load it up like a regular model/ checkpoint? Or do you need to get the source models and merge them with the merged ceckpoint?

[signal dome]

this server is so dead now

[loud cipher]

Because everyone is talking to AI instead of real people.

[signal dome]

humanity at its peak 🤣

[loud cipher]

And everyone admiring virtual girls instead of real one.... we are going to extinct 😄

[errant bronze]

I am your partner in AI-powered business transformation. My mission is to bring innovative, AI led solutions, to your business problems, through a personalised human led approach. Delivering excellence for clients and customers with demonstrable results and measurable return on investment.

If you are looking for AI engineer, I 'd like to discuss with you.
Thanks

[violet turret]

Can stable diffusion save metadata to jpegs or only to PNG?

[fervent thunder]

only png

[loud cipher]

Can i use FLUX.1 LoRA for video generation? Sorry if this is a stupid question, i am pretty new to AI stuff....

[tropic frost]

so, i started training a lora with onetrainer. If anyone else uses it, on average, how long does it take train a character lora?

[violet turret]

That depends on a lot of factors, but basically how many steps you're doing and what GPU you're using, off the top of my head

[tropic frost]

yeah, did a bit of calculations

kinda went a bit too crazy with my settings

would've been a 9 hour training time

[violet turret]

What GPU are you using?

[tropic frost]

Radeon RX 6800

i might have also made a tad too big of a data set. I had around 90 sample images and might have set the repeats at 10

might have to look and adjust settings to something more fitting form my gpu

[violet turret]

You could always set a save point and test the LORA halfway or something before you restart

I went out and found a used 3090 for cheap to play around with.

[abstract quarry]

for videos you use Wan 2.1
There is an image2video version of Wan where you can use an e.g. flux image as starting point for a video

[loud cipher]

Mkay so i have to use this workaround to use my loras... thanks 🙂 That will be a bit extensive work.

[wind kindle]

is there any model for stable diffusion or comfy ui that like openart?

[peak cypress]

**👋 Romeo | AI/ML Developer👋 **

Hi, there. I am looking for a paid job or work as a developer with 8 years of experience in AI/ML and WEB development.

Mainly, I focus on Voice AI agent, AI-powered chatbot, Automation, Data Science, Computer Vision and Web Development.

Voice AI agent: Vapi.ai, Retell AI, Twilio, Asterisk, 11labs, etc...
AI Chatbot & NLP: RAG system, Prompt Engineering, STT/TTS, LLM models such as GPT-4.5, GPT-4o, Claude 3-7 Sonnet, Llama-4, Gemini2.5, Mistral, and Mixtral.
Automation: n8n, Zapier, and Make.com, etc...
Model Deployment: Runpod, Replicate, Huggingface, etc...
Program Language and frameworks: Python, FastAPI, Flask, Django, Node.js, React, JavaScript, TypeScripts, Express, Next.js, Nest.js, etc... (Lovable.dev)

🌐 This is my portfolio: https://romeo618.vercel.app/

In addition, I always try to learn new and cutting-edge technologies, and I place great importance on collaboration with team members in development.
If you have any idea or project, plz 📩 DM me.
Thanks

[mental forge]

Hello

[floral umbra]

Use comfyui, go into wanvideo custom node, load in wanvideo_480p_I2V_example_02.json workflow, but alter it to where instead of "load image", you add flux image gen, then it's ;last piece being vae decode, branch then it over to where wan load image weere, that way, you get a 1 click "flux to wan".

As flux and wan are fundamentally different.

I actually need to experiment as flux is image gen, but if you set wan to only generate 1 frame, but use a flux lora, if that would work Heck, try that as well I'm currently training a wan lora atm, so i can't test it.

[loud cipher]

Thanks, i will try that later 🙂

[floral umbra]

For us with folders with 100's or 1000 of images, here's gpt's why windows breaks due to fucked up thumbnail cache lol.

Ideal Feature Reality in Windows
Split thumbnail cache per drive/folder Global monolithic .db files
Auto purge based on size/age Manual cleanup only
Resilient against corruption One broken thumbnail = chaos
Async previewing for networks Often stalls whole Explorer
Indexed structure (SQLite, etc.) Custom binary format (fragile)

Makes me tempted to yet again try jumping to linux. As that way, i can set it up at least to do SQLite to do the active thumbnail caching lol

[quiet nymph]

hello 🙂

[worldly ice]

Ju

[loud cipher]

Yup, windows is and always was total crap... at least run it in container if you insist to torturing yourself with windows

[abstract quarry]

you cannot use flux loras for wan...

[floral umbra]

I do these just for kicks myself with even 5 or 10 lora strength just to see if output changes at all

I know they won't work if not trained strictly for that model :P I just like to experiment.

[abstract quarry]

they are not even applied cause the matrix names are differently

[simple dawn]

i need halp. im having tons of issues with SD WEB FORGE UI and stuff. on my 5070

[valid aurora]

Hello, guys, I have a question. Is there a way to get Clip Skip 2 in SDXL Forge without going to the 'All' UI in forge? In A1111, I just changed it in settings, but I don't think that setting exists in Forge.
The reason is, when I go to the 'All' UI and generate an image with the same settings and prompts I used to generate an image in SDXL, the results and style vary so much. So if possible, I don't wanna use the 'All' UI and I want to just change Clip Skip 2 in the XL UI without switching. From left to right, starting from the blue background image: XL UI unknown clip skip, all UI clip skip 1, all UI clip skip 2 (I've posted images on "general-with-images" any help is much appreciated !)

[south canopy]

movie with

[vocal granite]

hello guys i have a question about Kohya SS how can i get a Windows Build with start .exe in it ? I can´t find it on GitHub 😭

Thanks for answer

[abstract quarry]

it's a bunch of python scripts, there is no exe file

[desert dagger]

you might want to ask for tech support help in the #🤝｜tech-support channel

[fast sage]

For Wan2 img to vid, lets say I had a computer monitor and only wanted the screen to change, everything is static including position and camera, whats' the best prompt?

[high swan]

Hey homies, i'm currently streaming my Veo 3 / LTX session if you guys wanna join 🤘
https://www.twitch.tv/decambra89

idk iuf you guys are down with that

thanks

[valid aurora]

does anyone know how to get hands and feet to be apart of their openpose map ? right now im using openposesea website and controlnet to generate an openpose but they only include the main body frame and not the hands. the displayed image from openposesea and the exported image when i download it is different.

[crimson belfry]

We have a bot again?

When did this witchcraft happen

[peak cypress]

**👋 Romeo | AI/ML Developer👋 **

Hi, there. I am looking for a paid job or work as a developer with 8 years of experience in AI/ML and WEB development.

Mainly, I focus on Voice AI agent, AI-powered chatbot, Automation, Data Science, Computer Vision and Web Development.

Voice AI agent: Vapi.ai, Retell AI, Twilio, Asterisk, 11labs, etc...
AI Chatbot & NLP: RAG system, Prompt Engineering, STT/TTS, LLM models such as GPT-4.5, GPT-4o, Claude 3-7 Sonnet, Llama-4, Gemini2.5, Mistral, and Mixtral.
Automation: n8n, Zapier, and Make.com, etc...
Model Deployment: Runpod, Replicate, Huggingface, etc...
Program Language and frameworks: Python, FastAPI, Flask, Django, Node.js, React, JavaScript, TypeScripts, Express, Next.js, Nest.js, etc... (Lovable.dev)

🌐 This is my portfolio: romeo618.vercel.app

In addition, I always try to learn new and cutting-edge technologies, and I place great importance on collaboration with team members in development.
If you have any idea or project, plz 📩 DM me.
Thanks

[still glacier]

Just in case, if anyone clicked that external support discord or whatever. They re scammers. Leave it.

[worn flame]

Hi guys ,
I’ve been using ShakkerAI as an inference platform to run Automatic1111 for a few months, and everything was working perfectly. I recently got a new PC and installed Automatic1111 Web UI locally. I loaded the same model, same LoRA, and same settings that I used on ShakkerAI—but the results I’m getting locally are completely off and not at all like what I used to get.

I’m really confused about what could be causing this. Could it be something related to backend settings, optimizations, or dependencies that differ from ShakkerAI?

Would really appreciate any insights or suggestions.

[errant bronze]

Hi, all.
If you have a project in mind or if you're just exploring potential enhancements to your website or web application, I'd love to chat with you as a full stack developer.

[grizzled shell]

Hi, I'm using stable diffusion and I need some help on adding a custom upscaler

[peak tinsel]

‎hello guys
‎i am a hacking expert and i can teach hacking real hacking for you with a much lower price than the online courses, plus practice and supprt and i will guide you through all your road to ethical hacking journey thank you.

[soft pivot]

hello

[unique ravine]

Whatsup

[simple dawn]

i got my issue fixed it took a long time. but its fixed

[unique ravine]

Glad

[unique ravine]

Have u tried using the same seed tho?

[unique ravine]

What are u training

[pale wedge]

Yeah, that can definitely happen. Platforms like ShakkerAI usually have tuned backend settings, optimized VRAM handling, or even different sampler behavior that can affect outputs. Local installs might miss subtle things like precision settings or even a slight mismatch in dependency versions. Not really sure what you're after, but if for consistency on hyperrealism, go with Savro, i think it's optimized for stable output quality. at least for me.

[scenic vapor]

Hi everyone, I need your help. I want to install stable diffusion, but I get error

okey

[mild sedge]

hey guyss

[still glacier]

Was it someone telling you to go to another discord or dm ?

cause those are scammers

[scenic vapor]

I've already guessed it.

I think they don't give a shit about my wallet.

[rotund tartan]

[warm junco]

in the pinned messages of the #🤝｜tech-support channel you find the first link to the setup guides

[grizzled shell]

I got a question about sampling size

[signal star]

Why isn't Stable Diffusion running on an RTX 5060 Ti? Can anyone help me?

[still glacier]

check pinned message in #🤝｜tech-support

A111 does not work nice with RTX5000 series

[valid aurora]

@still glacier hello bro, may i ask a question about openpose ? i understand using multiple control types is the best but is it possible to generate an image that follows the pose using openpose only ? cause for me, openpose only vaguely follows my pose. i tried with depth only and it followed the pose perfectly. i was thinking maybe its not for solo use thats why its distorted but im not sure

[still glacier]

openpose works just fine for me. Maybe if you're using very intricate pose, too many characters, etc it will struggle but other than that it works fine in most cases.

can't tell what's wrong without an example

[wet grotto]

@brittle slate you had said that controlnet "anime" uspacel make the image more smooth, right? Can I assume that "anime" is this? Because I said that controlnet anime, for example, it to be used to reply the anime style, is this wrong? Anime is to smooth the drawing?

[old saddle]

looking for some experienced creator for some commissioned work 🧐

[floral umbra]

Managed to have comfyui run on the steam deck, but duer to 16GB total memory, desktop insta crashes,, i',m gonna see if i can make a node that offloads vae, clip and model itself to nvme swap to only use ram, or video memory rather to only hold the actual model data it will generate with

[weary plover]

Does Stability AI have any plans to continue working on the image generating models? SD 3.5 prompt understanding was great, but the images were less coherent compared to SDXL. So, im just wondering what their next plans might be.

[peak pulsar]

hey guys i created a local ai playground with ollama integration and baked in image generation and video generation which support stable diffusion anyone willing to test can download it from https://samosagpt.vercel.app/ or clone the repo.
kindly message me personally if ur interested

[simple dawn]

can i ask a question here? im just tryin to get some help with prompts. i know theres a channel for it.; but , it seems dead? or slow to respond to ppls questions.

[grizzled shell]

I need help, my cow keeps getting human ears but I only want cow ears

I put in (human ears:1.4) and it still shows

[foggy jetty]

i have an amd gpu (7700xt), windows 11 and 32gb system ram with a ryzen 5 7600x, how can i try stable diffusion

[warm junco]

Check the pinned messages first link of the #🤝｜tech-support channel.
There you find all the AMD Guides.
Go for Forge with ZLUDA.

[mystic ice]

Hi guys, I'm new to research in the AI field, I want to do research on something that's impactful and also help me get a job in a few labs, do u guys have any suggestions? something that's hot currently and something that most labs look for in candidates while recruiting

[weak knoll]

Hello!

[foggy jetty]

ive heard foocus is easy to use as beginner but ill try forge if you suggest it

[warm junco]

Yea would recommend forge much more than fooocus. As fooocus is out of updates and not as good optimised for AMD

[foggy jetty]

thanks! im setting it up now.

[warm junco]

If you have any questions feel free to ask me in #🤝｜tech-support

[foggy jetty]

ok thanks

[last fox]

does anyone here generate ai models, and get consistent results regarding the face?

[clever yew]

excuse me, does one of you have a way to avoid consistency problems for weapons and accessories? (like cut sheaths for example)

[formal ridge]

anyone know of a good process for transferring a character onto a sketch? trying to apply a consistent character design across several poses/framings

[errant bronze]

Hi, all.
If you have a project in mind or if you're just exploring potential enhancements to your website or web application, I'd love to chat with you as a full stack developer.

[peak tinsel]

Hello guys

[abstract quarry]

Flux Kontext is probably the easiest way to do that

there are alternatives like faceid, ipadapter and so on, but Flux Kontext is easiest and most flexible

[spare crescent]

Hihi!

Could I get some help for responses in my form for my MBA project? 🙂 Thank you very much in advance.

The project involves me gathering data regarding the '3 high' that many people face. Diabetes, Cholesterol, and Hypertension are major diseases that have many clinical trials running to further improve the science and medicine to battle them.

My survey involves the awareness of these trials on social media and public interest on clinical trials.

https://forms.office.com/Pages/ResponsePage.aspx?id=DQSIkWdsW0yxEjajBLZtrQAAAAAAAAAAAAN__s791ZFUMThGVlJOTEFVUE43RVI4T0Y1R0Y1WTg2Qy4u

[formal ridge]

wow man this is working great. Thanks so much!

[granite wedge]

Hi, I feel Comfi hard to use as I'm a visual learner. But maybe if I read it enough? I'm Trying to slowly detox from the apps that make me Barbies. I take the walk of mid journey shame, I can't make a concisten human... 🙈

[sand hawk]

Hi all, any interesting style or any new models are u working with to do good images lately? getting bored with what I have.

[reef birch]

Hello, anyone offers freelance services to build custom workflows? I'm building something similar to botika.io and I would like to hire the services of an expert to build the right workflows

[frosty tide]

Hi, how can I train my LORA and generate nsfw content without comfyui, suggest any online tool please, thanks

[lapis bear]

Hey everyone! 👋

I’m currently working on a lip-sync AI avatar project (similar to SadTalker/Wav2Lip). I’m trying to identify the best model in terms of real-time performance and video quality.

So far, I’ve explored:

• AniTalker
• FLOAT
• MuseTalk
• SadTalker
• Wav2Lip

Has anyone tested these in production or near real-time setups? Any suggestions for the most efficient model in terms of:
✅ Fast inference
✅ Good lip sync accuracy
✅ Easy Colab/Cloud setup

Would love your thoughts or personal experiences. Thanks in advance!

[open garden]

hows this server been doing

just checking up on it

been a minute since i've stopped by

[left comet]

any lora training professionals here?

I could only find conflicting knowledge about this topic everywhere:

Tagging the training set
-some people say you shouldnt mention anything from the image that you want the lora to "learn
-some people say the opposite

and what about style loras?
Lets say I want to train an arcane environment style lora, should I describe the scene including the art style itself? (e.g a "rough handpainted wooden wall, painterly stylized wood, brushed beige to brown gradient, rough wooden grain edges, in the style of (lora name)"?)

[errant bronze]

Hi, all.
If you have a project in mind or if you're just exploring potential enhancements to your website or web application, I'd love to chat with you as a full stack developer.

[true shuttle]

Where are you from buddy?

[primal kelp]

Hey everyone.

[daring wasp]

Hey Everyone

[dense kraken]

Hello everyone! Can someone please help me!? I have the original photo, how can I rotate it into different poses? Please help ! Всем привет! кто то может мне пожалуйста помочь!? у меня есть исходное фото,как я могу его повернуть в разные позы? помогите пожалуйста

[torpid sail]

Subject: Custom AI Deepfake & Voice Cloning Solutions – Done Locally (No Cloud)

Hi ,

I’m Aitotts, an AI deepfake and voice cloning specialist who builds fully local, high-quality tools (no cloud dependency) for filmmakers, content creators, and businesses.

✅ Custom AI Model Training (for unique voices/faces).
✅ One-Time Local Machine Setup (ready-to-use tools).
✅ Training for Your Team (master ComfyUI, voice cloning, etc.).

Here’s a sample of my work: https://youtube.com/@aitotts?si=ieWKMunrp6XKSQTx

Are you open to a quick call this week to discuss how we can integrate this into your workflow? Let me know a time that works.

Looking forward to it!

Best,
Aitotts
Connect with Me:

Instagram: @aiexpart.ai
Telegram: @Deeplearning211
WhatsApp: Click to Chat
WeChat ID: wxid_8zbf3pkvgymv22
Discord: discord.com/users/1104445468257296474

[short crescent]

I have a question. What could be the possible reasons for the slow loading of "manage"?

[sick aurora]

hi everyone

[idle pilot]

hellooo!

[high moon]

Good morning everyone!

I'm trying to train a LoRa on Runpod and I've had several tests with poor results. I was wondering if anyone could help me...

It's a LoRa to replicate the phenotype of an Argentinian woman. I've prepared a dataset of 202 high-quality images. Generally, for image generation, I use Juggernaut XL as base model, but I read that the base model should be SDXL Base 1.0. In my last test, I noticed that if I use the RealVisXL checkpoint, it gets a bit closer to the images in the dataset. Could it be that if I want to generate images with Juggernaut, I should train the LoRa with that base model?

I'm sharing my training configuration to see if anyone can help me.

Dataset Configuration
Number of Images: 202
Repeats: 7
Epochs: 6

Base Model Configuration
Base Model: stabilityai/stable-diffusion-xl-base-1.0

Dim: 64
Alpha: 32
Base Resolution: (1024, 1024)
Enable Buckets: ✅ Yes
Min Bucket Resolution: 256
Max Bucket Resolution: 2048
No Upscale: ✅ Yes
Bucket Steps: 64
Optimizer: 8-bit AdamW
Learning Rate: 1e-4 for everything (U-Net and Text Encoder)
Gradient Accumulation Steps: 1
Batch Size: 4

I realized that by doubling the weight lora:argenta:2, it gets a bit closer to the desired result, but it always generates a similar-looking woman

Thank you very much in advance!

[floral umbra]

For motion lora training for wan, do i just need say trigger word, and name of the dance to make it trigger as it should?

[neon void]

Hi

[true shuttle]

For years, I was the lonely student—no friends in college, no connections after graduation. As a solo developer, I built my skills in isolation, with no one to share my struggles or victories.
Then, everything changed when I met a Polish friend.
We coded together, debated ideas, and supported each other like true partners. We shared everything—even personal details and income, because trust was that strong. For the first time, I had someone who truly understood the dev life. I was happy.
But suddenly, he was gone. No explanation, no warning. Maybe an accident, maybe something else—I don’t know. Now, the silence is crushing. The collaboration, the camaraderie… all of it, gone.
I miss those days. I miss having a real friend who cared.
If you’re a developer (Polish or European, American) who values deep, genuine connections—let’s talk. I’m looking for someone who wants more than just small talk. Someone who believes in trust, teamwork, and building something meaningful together.
I don't need to be dev or business man, actually I'm looking for normal friend with normal idea who want extra income.

[spring spear]

hello there, i have a short question, does anyone knows a website or Controlnet Model that can properly extract the openpose controlnet skeleton from images (anime) im using the "waiNSFWIllustrious" model and the "Ilustrious_controlnet_union_sdxl_1,0_promax" model. the controlnet model works but gives black pose preview and applys details to the final image (like hairstyle, face from etc.) i appriciate the Help. thanks in advance

[frozen basalt]

hi ı just want to learn whats going on here and ı dont have any idea to start where is anybody can help me about this theres a lot of terms like stable diffusion, contestantly chacracter, comfyuı . . .maybe a few tutor,als to get start ?

[autumn ibex]

hello, does anyone need dev's help?
if so, plz let me know ur problem and i can help u with any kind of projects in this field
i'm a full stack developer
thx

[leaden iris]

hey everyone, if you were to print your AI-art onto real products (say a t-shirt for example), what type of art would you put on it?

I wanna know what you think would look best

[proper totem]

I want to contribute something on diffusion community, but it seems FLUX can do everything well. Does anyone have the same trouble as me?

[eternal comet]

is flux still the most accurate and realistic model ?

[true shuttle]

Me

[floral umbra]

Search up either aitrepenour or nerdyrodent on youtube, and search their channels for "comfyui" and "automatic1111, but instead of autoatic1111, google up forge-webui. It's a fork of auto, but not outdated by nearly a year, and actually fixes stuff.

Comfyui if you dont't mind dragging nodes
Forge if you just want a simple web based page to drag sliders

Benefit of nodes is that you can setup a long "factory lane" where it goes from image gen to image 2 video, then immediately to "video to sound", then upscale, only availability of tools and you imagination is the limit there.

[latent adder]

Does anyone use suno ia? If so, does anyone know how I can make a version with lyrics of an instrumental song? From an audio type cover or something because when I do this the music is different from the audio I sent and I wanted it to be the same thing just with lyrics

[mellow meteor]

[peak pulsar]

try checking this https://samosagpt.vercel.app/

[fervent thunder]

Hello

[true shuttle]

Btw samosa means what?

nd what is the difference beteween the other gpts?

Nowadays there are lods and lods of gpt versions.

[true shuttle]

halo

[lucid bobcat]

I need some help with Flux Kontext prompting (ComfyUI). I'm using the nunchaku version (haven't tried the full one). I'm stitching 2 images together. The left one shows a cheese lion and the right one a tropical bird made of fruits. I basically want to move the cheese lion into the right image of the bird. But I always get the same original merged image as output. It's not that the workflow doesn't work, if I prompt things like "flip the image" or "add a red hat" it works just fine. But for some reason I can't prompt to merge the two images into one scene. I have tried short prompts as well as lengthy AI generated ones.

[lucid bobcat]

Update: I got it to work by changing the guidance. I discovered a different issue: The model introduces alot of JPEG artifacts. Even when prompting to clean up the original image and remove any (small) artifacts, it only makes them worse and more pronounced! Looks like the model was trained on alot of low quality trash images.

[past knot]

I’ve been working as a software engineer for over 7 years, mostly focused on web and mobile. Recently, I built a time management app — so I even managed my own time pretty well! I also have experience building apps in travel, news, POS systems, and integrating with weather APIs. I'm currently looking for new opportunities. Thanks!

[thin geode]

Message me for guide mate

[dawn jewel]

Hi

[errant bronze]

I am your partner in AI-powered business transformation. My mission is to bring innovative, AI led solutions, to your business problems, through a personalised human led approach. Delivering excellence for clients and customers with demonstrable results and measurable return on investment.

If you are looking for AI engineer, I 'd like to discuss with you.
Thanks

[random musk]

Good Day 🙂 👋,
just discovered this dc today, glad to be here.
I´m using SD A111 and generate basically only with 1.5.
I was wondering, where would be a good place to post 1.5 pictures?

[true shuttle]

#🏞｜general-with-images
Here.

[random musk]

alright, thank you 😄 👍

may I ask, are there other people using 1.5 or is XL / Pony more common here?

[true shuttle]

Of course. I've seen so many of them, but can't remember exactly.

[random musk]

ah super, thank you again 🙂

[true shuttle]

Welcome.

[true shuttle]

Please post your images.
I can't wait to see that.

Thanks.

[autumn locust]

Hello, wondering if Spar3d is something I can run offline or if it needs internet connection after the installation.

• I just got the code from git and ran the gradio_app.py file until it said I need huggingface token.
• Wondering if post the token creation and few other initial steps needed, does it still require me to be connected to the internet?
• Please direct me to the right channel if this isn't the right one.

Thanks!

[tribal nebula]

Hi guys. Could you please tell me how to install a working Automatic1111 Stable Diffusion on the new generation of video cards, like the 5060 Ti?

[drowsy sky]

yall know any good models to convert the style of a photo into looking as if it was a sketch?

[formal crag]

Free TradingView Premium, Full Version (Windows & macOS):
https://www.reddit.com/r/CryptoForexSyndicate/comments/1kxeejv/

[drowsy sky]

That seems a little massive, I don't think that can fit into my VRAM capacity, anything else?

like 12GB max, I am running off of a Radeon 780M at the moment

wait on second thought, 8GB*, every Radeon iGPU before 8060S gets capped at 50% of system total capacity

wait fr? i will look into it then, thanks a lot

I am essentially looking into remaking an image into something as if it was something like either a watercolor painting or a sketch, should be good

just one image at the moment tbf but I might need to on a regular basis soon

AMD kinda hates my iGPU despite advertising AI stuff on it so I am running off of a custom build of sd.cpp's Vulkan backend 😭

its p sfw image but its of a family member so like 😬

yeah id rather prefer it stays on my puter and not go anywhere else

wait which one do I grab

I think I can only run 4_K

think I will grab the 4_K one thn'

[quasi notch]

Hey guys what model or application can do pic to video for free

[peak tinsel]

Hello Everyone , I’m an ethical hacker offering any kind of hacking related services.
‎Feel free to contact me for help regarding hacking issues
‎

[atomic ridge]

Hello!

[peak pulsar]

halo

[hollow lava]

hey guys

[supple minnow]

Hey All, just looking for some help to try to fix my logo with SD, is this where I ask my question? If not, where should I ask? I am using SDXL in Gradio, BTW.

[faint timber]

Read the server guide at the top of the channel list.

[lucid bobcat]

What guide?

[lucid bobcat]

There's many. But you need a good graphics card.

[runic wigeon]

is there some good workflows to create lettering/script?

[junior temple]

Just a quick note—if anyone ever needs help with automation (bots, scripts, alerts, etc.), I’ve got some experience with crypto-related tools and I’m happy to help out. Feel free to check my profile or reach out anytime. No pressure—just putting it out there.I’ve found it useful, but as always, I recommend testing things yourself and doing your own research before fully committing."

Totally get it ,  Not trying to cross any lines. Just wanted to share a space that’s helped me grow. No one’s selling anything—just serious traders discussing strategies and helping each other out.

[cloud vault]

yo can anyone help me make a lora model for a character?

[drowsy schooner]

Hi

[vocal coral]

Gm

[errant bronze]

Hi, all.
If you have a project in mind or if you're just exploring potential enhancements to your website or web application, I'd love to chat with you as a software developer.

[faint timber]

@runic wigeon The way I do fancy fonts is to find a close match to what I want in Photoshop or any image app. Then I kick out an image with the font in location where I plan on using it in my final. Supply that image to a canny control net. Then I prompt something appropriate, like font made of gold or made of candy cane. Whatever you want.

[thorny ingot]

cat

[random musk]

I generated a picture once that I would print on a mousepad / I could imagine it would fit pretty good

[leaden iris]

that's a cool idea actually, how did u do it?

[random musk]

it just generated within my usual generating while I generated a winter landscape with a lady in it

[leaden iris]

@warm junco check my dms

[junior onyx]

hello, can anyone experienced in using flux kontext help me please. i just have a few simple questions. i haven’t used image generators on pc in a few months but i’m wondering if flux kontext has support for using many images as input before generation with new memory optimizations (say 8-15 images). i need to merge a set of designs i have into one but i need use all of them in context in one prompt. i see online there are some spaces with support for multiple images and it seems to run quickly. would i be able to spin this up in my local rig the same way? im on rtx 3080 TI with 16gb vram & 64gb ram and 14 core i7. does swarmui or forge or comfyui support this feature with my vram and specs? how?? please anyone guide me

please dm me or tag me so i can receive notification

[gray wolf]

been out of the local ai scene for a min,what’s the current best and go to model and web ui?

[tight crater]

Hi

[sour otter]

the only way i could be able get away is by getting out the house with the dogs ooooo

[soft tulip]

Guys should I get comfy ui Portable or installer?

[split parcel]

Hello everyone

[rigid hull]

Hi I have a pc without a gpu can I use a stable diffusion in any way

[still glacier]

Yes but it will be terribly slow

[random musk]

Hi 🙂, its not SD but I know a 'good' free website thats based on SD ans can generate SFW and NSFW
I wrote a guide on how to use it, if your interested, feel free to dm me for the link and guide

[junior onyx]

TLDR; want help finding solution to multi-input image generation

hello, can anyone experienced in using flux kontext help me please.

i just have a few simple questions.

i haven’t used image generators on pc in a few months but i’m wondering if flux kontext has support for using many images as input before generation with new memory optimizations (say 8-15 images). i need to merge a set of designs i have into one but i need use all of them in context in one prompt. i see online there are some spaces with support for multiple images and it seems to run quickly. would i be able to spin this up in my local rig the same way? im on rtx 3080 TI with 16gb vram & 64gb ram and 14 core i7. does swarmui or forge or comfyui support this feature with my vram and specs? how?? please anyone guide me

i want to run something like this locally with support for many input pictures, it seems to support many but it says it’s using Kontext Max, is it only a Max feature currently? https://replicate.com/flux-kontext-apps/multi-image-list

but i don’t know exactly where to begin on comfyui, can someone please give me a basic blueprint?

please dm me or tag me so i can receive notification

[leaden iris]

hey guys, is the AI art community big on TikTok? if not, which social media do people tend to hand around in (except discord of course)

[gusty ember]

is there a channel we can find celebrity loras?

[lucid bobcat]

I highly doubt that Flux Kontext is good enough to assemble objects/details from 15 different images into one final output.

[floral umbra]

Wish me luck, having vscode copilot try to speed up comfyui, or really any python based program by as many times faster as you have cpu threads available

[floral umbra]

This shit has me so far so damn exited! If this thing works as intended, i'm gonna yeet it onto git lol.

I'm making now a "steam", but for python a.i stuffs. And not only will it hopefully accelerate startup and loading of everything hopefully as fast as your cpu/storage can muster, but it will also save so damn much space too! As it will make symlinks, in other words, one dependency shared across all programs that shares it, but those programs only gets a few bytes/kilobytes arrow pointing at the actual file's location. Thus this way, you can save 10's of GB by using this "hub".

comfyui-sonic/
├── venv/ # Main Sonic venv
├── project_envs/ # Project-specific environments
│ ├── comfyui/
│ │ ├── lib/ → symlinks to shared deps
│ │ └── specific/ → ComfyUI-only deps
│ ├── forge/
│ │ ├── lib/ → symlinks to shared deps
│ │ └── specific/ → Forge-only deps
│ └── automatic1111/
├── shared_deps/ # Shared dependency pool
│ ├── torch/
│ ├── numpy/
│ └── transformers/
└── sonic_accelerator.py

🎯 Universal launcher for all Python AI projects
💾 Massive space savings (torch alone is 2-3GB shared!)
⚡ Sonic acceleration for every project
🔗 Smart dependency management with hardlinks
🛡️ Project isolation without duplication

And yes, copilot uses emojis in it's chats and makes it seem like your daily scam

[fervent thunder]

Is it possible for people to sell AI-generated art as adoptables?

[floral umbra]

Well, it's technically not sellers property to sell to begin with, as after all, it's image data used from others's hard work and thrown into a vector blender.

[fervent thunder]

So you're saying that I could or couldn't sell them, right?

What's that?

It can be debating for someone to sell AI generated art.

[median kettle]

Hello

[fervent thunder]

Hello

[jagged ember]

Hi everyone, great to be here.

I run an architecture studio in Bali, Indonesia and I'm starting to explore Stable Diffusion for architectural visualization ... ideally with tight control over the outcome. I know this will be a process of fine-tuning and iteration, and I’m up for it.

If anyone can point me to high-quality tutorials, demos, or workflows (especially around ControlNet, white renders, or structured img2img), I’d really appreciate it.
Also, if you ever have architecture-related questions, whether it's design, planning, or development: feel free to reach out. Happy to contribute back from my side too.

[iron current]

hi guys

i kinda want to get a little deeper into local picture generation but i am not sure where to start and how to setup the AI on my pc. Also id love to here some recommendations on which model to use.
Thanks for zhe help.

[floral umbra]

step=11440

for a 1100 frame video i extracted png's from to train with, i might need closer to 20-30k steps lol

[floral umbra]

So, first question. Do you prefer a simple ui with sliders? Or are you ok with nodes?

Here is node setup image example, here is forge webui

And what gpu do you have, and how much ram? As if you got a decent bit of ram, you can offload bits over to the ram, in which frees up vram you can use for higher res images for instance.

[iron current]

I think id like to go with the complex UI, since i gotta learn it anyway...i got a 3060 and 32gigs of ddr4

[floral umbra]

I'm currently working on a insanely smart-hub for all things python which does everything for you, even uses hardlinks where many python rograms will share the same dependencies if the same version, to save 10's of GB lol. If wrong dependencies, it will fetch correct ones, and detect which gpu you have and link you gpu drivers if you're outdated, and so on. I can't code myself for shit yet, but paid hurtful money towards the damn copilot to make my "dream program" come to life xD

[floral umbra]

Gotcha. Do you want a one click easy instaler? Or learn the python command way of setting everything up?

[iron current]

For starters i think i'll go with the easy way but I'll come back and learn to set it up the right way once im more into the topic xD

[floral umbra]

Gotcha.

https://github.com/comfyanonymous/ComfyUI Scroll down roughly half way down, and you see a blue "direct link download" clicky. That's for the AIO package.

Take this one with ye as well :P A bat file i vibecoded (told a.i what i wanted, and it gave code) It launches comfyui, installs necessary pytorch if not present and auto opens web page with it's gui.

Just been a while since i used it on a fresh comfyui, so don't remember if it properly makes venv and whatnot lol. but report if it doesn't, and i'll fix it. It works perfectly fine after virtual env has been made though

https://image.duckers-web.site/LOGA8/bOLIKiYa03.bat

Click the download button top right, and it downloads the bat file.

Code is right there when you open the link, so nothing hidden xD

[iron current]

thank you man too kind

[floral umbra]

Aye I've gotten plenty of help here in the past, so returning the favors, plus i'm also in a quite neat mood too x)

Oh, take this one too and toss into "custom_nodes" and fetch it by double clicking on a empty space when comfy is open, type "force" and "force set clip device" pops up, click it, and move yellow node between "checkpoint loader" and "text encode", that way, you offload the text encoder to ram, and let gpu only handle the main model only. It only needs text encoder when it processes your text after all. And yes, i've multithreaded the node, so it'll use 100% cpu on any cpu to process the text encoder as fast as it's able to x) The more cores, the faster it processes :P (chews through my 5900x 12 core)

https://image.duckers-web.site/LOGA8/poHePitu40.py

@iron current

[iron current]

okok i'll try it xD...my poor 5600 😭

[lucid bobcat]

That's total nonsense. AI art is no less art than any other form of art. You are like a caveman complaining about brushes.

[floral umbra]

If it's nonsense, can you confirm none of the images used to train any of the image gen models used any copyrighted/imagery without asking the artist? As it pretty much just grabbed billions of images online to use for training :P So legally speaking, none of the materials used for those were anyone's to sell x)

But then again, i'm not a police, so if you want to sell your wordsmithed generations, i am not one to stop ye

[lucid bobcat]

That's totally irrelevant. Can you confirm that any of those "copyrighted" images weren't just imitating the art of other "copyrighted" images?

[floral umbra]

It's literally relevant when those images are literally what makes up each model

though, of course thrown into a vector blender to be trained on each art's shape and color. Then computer hallucinates forth a image with noise by using said snagged image's image information.

[lucid bobcat]

No it's not because it's just a neural network, similar to our brain. If an artist make a painting he/she will consciously or unconsciously do the exact same thing that an image model does.

[floral umbra]

Except that artist used their own hard work and training, finding their own style. And trained ai models are trained directly on others's work with no "touch of myself to not be a direct copuright copy" Like how you can make a game or a movie that has a copyright, you can make a similar game, but by not using any of the character names nor their design directly.
Needs to be indistinctive enough to not literally be a "asset copy".

And the difference between ai and a human brain, a human brain can think for itself, do it's own thing with what it has learned. A.I? It does the exact same thing as the image it was trained on, or video. Direct motion or shape replica, as close to it as it gets that is. Type mona lisa, it will draw mona lisa as it has been trained on that painting. Thus you can't sell it, because it's not yours to earn money on, legally speaking. Like if i order counterfeit nike shoes, they will be confiscated in the tolls for counterfeit good for instance.

[lucid bobcat]

You always have a human being with a thinking brain do the AI art, not the AI itself. It's just a tool. And whether or not a copyright has been infringed needs to be evaluated on a case by case basis. Just like with any other form of art. And styles are not "copyrightable". Also AI art creates its own new styles that didn't exist before.

[floral umbra]

A,i itself is a computer. It won't do squat without a human telling it what to do

And whether or not a copyright has been infringed needs to be evaluated on a case by case basis
Very much true. Cause A.I is still a very grey area. Neither legal nor illegal. But it doens't change the fact that the models has been trained on images without asking artist/photographer/person of permission. that's still a fact./

It doesn't create it's own styles. It mixes all the styles that it has been trained on from the webs :P AS otherwise, i wouldn't need to train my own loras to get the shape or outcome i desire :P If that were the case, I'd already be able to type "this celebrity stands in a mall", and it can't do that, because it hasn't been trained on that person yet. So you need to download images/videos of said celeb to make a lora to be able to noise up a image of them.

And take this book for instance. They used A.I here. And critizised a tonne for not using proper art by an artist, and instead just using A.I which is a blend of 1000's of other art to make up said image/shape

And it's why civitai got hit as they did by visa backing out of supporting them because of actual material of people and other imagery they found displeasing/disrespectful.

[lucid bobcat]

I'm not arguing with you whether or not AI slop is a bad thing. It absolutely is. I'm arguing whether or not training an AI on any art without permission is a bad thing. And I'm saying it's not, because again, style is not copyrightable and shouldn't be. Also every single artist is "guilty" of copying from others. For example manga/anime styles have become very popular in recent years. Not a single one of those artists can claim that artstyle their own. Any artist can copy any other artist. But that's not what a real artist wants to do. And I don't see this be any different with AI art. The AI artists that will make a name for themselfs will be the ones that do something unique and creative. Not the ones telling ChatGPT to write a prompt.

[floral umbra]

Aye. Style is looser on direct copyright. But models are inherently trained on everything within the image. It's artstyles it hasn't been trained on for the base model that made lora become a thing, and if people use loras that was directly trained on everything you can describe in a painting, that's where it will have data containing small bits of copyright, and where it's grey.

The people who makes mindblowing impressive A.I images are more what i noted before, wordsmiths :P They know how the models ticks, they know their vocabulary and are able to make the wildest of images.

[rancid patrol]

Hi everyone! Does anyone work with or know if stable diffusion's good for restoring old images? I have a before and after but can't post it here. Sorry if this is the wrong channel! I'm totally new to this

I've been working in restoration of pictures for quite some time, and I'm starting to increasingly use AI for some parts of the process, always keeping the fidelity of course. I recently saw an editor that achieved some impressive results in less than 10 minutes of work, and I'd love to learn how to do it!

I've been testing several online AIs, like FAL, TopazLabs and the new GPT model that respects the composition of the image a lot more than before. But none of them reach this level of detail and image fidelity as this one editor. I thought maybe a hyperrealistic model in SDXL could be the solution. What do you think?

Let me clarify: I can't currently use SDXL because I only have a GPU with 4 GB of VRAM. But if it's possible to use it for this purpose, I'm thinking about buying a new GPU with 16 GB of VRAM to be able to work on it.

You guys know more than I do, what do you think? I used SD when the 2.0 version dropped so I'm not an expert but not a total noob either!

[robust cave]

Hello
How are you doing?
I am a passionate developer, so far attended various kinds of projects.
so if you have some recommendations or looking for extra developer, I'd love to collaborate together. 😇

[peak tinsel]

Hi guys, Experience the best and safest ethical hacking and cyber security services; contact me if you need help securing/recovery your social media accounts/BTC and lost FUNDS.

[amber snow]

hello

[neon brook]

Hi

[lucid bobcat]

Flux Kontext should do the job. Stable Diffusion is useless for this as it will alter the original too much.

[high cobalt]

Any one use Stable Diffusion to segment?

[stray laurel]

Hi, can I purchase an NVIDIA RTX 5070 and use it to run ComfyUI and Flux smoothly?

[narrow imp]

hello

[gleaming bloom]

sure, why not

[gleaming bloom]

although you may want to invest in GPU with more VRAM

[stray laurel]

gpt says that flux doesn't work on 5070

[gleaming bloom]

flux-dev uses 24gb of vram
flux-schnell uses 16
you can still run quantized model on 5070 which has 12gb afaik

according to google

actually, you can even fit it into 8gb on low precision at the cost of speed and quality ofc

btw you can rent 24gb 4090 on the cloud to run comfyUI as low as $0.3/hour

pretty sure you can find 4070 on vast, simplepod or runpod to test it out

[still glacier]

Nowadays you want 16gb to run things without having to worry about VRAM. More than 16gb and you start running into enthusiasts++ territory. And if you were at that level you probably wouldn't ask this question or indeed consider renting bigger GPUs instead.

Iirc 5070 has 12 gb. You can make it work but you ll have to tread carefully if you don't want to use "medvram" options (or equivalent)

Flux is another beast. But gguf-q8 version should work on such """low VRAM""" gpus

There are many variations of it quantized differently.

[worldly marsh]

Ah yes 8 years in a product thats only been out for 2 lmao

[leaden moat]

Yo anyone need help with flux and sdxl loras?

[junior urchin]

You're gonna get banned eventually 👀

[quasi pendant]

Hello!!

[naive heart]

Guys, there is any software that ralistic simulates comfy ui, like a game, for beginners to understand how the ui works? That way it would not consume so much time and processing power to learn how to use and without extra costs, because the renting for a GPU would not be needed

[radiant tusk]

Anyone here tried using loras with the wan 2.2 model?

[oblique elk]

You could simply build and run your workflows local on your CPU (will take years to finish) and if your happy with the setup
Copy the workflow into your rental gpu

[autumn ibex]

WhoAmI?
an experienced blockchain developer
have experience working on various networks- Ethereum, Solana, Cardano, Tron, Celestia, Omni network
ensure excellent quality of all projects
tight timeline
not require so much money
professional at Solidity, Rust, Go, Move.
additional stacks : React.js, Vue.js, Next.js
Node.js, PHP
Python, C++.
Figma
LET US BUILD DECENTRALIZED WORLD TOGETHER.

[leaden moat]

Anyone need help in flux or lora training let me know

[round siren]

Has the stable diffusion install process changed for amd users recently? I saw that amd recently released rocm etc so maybe zluda isnt needed anymore to use reforge or something

[stark lark]

Hello. New here

[glad widget]

Hello, everyone. Newcomer reporting for duty.

[floral umbra]

Actually yes. As i've been training a good few wan loras now that i got the gist of how it works, but i can't seem to properly figure out how to just do motion, and not literally everything else as well.

As when i just describe "person dances the m3l0dy dance", the training also annoyingly brings the character with it. Is it the alpha that is off? Or training speed at 1-e8? And noticed now that the training that brought most of the original video including character doing the action was done at 1-e8 and not 2-e5 that the others did

Maybe that's why.

And if the dataset all is of the same motion frames, is there a choice to only use 1 dataset text file? Or does diffusion-pipe not take single-text file for dataset?

[leaden moat]

Do you only want to train on wan?

i dont have much expireince with wan lora training sorry

[floral umbra]

Nope, i want to learn to train them all. I've already somewhat nailed flux, although, i hated a update they did to ai-toolkit around a year ago iirc, as my whacky porcu hair worked perfectly, but after a few updates of ai-toolkit, it never looked the same since

And with SDXL/sd1.5, i never got those to work.

So it's more just getting the gist of how to format the dataset, text files for each image, training parameters etc, those i haven't nailed for sdxl/1.5, or even hunyuan yet. But wan and flux went nice.

Also currently having copilot make a fork of diffusion pipe to add support for optical flow to hopefully achieve better, or even only, but "perfect" motion loras :P

[surreal plover]

Hello Guys! I have a general question for someone that is relatively new to all the workflows and technical sides of image generation.
Are ai-artists like "ohneis" scam-artists? I stumbled upon his account a while ago and was greatly inspired. Diving deeper into what's needed to really create consistent and great image generation pipelines i obviously came across the technical side of it all. However, lots of ai artist claim that they can build production-chains with only chatgpt and midjourney and do all their magic only with prompting. They sell really expensive courses on those topics but i can find nothing about sdxl, comfyui, ipadapter etc.
I am geniuinely confused whats true because this completely contradicts with everything i concluded from my deepdive research.

[lucid bobcat]

If they sell courses it's a scam. There's nothing you can't find in free tutorials or with ChatGPT.

[vivid sinew]

Hello Guy, what is your go-to API aggregators for using models that you don't want to run locally?

[undone lion]

Guys is sdwebui still a thing, there's forge, reforge, classic. Idk what else. Or comfyui is the goto?

[clear pilot]

guys can someone help me setting up comfy ui?

[warm junco]

Depends on what you want to do. For normal image generation Forge, Reforge and Auto1111 work just fine

[warm junco]

In the #🤝｜tech-support channel, the first link of the pinned messages contains the setup instructions

[rancid patrol]

Thank you!!

[brisk urchin]

Why can't I use SD1.5 LoRAs in Automatic1111? It seems like they're not working properly, or at least not being respected. Which checkpoint should I use? Because I don't have any, I'm just using the pruned one..

[true shuttle]

Plese let me know which error you are experiencing.

[brisk urchin]

Oh... it's not a console error, I'm just testing a lot of LoRAs and it doesn't have the style.

@true shuttle you know what might be going on=

[warm junco]

1.5 loras only work with 1.5 based models

And then there are loras which could be broken.

Or some where you need a trigger word to see the effect

And this model should work with most 1.5 loras:
https://civitai.com/models/23900?modelVersionId=95489

[inland wedge]

I'm trying to create an image of a demon and three human people but it keeps making every look like the demon. How can I get the 3 people to not look demonic please lol

[leaden moat]

Anyone need help with lora traing let me know

[naive cipher]

Ooo don't need help with Lora training but I've been absent from the AI game for a while could you tell me if there's any new really good anime models that have released in the past few months?

[leaden moat]

Flux is best for everything almost you can use lora for style

[naive cipher]

Ye I used to make loras, been a while though

Mmmm I don't know if my laptop can handle flux though

It can handle illustrious but it's slow AF

😅

[leaden moat]

Then dont get into the game again lol

jk

[naive cipher]

Well what's a good anime flux model then

[floral umbra]

Oof, my lora sorter main script is nearing 2500 lines Poor copilot lol. (

TLDR, auto sorter and de-duplication script for loras and checkpoints, as well as images (by workflow and no workflow, and automatic1111/forge and comfyui as they use different format)

[floral umbra]

Just like i mess with my steam deck to generate images for kicks lol.

But i need to find out how i can alter rocm, or whatever files needed to alter to make my steam deck's gpu appear as a different gpu that is rocm supported xD

[floral umbra]

Any additions you guys would want for the lora and image sorter and de-duplicator to try if i still got copilot "allowance" left when all is tested and working of current core functions?

============================================================
           LoRA Management Suite
============================================================
1. Categorize and Sort LoRAs
2. Sort AI Images by Workflow/LoRA Usage
3. Deep Scan and Correlate Files
4. Generate Metadata Only (No Moving)
5. Exit
============================================================

Select an option (1-5): 1

============================================================
           LoRA Categorizer
============================================================
Enter source directory containing LoRAs (press Enter for default: .):
Enter target directory for sorted LoRAs (press Enter for default: .\test_data\loras): 

Sorting Options:
Sort by category (Character, Style, Concept, etc.)? (y/n): y
Sort by content rating (SFW/NSFW)? (y/n): y
  [FULL] Structure: BaseModel/SFW_or_NSFW/Category/ModelName/

Other Options:
Run in dry-run mode? (y/n): n
Enable deep scan for messy folders? (y/n): y

Metadata formats to generate:
1 = .metadata.json (comprehensive)
2 = .civitai.info (Civitai compatible)
3 = .md (documentation)
4 = .rgthree-info.json (RGThree nodes)
5 = .html (web-viewable)
Enter format numbers (e.g., 1,2,5) or 'all':

[pastel arch]

Hi everyone, I’m an AI developer focused on LLM workflows, agent-based tools, and MCP integration. Recently built AI sales assistants and RAG pipelines using LangChain and FastAPI. I mainly work with Python and Node.js.

I’m open to collaborations, contract work, or anything exciting in the AI space. Let’s connect!

[zealous roost]

Does anyone have a link to the unstable ai discord?

[warped sedge]

unstable diffusion?

[zealous roost]

Yes sorry I meant that one

[high peak]

hello

[flint vortex]

hello

[shut scaffold]

Hi

[wet grotto]

wich checkpooint on sd3.5 is good for atchitecture?

where is the clip and vae for sd 3.5? On github they are offline

[lucid bobcat]

No you're not an AI developer. You're a scammer that spends his day posting on Discord and other platforms.

[bright ember]

Hi guys, I'm curious. Is there a way to generate multiple images with different styles at once without having to do them one at a time? I'm especially interested in Forge.

[junior temple]

Just a quick note—if anyone ever needs help with automation (bots, scripts, alerts, etc.), I’ve got some experience with crypto-related tools and I’m happy to help out. Feel free to check my profile or reach out anytime. No pressure—just putting it out there.I’ve found it useful, but as always, I recommend testing things yourself and doing your own research before fully committing."

[brisk urchin]

Can I train LoRAs with Illustrious on OneTrainer? I don't see it among the profiles, and I don't know how to do it or if it's even possible

[unique sage]

anyone know how to make a lora using kohya_ss?

[leaden moat]

I do i can help you out if you want

[normal summit]

Anyone know if its possible to run 2 instances at the same time without needing a second comfy backend?

[leaden iris]

can anyone recommend some great models for text image editing?

The popular ones are so expensive at scale so Im looking for a cheap one or a self-host if possible

[fervent thunder]

hey all

i am trying to get a 5090 laptop for travel reasons if anyone is godo at computer stuff and knowlegable pls dm me and it would be for ai creation

[wet grotto]

you can rent a gpu service and run on any computer

[fervent thunder]

even laptop?

[wet grotto]

sure

even pc

[abstract quarry]

what do you mean? Editing images via prompt? There is Flux Kontext for that.

[wet grotto]

flux krea looks like is good to edit too

[floral umbra]

Oh boi, the lora sorter and de-dupe program i'm copilot fumbling up will be wild. Will even use LLM's to translate lora names and metadata from all [insert language] to english, or to any language really. And will also later if i got copilot budget left, have it also use a heavier llm based on their free gpu vram and initial installation translate the entire project's language to whichever language user would prefer.

[brisk urchin]

How I can train a lora qith illutrious local???

[leaden iris]

are there any that arent by black forest labs? their licencse is expensive

[robust axle]

hi

[warm granite]

I'm trying out SD for the first time and was curious. Where do people get models for their specific needs? I need one that is more like disney princesses style for a project i'm working on

[oblique elk]

the second text editing model would be highdream e1.1 (MIT as far as i know)

[warm junco]

Hey, they get then on Civitai.com

[patent ledge]

Which site do you recommend to start with?

[quiet isle]

Hello, StableDifussion Newbie here! Nice to meet you all.

[bitter orbit]

hiiiii

[terse glacier]

Looking to Hire (Paid)🚨

I’m building an AI-powered cover art generator platform.
I need a dev who can:

•Automate training LoRA models from uploaded selfies

•Integrate identity embeddings into a Stable Diffusion pipeline

•Build UI flow for upload → generation → output

DM or reply with portfolio examples or past work.

[odd hull]

hello

[stuck wing]

hello

[errant bronze]

Hi, I’m an an AI Engineer specializing in machine learning, NLP, and generative AI. I build scalable, real-world solutions that turn data into intelligent products.
Open to new opportunities, let’s build something impactful together.
Thank you

[floral umbra]

Doing a few test runs of my actual lora/model library, and even with a 5900x and seagate exos, it will take a few days lol. Processing 23338/80821And this is after 12 hours ish Currently logging all model's path and hash for test runs, and will after the has scan test the de-duplocator, to see just how many dupes i have, and how much unnecessary space they take lol.

And as it's also logging every file's path, even if i move them and delete source folder, i can make a script if user only wants to share loras made by "this artist" if the path originally was in a creator's folder, then a script can make a copy of those to a separate folder of choice in the original folder structure

[leaden moat]

Anyone need lora training let me know

[brisk urchin]

Hey, I’ve got a question I have a selfie and I wanna turn it into a 2D digital drawing using an Illustrious checkpoint and a specific LoRA… but I want it to keep the same pose as the original photo. How can I do that?

[kind whale]

Hi all, new to Discord... I’m looking for a professional LoRA trainer to create a photorealistic SDXL LoRA for my AI influencer. I have a 18-image dataset ready. I need the LoRA to lock her face/body, support NSFW, and work in AUTOMATIC1111. DMs welcome.

[leaden moat]

check dms

[leaden moat]

can do img 2 img

[brisk urchin]

Is it the same one that shows up in the Automatic1111 interface? Or do I need to do it some other way?

[leaden moat]

there are many ways to do it can look up a simple tutorial

[brisk urchin]

can anyone tell me where I can find a reference controlnet for sdxl?

[lyric turtle]

Hey is wan 2.2 allowed to be discussed here!

?*

[lyric turtle]

Cool

Anyone know how to make img to img stick closer to the image style? I send it a real image and it generates a cartoon output with a similar style to the image but very different characters almost as if it’s doing it through control net

Mmm think I may have worked it out lol

I tried to use it as text to video so bypassed the positive negative prompts from image to video node

lol ok yeah working great now 😅 thanks for my help

[brisk urchin]

anyeone know how I can make a dataset of a character with a single image? generate more consistent images of the same, if possible with automatic1111 and if not with whatever I can.

[reef plover]

I found Akool is very good at generating characters with images and videos.

[deep stratus]

Hi guys, i need some guidance with Stable diffusion, i am very new to these sort of systems. if anyone can help me please DM me.

Basically, i want it create an image and then edit that image, for example it create a pic of a lady in orange dress, but then i want it keep everything the same but change colour of the dress, or change a few things in it.

[thorn hare]

hello world

[bold obsidian]

is there a easy tutorial on how to train a model

[eternal oriole]

Hey, I'd like to try and use StableDiffusionInpainting to take a clean image and generate a new image that simulates having dirt and mud on the lense. Anyone think they know how to get this kind of thing to work?

[serene nest]

hello

[lyric turtle]

Hey are there any wan 2.2 nsfw loras yet?

[keen niche]

Yes

[lyric turtle]

@keen niche ok where do I find them please? Happy to DM if you get a moment

[lyric turtle]

?

[pine path]

@vapid dove scammer

do not click on that link, it will just tell you to install malware or something

[still glacier]

Most likely they re after people's wallet from what I ve seen.

[valid aurora]

hello guys i want to use wan 2.2 but i heard you can only use it with comfy but i dont wanna learn comfy cause it looks too hard, so my question is can i use wan 2.2 only in comfy without much complications (not interested in learning anything else) then ill just produce images in forge. like if i learn only the video generating part, will it still be hard?

[floral umbra]

There's a few on civitai.

Only a handful so far. But if you wanna make your own, diffusion-pipe supports it now

[floral umbra]

They are after the discord tokens most of the time. As if they can get the token, they can get the account and sell it.

[still glacier]

Maybe that too. But most of the ones I see active here are just trying to get you to install ransomware and scan for wallets

[kind whale]

Hi guys, i'm looking for a sdxl lora trainer please. paid job. If anyone is interested please DM me. Proof of previous high quality work required, please no time wasters

[floral umbra]

Oh, don't mind those. They are just gullible people who hasn't been taught in internet safety of current method of account phishing and got their account's token hijacked from falling for the scam.

I just ping Maxfield if they appear. And i'm a mod on another discord, and we filter 10's of those daily.

Oh, on that note, @vapid dove if you got a bot, or can get one that can filter/delete links automatically, next time you see one of those multi image spam ones, have it filter only the first link's ID itself, and it filters that entire server.

Like .gg/356i4239684/354325493,jpg or something (just spammed keyboard btw), and take that first id as the filter. That's what we do on the server i'm on.

[halcyon cliff]

👋

[floral umbra]

Hayo

[pine path]

A very good paper recently dropped. Same team behind DDT. It says it's for pixel space, but I know someone that trained a test model with the SDXL VAE
https://arxiv.org/abs/2507.23268

[thorn hare]

Looking for good documentation on running I2V on WAN 2.2. I am having some success but I think my prompting is bad and I need to add LORAs to get closer to what I'm trying to accomplish.

Anyone have and sources for me to read up? I tried youtube but every link is someone trying to get me to join their patreon.

FYI: I'm a total noob. This is day 1 for me. I am running comfyui with kaijin's workflow that I found on huggingface.

[vapid dove]

Can get a bot up and running pretty quick, I agree the weird scammers have gotten out of hand

[leaden iris]

can anyone tell me if you need a flux license if you use a flux model via Replicate API commercially?

or do Replicate just handle the license so you dont have to?

[vapid dove]

my understanding is yes, if you meet the criteria for commercialization you still need to get a commercial license from BFL

[leaden iris]

how do you know?

[vapid dove]

having read the BFL license?

but I'm not a lawyer

[faint timber]

That's the way I read it as well. The only Flux that's commercially free is Schnell.

[near silo]

Hey guys, quick question. I had my confy download a random addon once, which was basically a side area that auto-fetched all the LoRA's on my PC, and would link to them on civit, along with any info on them. Anybody know what its called? It helped me with trigger words and everything

[near silo]

https://github.com/hayden-fr/ComfyUI-Model-Manager

Found it

[hollow imp]

for a dataset of 60 character images, is 20 epochs 1ith ten repeats too much, do you think? I've never been sure how to tell if a model is overtrained.

[strong smelt]

how the fuck do i install ts

[floral umbra]

https://bfl.ai/pricing/licensing Here's their pricings for commercial use of their model.

Though, i think that's for using their model as a resource. If it's for image gen commercially, it's this section most likely

https://bfl.ai/pricing/api

[floral umbra]

What's a TS? Teamspeak?

[floral umbra]

In my experience with flux and wan, overtrained tends to overfit, and you'd end up with more source material than OG gen with bits from trained.

Sadly i've yet to understand how wan properly works, could be my training params or even my wording per text, as mine always overfits lol. As i wanted to train motion loras, but i always end up with the source content as well as when motion is right lol.

[oblique elk]

You need to be more precise. Using the output of the models (images) are covered for non-commercial and commercial purposes. The system where it runs on (api, self hosted,… ) should not matter.
If you use the model or the base weights for training, own image generators etc. you need a license. This is at least what I see from the different terms of use pages from Black Forest labs

[short thunder]

meow

[robust cave]

Hello
How are you doing?
I am a passionate developer, so far attended various kinds of projects.
so if you have some recommendations or looking for extra developer, I'd love to collaborate together. 😇

[sudden vapor]

raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

Stable diffusion model failed to load can someone help i will pay :/

[formal imp]

Heyyyy, im a new artist trying to learn new stuff, i recently discovered the world of SD so i joined the ds. Im also a gamer, i will be asking questions here probably a lot.

(Im from Argentina)

[runic patrol]

Question, if its okay to ask. What tags do you all use to stop your images from being too bright and washed out. I feel like my renders could blind people with how bright they are.

[warm junco]

Can you show an example in #📝｜prompting-help ?

[dapper jungle]

If anyone who want a beautiful anime style character plz DM me and I can design you one FOR FREE

[sterile heath]

Hey guys!

Could someone make me a video out of this picture?

oh my bad i cant send it nvm

ill send it in general with images

[undone garden]

Has there ever been a tool that allows to apply different levels of denoising strength and or controlnet strength to different areas of a picture in img2img?

[smoky elk]

hii

[grand pollen]

hello all, how has this worked out for everyone so far?

[abstract quarry]

comfyui's differential diffusion is closest to what you want

[robust cave]

Is there anyone looking for dev?

[leaden iris]

can someone recommend me a model thats good at inpainting? (via API)

[nova fable]

I have a large dataset of ~1.6 million images, many of which have watermarks that need to be removed so that I can use them as training data for an SDXL fine-tune.

I am interested in hearing about the workflows that all of you are using for large-scale batch watermark removal.

There are tools like Inpaint-Anything which can remove individual watermarks, but I have to manually locate the watermark for each image and enter the coordinates it so that I can remove it.

What I would prefer instead is give a text prompt like "watermarks, text, logos", and then have it locate/mask these objects and inpaint them out of the image automatically, instead of needing to manually specify coordinates (or click on the object myself via a GUI).

How are you all achieving this? Can some of you share code that would demonstrate clearly how to do this?

[desert dagger]

@vapid dove spamer alert

[quiet isle]

I'm been using ComfyUI and some of the workflows allow for some fine-grained control.

[quiet isle]

Depends on the tool you're using.

Hello Friends, I'm new to this world and already liking it. I managed to get ComfyUI working on an old Xeon workstation with a Quadro M5000 (yes, you're reading it right) 8GB non-RTX GPU. The great majority of my images use upscalers to account for the low VRam. Great to meet you all!

[floral umbra]

Either use different base models, or loras that targets light changes. There's loras to make brighter, and ones to make darker, almost film dark too. Or makes target the right brightness to stick out of a otherwise too bright photo

[dusty glacier]

Any prize events here?

[errant bronze]

Hi, I’m an an AI Engineer specializing in machine learning, NLP, and generative AI. I build scalable, real-world solutions that turn data into intelligent products.
Open to new opportunities, let’s build something impactful together.
Thank you

[weary mantle]

🤰🤱

[atomic lynx]

Hey ^-^
I'm not sure what is allowed/prohibited here.
I am looking for people to work at Stable Diffusion, with at least one year of experience.
The projects are interesting, the estimates are reasonable, and the team is friendly.

[dapper spoke]

Does stable diffusion use a diffusion transformer or still based on u-net architecture?

[leaden moat]

if anyone needs loraa training let me know

[plain drift]

Is seaart a good place to get loras there's some on there I want that I can't find on civit

Nvm just found out you can't download from seaart

[undone lion]

guys what sdwebui version do I use for 50 series gpu? or are there better ones?

[warm junco]

you definitly can use Auto1111, Forge or ComfyUI

[leaden iris]

what's up guys. Im building a project that integrates a bit of image/art generation

its gonna have a style selector

but i have no idea what styles to add

what are the most popular ones? what are your favorites???

NAME ANY

[shy forge]

maybe not

[lean mesa]

does anyone know which stable diffusion is the on where I can select pony, sdxl, sd 1.5 etc in one web ui? it was a fork I think I just dont remember it anymore. I could change it instnaly so I can use diffrent checkpoints that required difrent things

[still glacier]

Forge/reforge

[lean mesa]

thank you!

[abstract quarry]

stable diffusion was always based on transformer architecture regardless of unet or not. But since SD 3 it's a mmdit architecture, so not a pure diffusion transformer but a transformer architecture where text and image tokens are transformed together

[somber orchid]

Sorry if this isn't a stupid question but does wan have a discord?

[undone lion]

im getting error when installing forge classic and reforge. a1111 and forge seems outdated. comfyui seems fine

[warm junco]

you need to follow the install guides from here to get it Auto1111 and Forge working with RTX50xx series:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[runic cape]

is there a discord with a focus on Wan 2.x?

[serene sandal]

Hi I'm trying to generate Character Sheet. I'm new to doing it. Can someone guide me about how to do it? May be you can share me some resources that has worked for you.

[dull patrol]

i'm in comfyui for the first time after using Automatic1111

was using Illustrij just fine in Automatic1111, but now trying ComfyUI, illustrij isn't working. it keeps producing a black, blank image. and the checkpoint file is in the correct directory, too. any ideas?

[errant bronze]

Hi, I’m an an AI Engineer specializing in machine learning, NLP, and generative AI. I build scalable, real-world solutions that turn data into intelligent products.
Open to new opportunities, let’s build something impactful together.
Thank you

[quiet isle]

I had the same issue, I found a fix where I had to use this: https://docs.comfy.org/built-in-nodes/ClipSetLastLayer and set it to "-2".

[dull patrol]

thanks! where do i do this?

i'm only a couple days in with SD. a noob.

[quiet isle]

In Comfy, look for the "CLIP set last layer" node and put it in your workflow, the connection is as follows when doing a simple workflow: "Load Checkpoint (clip, little yellow dot)" -> Clip set last layer -> Conditioning/Prompter...

Basically, instead of going directly from the CHeckpoint loader to the prompters, you go to the CLIP Set Last Layer node first, and from there you connect its output to the CLIP input of the prompters.

Sadly I don't have my rig up so I can't share a workflow 🙁

[dull patrol]

ok looking around. i appreciate this. gonna try...

[dull patrol]

wow it was already set as you described, and just simply changing -1 to -2 fixed it!

it was -1 by default. that's amazing. what does that even do and how did you figure this out?

[leaden moat]

If anyone needs lora training let me know

[dawn mulch]

Do you know any lora or model that geneates good sketches?

what do you all think of qwen3 image

[eternal oriole]

Anyone know of a good way to inpaint dirt / mud on a camera lens?
I find that every inpainting model does not know how to add that kind of effect if I want it.

[vital charm]

Sometimes I want to prompt something a little odd, like "holding a barbeque fork" or "uncovering a Clovis point" or "walking over caltrops" and the model doesn't really know what it is, so I get a regular fork, a regular arrowhead, or sort of a mess.

If I had to guess, I'd say Pony models are the absolute worst on this (probably overtrained?)

Does anyone know which models are pretty good for finding good visual representations of odd things that are maybe a little too specific or out of the way for anyone to really have a LoRA of?

[abstract quarry]

yes, pony models are extremely dumb regarding general text understanding. But its already a limitation of CLIP, I don't think you can get around this with a CLIP-based model

[vital charm]

surely not all models are equally bad?

[abstract quarry]

use a non-CLIP model like Flux, SD3, Wan, Qwen Image

[vital charm]

Are these all compatible with Forge?

[abstract quarry]

I don't use forge, so I don't know

but Flux and SD3 are very likely supported

[vital charm]

Alright I looked into it, and I can support SD3 and Flux and maybe Wan but I probably have to swap to Comfy-UI for that, not sure. Qwen seems proprietary through their site? Not sure yet.

So now I'm looking for a good Flux model. Do you have any favorites?

[abstract quarry]

Depends on what you want. For artistic stuff I would recommend PixelWave, but the new Krea checkpoint is also quite good

Wan is a video model, but it can also do images (basically videos with a single frame).

Qwen is probably the best but it has extreme hardware requirements

all these models are open weights and can be run locally

[vital charm]

oh I didnt know they were all open
I'm giong to upgrade video card soon so maybe qwen would be nice, but where do I find it? I usually use civitai but they don't have anything for qwen

[abstract quarry]

https://huggingface.co/Qwen/Qwen-Image

but Qwen image is too large for any consumer gpu

you have to quantise it, otherwise you cannot use it

and it will be very slow

definitely the best model, but you will have to decide if you want to generate images within a few seconds (flux) or within a few minutes (qwen)

(I haven't tried Qwen myself yet, so the time estimates of several minutes is just what I read from other people)

[faint timber]

@vital charm First, update your Comfy, then visit the examples page which has the links to the models and a workflow for running qwen. It works on my 3060 12GB.
https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/

[cedar umbra]

Question regarding 'image editing models'; do you need a different model when you're doing image to image as compared to text to image? Is that what these image editing models are for? The aim would be something like making a family portrait into an anime-styled image, or a different style in general. I tried to do an image to image workflow using Qwen, but the output isn't too great.

[abstract quarry]

image editing and img2img are two different things

img2img is basically what you always do in diffusion. Therefore, you don't need an extra model. When you do txt2img then the tool you use is just doing internally an img2img on a gray image with 100% denoise

img2img with low denoise can only change small details in an image, with high denoise it will change a lot of things in the image and you need control nets to keep the composition intact

image editing is a different thing. It's more related to control nets but still different. Basically it's an extension (separate model) of basic diffusion such that your txt2img does not only get a text prompt but also an image input. edit models are trained such that they understand a lot of editing tasks and excel in only modifying parts of your image

for transforming a family portrait into anime style, edit models are the preferred way. You can alternatively also try control nets, but editing models are better

Qwen is special in this regard, because it uses a text encoder that can understand both images and text, so you can simply add images to your text prompt without the need of an extra model

if your image to image workflow with Qwen didn't worked so well then you probably used img2img instead of an image edit workflow

[unborn hedge]

Mods we got a spambot here

[still isle]

Has anyone submitted to any AI art contests?

[shy forge]

yes

[quiet isle]

Wow, thanks for the info. Are there any editing models that you could recommend, and are those able to run in ComfyUI?

[abstract quarry]

there are not many at all

• Flux Kontext
• HighDream edit
• Qwen Image

Omnigen

I only tried Flux Kontext so far

it's good, but it has the same issue as flux: it's not good with styles in general

[quiet isle]

Understood, are those available in CiviAI?

[abstract quarry]

probably. They are definitely on huggingface

huggingface is usually the #1 place to download models

there is even a ghibli style kontext lora on cititai: https://civitai.com/models/1732367/ghibli-style-flux-kontext?modelVersionId=1960649

in general Flux Kontext should be able to do most of the stuff without extra loras, but they can improve results

[quiet isle]

Nice!!!, I'm a big fan of Ghibli, tbh, the reason I started (still a very noob) on this path was because I wanted to generate my own storyboards with "original" (and I put it as such because all AI generated images are technically original) characters. Thanks for sharing the info.

I have a 12GB GPU with a 16GB on the way, a 5060ti, not the greatest, but the price couldn't be beat.

I'm hoping I can run Flux Kontext in it.

[fervent thunder]

hey heya

[lucid bobcat]

Has anyone managed to get Qwen Image running on 8GB VRAM in ComfyUI? Even the 7GB Q2 model takes 35s/it. Yet I can run the 13GB Flux Q8 at 5s/it. Something's not right.

[quiet isle]

I'm impressed that the Flux Q8 can actually run on 8GB, what's your GPU model? Does it require RTX-class tensor cores?

[abstract quarry]

quantisation reduces memory but it does not make the model faster

Qwen has many more layers than flux, therefore it takes more time

also it has a similar architecture as SD3 which is slower than the Flux architecture

[lucid bobcat]

RTX 2070

Shut up you're just a scammer and not a softwre engineer.

[lucid bobcat]

I was expecting Qwen to take roughly twice as long than Flux and 4x with CFG > 1. That would roughly be 10s with CFG = 1 but it takes 19s.

[abstract quarry]

checked the flux params again and Flux and Qwen are indeed same size

just Qwen is using the slower SD3 double block architecture

[robust cave]

Don't insult without evidence.

[desert dagger]

@vapid dove more spammers - this is how they are getting around the filters now

[lucid bobcat]

Don't tell me what to do!

[hollow lava]

Hey guys, are there any freelance digital artists who have experience working in the video game industry? I’m curious to hear your thoughts on the current AI debate.

[lucid bobcat]

Here's my thoughts as a gamer. It's great for small indie projects, especially with future AI developments. It's really bad for curation, because there will be a flood of AI slop cluttering shops.

[hollow lava]

I’ve been interviewing with some freelance visual artists working in game studios, and they say the AI situation right now is incredibly polarised. There are anti-AI vigilantes doing witch-hunts on LinkedIn, putting artists who use AI on blacklists. Just speaking about AI is already a huge taboo within the industry for a lot of game artists.

[faint timber]

Same in the world of writing. Talk about a toxic group.

[deep axle]

Anyone need an automation set up for their business.
Please contact me.
I need a job

[half raven]

where can i get support?

[desert dagger]

in the #🤝｜tech-support channel

[tacit schooner]

hey all I’m trying to help my mother with this project where she’s a potter and she has all this stuff pottery that’s not glazed and she wants to use AI or something to not only isolate the pottery in the picture but also show her a preview of a bunch of different glazes she could buyI tried using a custom stable diffusion but I failed any ideas

[lucid bobcat]

Use Flux Kontext. It can isolate the pottery (if the photo is good) and add glazes.

[meager jasper]

Anybody used SD for creating game concept art/assets? I'm curious to hear about how that went

[sly silo]

has anyone here had success with using controlnet and openpose to simply change an existing image's pose? for some reason anything i generate is just a bunch of distorted noise idk what im doing wrong 🙁

[valid aurora]

@sly silo haha bro im a big noob too and i encountered so many problems. couldnt make it work for the life of me, in the i resorted to depth and canny, you have to adjust the time step tho to make it work properly as there will be bleeding. i asked other ppl tho and they all say openpose works tho. this is just my experience

[sly silo]

i uninstalled everything

gave up 🌝

probably my pc b.c the more i generate the more laggier it becomes and the slower the generations are

[valid aurora]

ah thats true vram is a thing. oh well bro

[desert dagger]

that's not what it's for. you're using the photo with openpose to create a new image that's posed the same way

[oblique elk]

Do not bother, he left already. Guess not enough patience at all to learn new things....

[vague lodge]

Hello

[desert gorge]

hello

[sweet bobcat]

hello guys. im trying to go into Ai generators and i have extrem hard time instaling the Stable Diffusion on my PC. i have a AMD Radeon RX 6650 XT and i start tinking is inpoible for me to make it work

i try all the COMMANDLINE. and now im trying to find a way with DirectML or ZLUDA but i find so hard 😄

someone can help me in any way ?

[tranquil mountain]

Anyone here make ( Train Lora ) ? If yes dm me I'll pay

[desert dagger]

For which model?

[dry tapir]

Hello everyone!
I've been working my way through stable diffusion and ComfyUI and have taken a rough look at the new models for SDXL and Flux, but somehow I don't understand which combination is best suited to the following scenario:

I have a 3D model of my character, and I have a character sheet of him where you can see him from all angles and so on. If possible, I would like this image to be used in a prompt describing how this character can pose and express themselves, and then I would like to receive a 2D anime drawing of it.

What do you think would be the best combination for this scenario?

[tranquil mountain]

illustrious

[obsidian kindle]

hi

[tranquil mountain]

Dm if you interested

[abstract quarry]

Flux Kontext

[warm junco]

Check the guides of the first pinned messages in #🤝｜tech-support

There are Guides for zluda

Dont use directml

[dry tapir]

Oh Okey thank you! Is a Lora (training) also needed?

[abstract quarry]

usually not

important thing is to generate the image stepwise

e.g., first generate the character in the right pose, then change expression, then transform into anime

[dry tapir]

Thank you very much! Ill look into that

[desert dagger]

Im not, that model is a pain to train for

[tranquil mountain]

oh

[sweet bobcat]

@warm junco i manage to make it run with directml... but is slow af, and to train any Lora is brutal😅 i will try zluda

[bronze cove]

What's the best model for fantasy oil panting character art?

[warm junco]

yea thats why directml is not recommended xD

zluda is much faster

[wanton garnet]

anybody know how to use control net settings with xyz plot? It says online just to select from the dropdown menu but when i open the dropdown menu its not there.

[prisma smelt]

Will this algorithm exhibit any peculiar reactions on the diffusion model? https://www.alphaxiv.org/abs/2508.02124

[young bronze]

PLAY CAVE OF THE CURSED SKULLS 💀

https://rodrigotoller.itch.io/cave-of-the-cursed-skulls

Master your class, forge broken builds, and slay monstrous bosses in a brutal pixel-art dungeon.
Loot upgrades, become unstoppable, die trying—repeat.
Download & tell me what you break. ⚔️

[lapis swallow]

what's the difference between img2img and controlnet and can someone give me an example of when would I want to use img2img + controlnet instead of just text2img + controlnet?

[tawny mauve]

what's the latest people are using? sd? flux? wan?

[oblique elk]

I would say that img2img is more an overall approach which includes colors, composition, pose, etc. The main parameter is the denoise value to influence the result (staying close to the original or having more creative freedom).
Controlnets on the other hand let you use specific elements like depth, lines, pose which are used for the complete new creation of an image.

[oblique elk]

As usual it depends on the goal you would like to achive and your gpu. Most current models for 16g+ cards would be WAN2.2 for videos and for images the most currents are flux krea, chroma model and qwen

[tawny mauve]

Thanks. I stopped paying attention to this stuff like 6 months ago. Then I got a 5090

[oblique elk]

No Problem, with that card you should run into no problems. Still you might look at all the lightning lora stuff for wan to keep the steps low. It can be frustrating waiting five minutes for garbage 🙂

[tawny mauve]

yeah I've only just got into Wan today. I was trying to build pytorch for cuda 12.9, and over the course of the 3 or so days I tried to get it working, pytorch released official wheels for pytorch for cuda 12.9

[thorn hare]

how are AMD cards right now ? does it still require a lot of work making them work on text and picture generators like Forge and Silly Tavern/Kobold ?

[abstract cairn]

Hi

[abstract cairn]

Hi everyone,
I have completed the verification steps but haven’t received the Verified role yet. Could someone please help me?

[warm junco]

Its a bit more setup than for nvidia but its okay currently

[buoyant stone]

your project sounds amazing.
I’ve worked on computer vision and generative AI projects before, and I can definitely help you set up a smooth workflow.
We could use a segmentation model like Meta’s Segment Anything (SAM) to precisely isolate each pottery piece from its background, then apply different glaze styles using a custom Stable Diffusion model or ControlNet. This way your mother can see realistic previews of each glaze before deciding which to buy.
If you’re open to it, I’d be happy to collaborate directly — from setting up the AI pipeline to making it easy for her to upload photos and get instant previews. This could be built as a simple web app so she can use it anytime without complex tools.
If you really want my help, let's have a call and discuss about that with more details.

[buoyant stone]

hello, I'm a full stack blockchain developer and I'd like to collaborate with you on your current project.
now, I mainly focus on blockchain development but when I started my job, I was just an AI developer, I'm good at Python, C++ which are used in AI and also other stacks for frontend and backend development.

I think I can collaborate with you.

If you are interested, please contact me.

thank you

@terse glacier

[autumn ibex]

can you build handsome website for me?

@buoyant stone

[buoyant stone]

sure, no problem

send me dm request please.

[autumn ibex]

ok

I send it

[bright ember]

Retsubu Are you there? I need to ask you for help.

[queen steeple]

Hello everyone, I don't really see where i could ask this so I try here. I'm playing with IP-Adapter and Reforge, and I try to understand what exactly each layer does to the generation. I asked Claude, I checked some videos and the github repo, but it's like I'm just suppose to tweak values without knowing what it refers to.

Anyone with some experience to help me understand ?

[peak tinsel]

Hello Everyone , I’m an ethical hacker offering any kind of hacking related services.
‎Feel free to contact me for help regarding hacking issues
‎

[abstract quarry]

what exactly is your question?

[queen steeple]

When I play with IP-Adapter Controlnet weights, there are 11 values I can modify. I've found on internet they're layers, or channels relating to specific concept IP-Adapter will use. But I don't know exactly what layer 1, layer 2... do. So I'm quite blind, I tweak things but it will work for specific cases then it won't work anymore.
Is there a documentation about these layers so I can understand what is there effect ?

[abstract quarry]

this are the unet layers. Which model do you use? sdxl or sd 1.5?

[queen steeple]

sdxl

[abstract quarry]

so the unet architecture works by step-by-step shrinking the latent image into lower resolutions and then growing it back again

[queen steeple]

I've found this doc, it's that kind of info I seek, but more complete : https://github.com/cubiq/prompt_injection/discussions/8

[abstract quarry]

basically you have down, middle and up layers. You cannot say that a layer does a specific thing. They all do everything into some extent

but you can roughly say: middle layer is for image composition, up layers is for textures and fine details

down layer is a bit image understanding

usually there is no reason to influence the layers individually. But in some cases it can be helpful

[queen steeple]

down, middle, up is in the same order in IP-Adapter controlnet extension ? Meaning the first block is the down one etc ?

[abstract quarry]

very likely, yes

middle is the largest

I think it was 5 layers in the middle and 3 up and down

but I am on my mobile, cannot check the source code currently

[queen steeple]

I try to get consistent faces thanks to it. It worked to generate a bunch of quite similar faces, but now i want to use this batch to keep the same face for the character I generate in various situations

[abstract quarry]

I don't think it makes much sense to change the individual layers for that

it's more like: you want to use ipadapter to transfer the style of an image but not its content, then you only use the last few up layers

or you want to use content but not style, then you only use the middle layers

[queen steeple]

My attempts helped me understand some layers will keep the background or the composition, for example, if I lower them I keep details and the composition is more creative

hmm

[queen steeple]

So number 4 to 8 ?

[abstract quarry]

yes

everything else like layer X is for faces and layer Y is for beard is just "empirical". It will work on some images and be totally different on others. I would not trust these claims

[queen steeple]

so it's better to keep reasoning it terms of "blocks" with these down, middle and up ones you mentionned

[abstract quarry]

yes

[queen steeple]

Thank you for these explanations !

[mossy ridge]

hello!

[fervent thunder]

I got rtx 2060 ko is there any point of trying wan 2.2 or will it take to long to generate image to video

[fervent thunder]

Hi, looking forward to getting this set up and doing some diffusion! 🙂

[desert dagger]

not a chance

[quartz pelican]

Hello, I guess

I just got into working with SD. Where can I ask some very basic questions about working with it?

[tawny mauve]

https://dontasktoask.com/

[quartz pelican]

Right, I almost forgot what kind of people I'm trying to deal with

[tawny mauve]

just ask your question, you have permission to already

[quartz pelican]

What is LoRA?

[tawny mauve]

if someone is around, and wants to answer your question, they will

low rank adaptation, it helps an SD model produce an output with some specific feature, like a specific character or object

[tawny mauve]

a better explanation than I could write, first result in google
https://www.reddit.com/r/StableDiffusion/comments/196ikpo/what_is_a_lora_and_how_do_i_use_it/

[quartz pelican]

Ok, thanks

Glad to know that this server is not entirely infested by bots and scammers like many others

[spring spear]

hello there im looking for a way to control the length of hair does anyones knows a way? im using the Danbooru method (very short hair, short hair, medium hair, long hair etc.) and im looking for a method to use a already existing Image as background if some one knows how i would appreciate the solution (im using a Illustrious model no FLUX)

[oblique elk]

Hi there, normally i would just regenerate as long i get closer to the required image. As it seems you got some pretty clear idea of what you would need the options would be inpainting (mask the part where you want more hair) and hope for the best...
Another way would use gimp, krita, photoshop to clone stamp the hair and use image 2 image with a a denoise level that gives the ai enough room for removing the cloning errors... Another way but you excluded it would be flux kontext and ask to change the hair style, length etc.

[untold moth]

is there any webui that support cloud gpu via ssh?

or remotely via colab but the webui and models are local?

[tawny mauve]

just imagine how long it would take to load models from your machine to the cloud machine

[untold moth]

i just wanna test though

[tawny mauve]

it's possible, but not advisable

no sane application would ever have this functionality

if you have access to a remote machine, load the models and webui on there, and then access it from your local machine

[untold moth]

Let me guess it'll take hours to load the models?

[tawny mauve]

yes exactly, if you have a 16gb model locally, you need to send the data to the cloud machine before it can use it

[untold moth]

So i need to upload it on drive?

[tawny mauve]

what are you actually trying to achieve?

[untold moth]

I wanna test the app if it really work cause my gpu is really worse

Not suitable for gens

[tawny mauve]

you want to test comfyui or a111 ui?

[untold moth]

Do sdnext had the cloud option?

Or not?

Cause that's what I'm using now

[tawny mauve]

you can deploy it the cloud if you wanted to. It may take some work

[sterile plover]

helllo

[dapper jungle]

[Looking for beta testers]Hey guys, have you ever wanted to create/design a character and put it in a game or anime?

My friends and I are trying to create a tool that will allow you to quickly generate your own characters and worlds using AI. If you are interested in participating in the beta test, plz let me know.:))

[quartz pelican]

Huh
good luck

[untold moth]

The thing is i can't find the settings for it

[sterile plover]

hello

[untold moth]

Or it doesn't have that setting at all?

[quartz pelican]

What's your thoughts on Automatic1111?

I heard that Forge is faster, but I cand of can't download it for some reason

[warm junco]

Check the first link of the pinned messages in #🤝｜tech-support for all the setup guides.

[quartz pelican]

I can't download the Forge archive. I've already installed A1111 and now useing it.

[warm junco]

Okay Auto1111 works just fine. Its just a bit more outdated than Forge

[quartz pelican]

Fine by me

[autumn ibex]

hello, are you looking for beta testers?
it is my special field,
can I have a discussion with you on that?

[quartz pelican]

No, buck off

[autumn ibex]

sorry, I sent a wrong message, it was for Norko

[autumn ibex]

Hello, I can help you with beta test, as an experienced developer.
if you are interested , please contact me.

[buoyant stone]

Hello, how can I help you as a senior blockchain developer?

[untold moth]

Did you ever hook a cloud gpu to locally hosted webui?

Cause i really don't know if my webui supports that.

[gaunt salmon]

[rain yarrow]

helllo, i need some help

[versed notch]

I am looking for a business USA paypal.

[buoyant stone]

hey, what problem with you?

[rain yarrow]

Well Am using Forge UI and it does not want to do png info to txt2img WITH the picture. The txt2img TEXT works.

[buoyant stone]

Ok, so you need to update the UI?

[rain yarrow]

i did update it i think

[buoyant stone]

In my opinion, txt2img only works from text; it doesn’t take an image as an input. If you want to start with an existing PNG and then modify it with text, you’ll need to use img2img instead.

[untold moth]

@buoyant stone?

Is it possible with sdnext

[rain yarrow]

it should do that. You can print from PNG info to txt2img

because that is where my controlnet is ;c

[untold moth]

Like running ssh on colab and use that gpu on locally hosted sdnext

[buoyant stone]

sorry but I saw your words now

[rain yarrow]

is i maybe because i am running a wrong version of python?

[untold moth]

No worries

I'm just curious though

[ionic onyx]

Hello all! new here stopping to say Hi!

[untold moth]

Since i can't find settings to enable remote gpu access from colab. And i didn't mean running webui in colab.

[rain yarrow]

@buoyant stone https://youtu.be/fJsi-swa5ZY?t=69

AND IT TAKES THE image to styles D:

also what does this mean? RuntimeError: mat1 and mat2 shapes cannot be multiplied (154x2048 and 768x320) I picked the wrong model.. it should be illustrious..

[buoyant stone]

sorry for responding late but I'm very busy because of my project.

[plain ledge]

Hey guys, im new to stable diffusion, kohya and LoRA, im having difficulties setting things up. If there’s anyone who could help me it’d mean a lot. I’ve been stuck with the same error message for two days and ChatGPT can’t even help

[buoyant stone]

I think I can solve your problem based on my experience, but first detailled information is needed
dm me

[plain ledge]

Dm me if anyone’s free

Could we get added?

[limpid ravine]

Where do I find all the models? Is there any place I can see what the different models do?

And which version do I pick with 32gb RAM and 5070TI?

[rain yarrow]

you cna find models on hugging face :). Depends what you want to make

orr! on civitai.com

[limpid ravine]

Any way to sort on sites like Civitai? It's riddled with shit.

[rain yarrow]

yes, You can set filters

[limpid ravine]

God so much furry

[rain yarrow]

yea...

make an account or gib email and set a filter for that shi

[limpid ravine]

How long does it usually take to produce an image?

And which version do I go for with 32 gb RAM and 5070TI? Flux?

[rain yarrow]

oh

right-click on webui- user (BATCH, not shell script) and edit with notepad, something with arg add this --cuda-stream --cuda-malloc --disable-gpu-warning

[limpid ravine]

I still need to figure out which version to use

[rain yarrow]

this is at the very beginning 😄 it is really REALLY necessary.. depends how long it takes really, what you want to make and how detailled etc

[limpid ravine]

It depends on hardware doesn't it?

[rain yarrow]

kind of

[limpid ravine]

My question is, do I use Flux with what I have?

Or XL?

[rain yarrow]

I am not too helpful in this, I've started 2-3 days ago with all this 😄

[abstract quarry]

with a 5070 you can use flux

[red veldt]

What’s up any Lora chefs? Need help please

[fervent thunder]

Whats the minimum GPU memory you need to be running fun models?

[glacial glade]

not sure

[red veldt]

[untold moth]

Is it sdxl (pony, illustrious)?

[opal niche]

howdy

[opal niche]

pretty new to this was trying to find models for fooocus off civitai but didnt understand how to find them are they paid for under the membership or an i just missing it

[random falcon]

Can anyone help me, pls? I installed flux1.1 dev in comfy ui to run it locally. Which installed some 23GB. Got a working prompt to image model in it, but don't know how to run an image-to-image model in Flux1.1 dev. How to change it to img -img in it?

[warm junco]

Fooocus doesnt have any extra models.
You can use any sdxl based model with it.

[abstract quarry]

just use an img2img workflow template if you don't know how to do it yourself

[rain yarrow]

no, not all models are paid, you can just look at it on hugging face or civitai 🙂

[abstract quarry]

Comfyui has a list of default workflows/templates

[untold moth]

Is there any extension for sdnext that can download civitai models

[lunar solar]

hello

[lunar solar]

fine

[still glacier]

Just in case you got dm'd by this @rapid summit guy, don t join their external discord server whatever. It s a scam.

[oblique elk]

Nowadays it seems easier to identify them as they using the mod server tag to make an more legit impression 🙂

[lunar solar]

OK

thank you

[still glacier]

shhhh don t tell them.

but yes, some do make it extra obvious.

[brittle musk]

helllo

[rain yarrow]

hello 😄

[rain yarrow]

nono, you don't need that 😄

[floral lance]

Is SDXL still be best base model for training a graphic art sort of style?

specifically on art styles that it's never seen before

[rain yarrow]

Interesting question, sadly I have no notion about this 😭

[rain yarrow]

what is that weird thing i see on youtube when people are typing and then it autocompletes their search queries in forge

[young cliff]

Hello hello people.
I was wondering if there is a place (ie. website, youtube channel) with good tutorials to properly use Stable Diffusion and/or ComfyUI?
My ultimate goal would be to create good comics in the end (I know Stable Diffusion/ComfyUI won't be enough and some Photoshop/Gimp skill will be needed later). Is there a good "tutorial path" to follow to attain such dream? Like "Writting good prompts" "Creating loras/Consistant environment, characters, etc." "How to use ControlNet" etc..
I'll be looking forward your answers. So far, I've only found a few tutorials, there and there, but it seems like I'm always missing some skills at some point. If you guys have a good path for me to follow, I'll gladly take it.
Thanks

[opal niche]

thank you for the advice and info

[opal niche]

thank you, the only other question i gues at this point i have is long term should i branch out i just found fooocus on accident and gave it a try out of boredom now im hooked playing with prompts left and right is it possible to make videos to or would i have to branch out a bit

[rain yarrow]

no problem hope it is useful 🙂

[inland wedge]

How do I generate an image where 2 characters are facing each other? e.g I tried to generate a 1940s bar image where the bartender is serving a customer but it kep fusing them together?

[abstract quarry]

on civitai there are several models that were trained on multiple checkpoints (sdxl, illustrious, flux, ...) and I never found a huge qualitative difference (although they rarely share their training data, so hard to say which model follows the training data best)

flux is definitely really bad in styles, but that doesn't mean its bad in being trained on styles

[gloomy swallow]

chat anyone knows how to convert SDXL to Onnx model?

[warm junco]

Are you on AMD or why do you need them in onnx?

[gloomy swallow]

Yes. Trying to run on Intel and AMD hardware.

I tried the existing safetensor ckpt --> diffusers (using hf diffusers lib) --> onnx (using optimum lib) pipeline, but its not working.

I want to port to Onnx specifically cause its hardware agnostic, so i can also run on other CPU / Hardware.

[warm junco]

What's your GPU?

[gloomy swallow]

CPUs / iGPU. not mine, but for users of my project basically. I'm creating a GUI tool and want to integrate SDXL to it.

[warm junco]

Ah okay I think stable-diffusion-webui-amdgpu fork with directml or AmuseAi have boath onnx converters integrated

But for AMD using Olive+onnx makes not so much sense currently

[gloomy swallow]

yeah i was looking for converters and only found olive ones.

I wanna integrate a completely standalone one, so i think i might have to copy parts of that.

[rain yarrow]

hi guys

i tried to upscale a model, but then got this error RuntimeError: Given groups=1, weight of size [128, 3, 3, 3], expected input[1, 4, 1024, 1024] to have 3 channels, but got 4 channels instead
Given groups=1, weight of size [128, 3, 3, 3], expected input[1, 4, 1024, 1024] to have 3 channels, but got 4 channels instead

[gloomy swallow]

can you link me the fork url / source?

nvm. found it

i assume this codebase https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu-forge ?

[rain yarrow]

aa nevermind. Used the wrng upscaler for IlluXL

[gloomy swallow]

cant find any script in this codebase. do you have a link?

[warm junco]

Nope sry I can recheck later if I find it or if it got removed

[gloomy swallow]

found a a1111 extension that suports runtime

https://github.com/tianleiwu/Stable-Diffusion-WebUI-OnnxRuntime/blob/main/ort_model_helper.py#L30

they a SD Unet that is compatible with 1.5 and XL

!!! cool and sweet !!

[warm junco]

Oh nice!

[haughty talon]

Good day to you all 👋👋
I'm Adebayo, i'm planning to venture into Agentic AI.
I hope you all support and guide me through the journey 😀😊
Thanks in advance 😎🤗

[errant bronze]

Hi, I’m an an AI Engineer specializing in machine learning, NLP, and generative AI. I build scalable, real-world solutions that turn data into intelligent products.
Open to new opportunities, let’s build something impactful together.
Thank you

[stone locust]

I want to write about it, but I can't...

Are there any other servers besides SD?
Help desk, etc.

[still glacier]

if you got asked in dm to join another "help desk/discord" or whatever. Don't join, it's a scam.

[untold moth]

Is there a way to run remote gpu from colab on local webui?

[leaden moat]

if anyone needs lora training let me know

[brittle musk]

Hello, I want to create a workflow that will allow me to make cartoon/animated videos easily with a couple of prompts.

Here is the breakdown of steps i was thinking

1 - Creation of all the characters and their expressions and different angles.
2 - Animating the characters made in the first step by giving prompts. and creating multiple clips of these
3 - editing them together in a video editor.

i want to do all the steps locally in my computer as it has a nice GPU. it would be preferred for the tools to be free no issue if they are paid

Can you suggest me the tools and any additions to my workflow

thank you

[valid aurora]

@brittle musk https://www.youtube.com/shorts/r04t9q9NULE not a pro but from my own basic research it seems wan 2.1 vace can do this (no 2.2 yet afaik)

[open night]

I upgraded my gpu and now I can't really use my old automatic111.
I'm thinking of just upgrading my stable diffusion. What's the easiest one to upgrade to?
I want to reuse all the models and lora's I have. Ideally same kind of ui, forge can do that right? I can just drag and drop my models and lora's to forge's one

[oblique elk]

Fast mods today.

Thanks a lot

[valid aurora]

@open night wdym by upgrade? changing ur ui or changing ur stable diffusion model? i went through the same path as u started with 1.5 in a1111 then forge with sdxl. forge is good. similar ui and ya afaik u can just drag and drop ur loras and files (confirm in tech supp tho)

[open night]

I mean straight up installing something new as an ui, once that's done grab the models and lora's I still have on my drive and put them somewhere else. Although I would like to try flux which people talked about which is a nice bonus

ah that'd be nice

ill check thank you

[valid aurora]

@open night "Linking Models, Loras, etc. from other Webui's or Folders to Forge:
If you want to link all models from an other Webui you can do that by editing the webui-user.bat like this where you set the A1111_HOME to the Path where your Automatic1111 is installed or where you store your models. Here is an example:" u can also do this but tbh id rather just copy and paste xD

its pinned in the tech supp for more info u should check the guide there was super helpful for me

[open night]

Yeah I was hoping to avoid weird manual windows linking or having to keep my old installation folder. I just want to drag and drop if that makes sense, but if I have to i'll look into this 🙏

[valid aurora]

dw i feel u i myself am dumb af if i got this far u can too. good luck with the journey

[open night]

thank you

[robust cave]

Is there anyone looking for dev?

[rapid vector]

Hey! I'm looking for someone who can give me advice and a little guidance on image generation. The truth is, I'm new to this and would like to learn more. If anyone is willing to teach me, please send me a message.

[sonic forge]

**👋 Romeo | AI/ML Developer👋 **

Hi, there. I am looking for a paid job or work as a developer with 8 years of experience in AI/ML and WEB development.

Mainly, I focus on Voice AI agent, AI-powered chatbot, Automation, Data Science, Computer Vision and Web Development.

Voice AI agent: Vapi.ai, Retell AI, Twilio, Asterisk, 11labs, etc...
AI Chatbot & NLP: RAG system, Prompt Engineering, STT/TTS, LLM models such as GPT-4.5, GPT-4o, Claude 3-7 Sonnet, Llama-4, Gemini2.5, Mistral, and Mixtral.
Automation: n8n, Zapier, and Make.com, etc...
Model Deployment: Runpod, Replicate, Huggingface, etc...
Program Language and frameworks: Python, FastAPI, Flask, Django, Node.js, React, JavaScript, TypeScripts, Express, Next.js, Nest.js, etc... (Lovable.dev)

🌐 This is my portfolio: https://romeo618.vercel.app/

In addition, I always try to learn new and cutting-edge technologies, and I place great importance on collaboration with team members in development.
If you have any idea or project, plz DM me.
Thanks

[quartz pelican]

What is considered a good prompt?

What is the extend of possibilities of basic SD 1.5?

[abstract quarry]

the exact same way you would do it with flux dev

[worn burrow]

Hi everyone

[thorn hare]

if i cant get into ComfuUi, is forge the best alternative ?

[true shuttle]

Why can't you get into comvyui?

[fervent thunder]

Hi. I am new.

Where can I learn the basics from?

[true shuttle]

Just from the readme file on Github.

Or can ask me for a help. 🙂

[warm junco]

whats your gpu?

[thorn hare]

i tried and its just too open, i like the simplicity of things like Stable diffiusion

[thorn hare]

3080TI 12GB ram

[warm junco]

Then yes forge is good alternative as with 12gb vram your probably not going into video generation anyway

In the first link of the pinned messages in #🤝｜tech-support you find the setup guides.

[prisma horizon]

16 gbvram are enough to gen videos?

[tawdry stirrup]

Hi, whats the best way to run SD on a mac these days.. i've ran it with automatic like 2 years ago but that project seems sleepy..

[prisma horizon]

damn, they almolst got me this time

luckily discord warned me

[oblique elk]

Depends on your taste for UIs. I would say the easy way would be tools like draw things. Invoke (community edition) works too on mac and finally comfyui works also. Still not very pleasent using comfyui to get most current models just to see it takes 30 minuten for one generation 🙂