💬｜general-chat

[foggy ice]

I hadn't been in there in a few years.

[atomic mortar]

not to mention their model wasnt the greatest iirc all the models you can find on civitai in the sdxl-sd3 are performing better imo

[warm junco]

It still exists

[grave stag]

I have bounced on and off of setting up image generation and twice I tried but idk if I can just say in plain English bla bla xyz and then it makes those exact fine tunes precise edits like an artist would when getting into image generation for my specific refs I'd use for RP and especially since it would have some way of telling OK you look for it it be AI (not like the 4 hours comb through pixel by pixel though). Or am I wrong?

That said realizing how much video generation has caught up...can I use it to uncensore the anime shots that just have unrealistic amounts of soap or steam just to be TV compliant? I'd edit those clips for sure and make a lot.

[atomic mortar]

Hmm question 1 i don't fully get what you mean.

i do know there's a video edit model out now by LTX but the hardware required is probably out of most peoples price range. Personally i haven't messed with that but could work but manual inpainting per frame would be faster

[grave stag]

I don't know, sorry half the time my brain types what everyone reads as gibberish but my brain says makes sense to me...so I press send not realizing how bad it is.

TLDR, for question one, basically how granular on random fantasy race that isn't just an elf can you get for edits, like can you say change xyz and really fine details if it generated an image you like mostly already? Because to my knowledge you can't only broad generate me x girl or guy in x outfit and that won't be enough to sway mW back to AI for image generation.

[atomic mortar]

hmm so let me rephrase a little:
first question:
if you have a image you like already but you want to edit it further? like small details for lets say from elf ear to demon horns?

second question?: are you assuming that AI can only do vague descriptions? like eld in a red dress and not more spesific? and that youll only use AI gen if you have more control over it?

do i get it right?

because 1. you can edit fine details on existing images you generated yes.
2. well with the newer models you can get very spesific on what you want but for older models you probably have to use a lora if its a custom race but you want elves with pentagon shaped pupils and sparkling wings? go for it

[faint timber]

I'd use Qwen Edit. There are other edit models.

[ember crag]

heyo brother are you still kickin it around

[warm junco]

Hey

[errant swan]

.

hi

[atomic mortar]

Hi hi

[deft topaz]

Hello guys

[carmine sandal]

hi

[loud hawk]

Hello

How are you doing

[atomic mortar]

🕵️

hey guys

[carmine sandal]

Hi guys im new how can i make it actually listen to my prompts? Im using sdxl with animagine xl 4.0 with an anime Lora

Like i want it more nsfw

[atomic mortar]

hmm well i cant help with the nsfw part but i can give you some tips on prompt adherence if you got some sfw examples

[carmine sandal]

Cant or wont?

[atomic mortar]

and ngl your better off using illustrious for anime models

cant help in the server due rule 2.
however i do think by switching to illustrious and following that prompt style youll get what you need

[shadow ingot]

Could anyone clarify for me the origins of illustrious and flux 2 checkpoints?
Is illustrious refined from pony or from XL base?
And is flux 2 trained from zero or is it derived from flux 1?
It is to more or less know if LoRAs may be compatible.

[atomic mortar]

you can use pony / illustrious loras together somewhat because they share common origins but youll see a lot of quality degredation. in some cases it can be worth it to still use it however

for Flux 1 and 2 use a different architecture so the loras are incompatible there unless its a very spesific one but that should be mentioned on the page

you redirect him to a cloud platform for when he wants to run locally?

[shadow ingot]

Thank you, it is more or less as I expected. I guess I will keep my illus and pony libraries separated then.

[atomic mortar]

🫡 always a pleasure

[fiery ore]

Just an option btw

[atomic mortar]

Reminder guys people asking to DM and especially from people who just joined should be treated with extreme caution

If something looks too good to be true it probably is.

[frigid sundial]

is the equilivant of meeting someone in the pub and the first thing they ask to do is go and chat down that dark alley where nobody can see 🤣

[bright reef]

How to set up simple workflow (in ComfyUI) for flux.2 to fill up masked areas of image? I usually use it for outpainting, but sometimes I need to do outpainting of skewed (rotated) images. At first I tried to give it png image with transparent areas, but it doesn't work well with those. Also tried giving it image with white background with prompt "fill up empty space with background", but results are mixed (it either doesn't fill up white space properly, or fill it up but in a weird way), also tried to mask in inside of editor, but results are not good too. Currently working with default workflow

[bright reef]

Also it makes output images smaller

[atomic mortar]

Hmm personally im not too experienced with flux2 in comfy workflows since its heavy as hell but i do know inpainting works by surrounding context so using basic colors and minimal detail to fill in the missing area and inpaint over that sometimes does help (and increasing the creativity)

Making the output smaller is a workflow limitation but i dont use comfy, maybe someone else pitches in for that one

[bright reef]

It's surprisingly fast

[atomic mortar]

Ah depending on the model of flux 2 i suppose yeah

[bright reef]

Managed to get some good results after trial, errors and few crashes

[atomic mortar]

Imo it was outpaced by z image really fast for me

In results and speed

[bright reef]

Either giving it image with prompt "fill up empty space with background", or by using inpainting preprocessor with png image with transparency as mask with prompt "fill up empty space with background, continuation of the existing scene, matching lens perspective, style, lighting, shadows, color balance, and background detail,"

But still can't figure out why output image is smaller, have to writing size by hand

[atomic mortar]

If its the basic workflow it makes sense

Wanna show me the workflow in #🏞｜general-with-images

Also why comfy specifically if you dont mind me asking

I'm also doing some garden work so i might vanish for a second

[bright reef]

ummmm... I use content that doesn't belong to me, I take it and extend to 16x9 to use as wallpapers

So I can't post those images

[atomic mortar]

Just the workflow, not output!

[bright reef]

It's literally default flux 2 workflow

With inpaint node

[atomic mortar]

Hmm you could match the input size as the output size resolution

Or increase it as your outpainting

hmm yeah what kind of flux 2 are you using? klein? dev?

4b, 9b?

[bright reef]

image_flux2_klein_image_edit_4b_base

[atomic mortar]

ahhhh i see yeah the bigger problem of it not filling up properly is a 4b problem

and using the edit feature like that limits the size, i thought you were using the other non klein model

[bright reef]

Figured out why scale is weird. Now gotta figure out gow to stop it from altering image itself

[bright reef]

I've got 8gb...

That's the best I can afford

[atomic mortar]

yeah that will limit the size and what you can use im afraid. hmm well if you dont mind learning comfy a bit more or some manual work, you could split the image in half > process it > combine them again

but that really depends on the images you work with ofc

[bright reef]

nah, I'm kinda ok with what I'm getting now

And I ofthen had different problems with image segmentation

[bright reef]

Is it just me or flux 2 upscale is very noisy?

[bright reef]

what z-image model do you use?

[atomic mortar]

Turbo mostly but base is very nice

Havent used the anime one yet

[bright reef]

Turbo is the one that is not full? Have you tried the other one?

[atomic mortar]

Thats the base one yeah, i tried it on release but turbo was doing everything i needed it to

[plush plover]

hello

morning all

[plush plover]

Is there a consensus on the models that most of you guys use especially for mythical or fantasy stuff? And what is the policy regarding posting some work and stuff like that. LOL

[atomic mortar]

Anything sfw (rule 1 &2) but personally i like illustrious/ for most of my gens but i have to say z-image is really good
But nothing beats sd1.5 in terms of creativity

[plush plover]

well, i'm using the main sd app but i've downloaded a lot of diff models, juggernaut and zavy my fave. is this the channel people post stuff?

i'm doing everything local on my server

looks like i cant paste in here lol

[craggy escarp]

Hi, I have a problem with Flux2 Klein. I'm trying to use inpainting to remove text, but in the areas where I applied it, there's a blur issue, it looks kind of smeared. The second problem is that in those areas the color changes compared to the surrounding environment. Is there any way to fix this?

[steel ridge]

Hey @craggy escarp,

This is usually caused by high denoise and soft mask blending. Lower the denoise, reduce mask blur and make sure you’re using full-resolution inpaint. That will fix the blur.

For the color shift, keep denoise lower and ensure the inpaint has enough surrounding context so it matches the original tones properly.

[steel ridge]

If you want, you can share your workflow and I can help you fine-tune it exactly to remove both issues completely.

[atomic mortar]

general chat cant handle images, other channels can

[atomic mortar]

so what llm do you use to awnser everyone

[atomic mortar]

people generally use #✨｜sdxl or #🍥｜anime here but seeing you use juggernaut #✨｜sdxl is the way to go ;)

[pastel socket]

If you crank the learning rate you think you could manage a 1 image lora in qwen image edit?

[atomic mortar]

why not use a model like qwen edit or klein to make more images for a dataset to make a lora

[pastel socket]

Yeah that's another way

The benefit of one image is speed of course

[atomic mortar]

but technically YES its possible, even more possible to cause image burn and mess the lora up

[pastel socket]

I was kind of thinking about rag setup with auto lora

[atomic mortar]

not sure how much qwen edit will like it but in the sdxl era it was doable

rag setup? i only know koyha and onetrainer

[pastel socket]

Yeah you could search for the image that matches the prompt, train a lora, load it, then generate

Does anyone do that?

[atomic mortar]

first time ive heard of it

[rustic tundra]

Joining the group from Pakistan ❤️‍🩹

[raw current]

Yeah

[steel ridge]

Yes, some people do that but it’s not really the efficient way to train a good loras

Most workflows involve first building a clean dataset then resizing/captioning them properly and training with tools like Kohya or ComfyUI LoRA nodes. Searching individual images from prompts tends to be slow and gives inconsistent results. @pastel socket

[twilit mica]

@vivid hull am i allowed to send images made by models other than stable diffusion

[vivid hull]

i don't see why not. and i'm no admin, i've just been here a long while.

for most old-timers here it started with stable diffusion

but we moved on to flux, to chroma, to illustrious, to a variety of models

[atomic mortar]

Yes of course from what i currently understand. Though I'd keep the related channels for their intended purposes (like no flux in sdxl )

[twilit mica]

Cool thanks y'all

[unique ravine]

tbf i dont see channels related to those

[atomic mortar]

well yes but we have #🏞｜general-with-images and #🌶｜off-topic , #🍥｜anime
its first and foremost a stable diffusion server afterall

[unique ravine]

true

[craggy escarp]

I’m trying to use Flux2 Klein to remove text. When I use inpainting on an image, the area where the mask was applied doesn’t fully remove the text it becomes less noticeable, leaving ghosting. Is there a way to fix this?

[atomic mortar]

are you using 4b or 9b?

and what prompt are you using?

[craggy escarp]

9b, prompt i use: Remove all text and lettering. No words, no letters, no numbers. Fill with clean surface matching surrounding area.

[atomic mortar]

hmm thats strange, it should work normally, do you like hand inpaint or something?

like manually mask it

[craggy escarp]

I'm using Inpaint Crop node

I do it manually by hand

[atomic mortar]

what if you let the model go nuts on its own

probably some finnicky stuff going on

[craggy escarp]

Yes, processing the full image works, but it slightly changes the color palette, also it's difficult to process the full image at high resolutions, for example 2056x2880, without inpainting the model removes text perfectly on the full image, but when using the mask it partially restores the text.

maybe the problem is in the node settings?

[atomic mortar]

i think its partially because if you hand inpaint it, the area for it becomes too small probably

you could try to increase the masking area

because if its masked. the surrounding area outside it doesnt exist

could also try increasing creativity but if its masked like that you dont need flux klein perse tbh

[atomic mortar]

@pallid vault rule 5 no advertising.

[pallid vault]

Am sorry

[atomic mortar]

no worries

[oblique lotus]

hi

[atomic mortar]

hi hi

[fierce crag]

guys why is there no wan 2.2 in comfy ui browse template and how do i get it

[atomic mortar]

is your instance of comfy uptodate? i see the 2.2 14b templates just fine

[fierce crag]

actually the tutorial i followed before it was old and had a wan 2.2 animate something. i was looking for that

now i figured out ig wan 2.2 14b is the same

[fierce crag]

could u help me out a little.
if i want to make a character dance using a reference video
wan 2.2 14b will work right

[atomic mortar]

i assume so yeah

though wan2.2 is very very beefy to run

[fierce crag]

oh i need a crazy gpu??

[atomic mortar]

what gpu do you have?

[fierce crag]

😭 3060ti rip ig

[atomic mortar]

im afraid you cant run the 14b model unless you have a crazy amount of ram

but the gen times wont be fun

[fierce crag]

so i should get the 5b one right

i dont mind the gen times being very long tbh

[atomic mortar]

at 480p it should be possible i think using a Q4 GGUF

but

how much ram do you have

[fierce crag]

nvm ill just pay for some subscription online 😂

16gb ddr4 lol

[atomic mortar]

if you have 32gb of ram and usingf a Q4 K S or even a Q2 with a fully offloaded T5 it might be possible

ah rip

i recommend online services then yeah

[fierce crag]

yk any online websites free/cheap?

[atomic mortar]

hmm the ones i do think are trust worthy are civitai but theres also the option of renting a gpu online but i only recommend that if you know what your doing

[fierce crag]

also which models support nsfw stuff?

theres a wavespeed ai website, it has multiple models init

[atomic mortar]

i am not familliar with those extra sites and you should be careful to not get scammed but i recommend checking on civitai.red for more free model examples under the video tab though i think WAN is what you want

[fierce crag]

wan has a free website which can generate stuff for looong q times
ig thats my solution for now

[atomic mortar]

well for sfw you can use that but for nsfw i cant help

[gloomy solstice]

Hello I'm very excited to be part of this community and know too much content creators for doing collabs or proyect together

[unique ravine]

https://tenor.com/view/side-eye-dog-suspicious-look-suspicious-doubt-dog-doubt-gif-23680990

[latent sedge]

hello everyone im a beginner

should i buy the beginner memorial select gacha? or just roll on another banner

[atomic mortar]

the what now? i think your confused

about what kind of server this is for

[latent sedge]

wrong chat mb

💀

[atomic mortar]

lmaoo

[shrewd gorge]

any interesting new models? I'm seeing if theres anything better than QWEN

[atomic mortar]

Hmm depends on what you mean by better, honestly z image is really good but in terms of creative styles/realism i still like sd3 a lot

[atomic mortar]

But for purely realism i use zimage

[atomic mortar]

[queen ocean]

hi

anyone there who wants to read urdu novels?

[atomic mortar]

read as in how

like free?

[nova lily]

Is it possible to achieve more advanced image2image functionality on local GUIs?
(such as subject moving while keeping the image intact/easy inpaining)
or is local just too "one dimensional" for such functions still?

[wheat venture]

Cpu and gpu rendering sure does make a difference, 2 hours for a image off my cpu and 1 minute from my gpu

[cursive sage]

Yes, it’s definitely possible on local GUIs they are not “one dimensional” at all, just a bit more manual to set up.

[fervent thunder]

How to stop auto saving videos on iPhone 13 when posted

[atomic mortar]

Like on discord? Settings i would assume

[atomic mortar]

Yes klein does that but its a mixed bag since ai still is random at times

[fervent thunder]

I found it on tik tok was in posts thank you

[valid aurora]

is anyone else having problem with custom audio? i tried so many workflow but none of them sync to my audio. they speak fine without custom audio tho

[jovial sphinx]

What model do you think is better for I 2 v or v2 v? I am using wan 2.2 or wan animate

[nova lily]

Wats klien?

[atomic mortar]

its a different model! its basically an editing model where you load in a image or two and say "replace the outfit from image 1 with the outfit in image 2" and it does the rest for you

it makes mistakes here and then but nothing a retry cant solve

[nova lily]

Ive just been using chat ai to edit, copilot/grok/meta etc. for now until I find some dedicated sites or how to do anything close on forgeNeo

[atomic mortar]

I mean that works, but i think forgeneo can use flux klein 4b/9b depending on your GPU

[uneven plume]

Hi

[solemn otter]

What do you think about an RX 7900XTX with WSL2 and ROCm, does that make sense?
I can buy a new RX7900XTX with 24GB VRAM for €1000.
For an 3090 or 4090 I pay much more for a used one

[warm junco]

No need for wsl2,
im using the 7900xtx on windows with native ROCm

[solemn otter]

Is that fast? Gemini says, that it is way slower than on Linux

And can you use every plugin and modells, or can you only use programs that are optimized for amd?

[warm junco]

Its very fast and on par with linux.
Everything works but sometimes its a bit more to do than for nvidia gpus

If you want to just play around and generate some images or videos its totally fine.
But if its for a business project or something else then go for nvidia

[solemn otter]

So I wanna play around first, but if I’m happy with the results I wanna use it for online shops.
With cloud models I can’t get good results for some products (cause the ai uses the product wrong). So my idea is to train a model and use ComfyUI and I hope I can fix this problem.

But I don’t know how good the results from a lokal ai are compared with cloud models. I wanna make photo realistic images

[warm junco]

That depends on the models you use.
Newer models like Z-Image and Qwen or Ernie are very good at composition and realism

[mortal haven]

Hello is there anyone that needs instaraw workflow here

[atomic mortar]

if your selling thats not allowed here otherwise Hi

[mortal haven]

Sorry I didn't know it is not allowed

[cerulean zephyr]

Hey chat. So is this BYOM kinda thing? Can you connect your local/openai compatible endpoint/key?

Yes no?

Fuck it. Good bye

[atomic mortar]

Oops, mostly local yeah, servers sorta slow

But using UI's like comfy you definitely can use api aswell

[atomic mortar]

🧹 keeping it clean

[still glacier]

rule 5, cf #✍🏼｜rules-and-tos

[atomic mortar]

? Theres no sales here

[woven sequoia]

Sorry, posted that in the wrong server.

[jovial sphinx]

What is the best set up in k sampler for body swap

[twilit tendon]

你好

[atomic mortar]

Lets keep it English chat

[twilit tendon]

ok

[south basin]

hello community 🙂
im looking for a collection of generated video examples (say hundreds), from any model, where input references and prompts are available. anyone knows where i can find something like this?

[atomic mortar]

your best bet is CivitAI but not all posts have their prompt attached

input references however not really

but prompts yeah

[south basin]

thank you i think civitai does have a lot with prompts available but i kinda want references as well. any other sites like that maybe

[atomic mortar]

not that i know of

[mortal cloud]

All of my generated .png files have the complete meta dat including, model, seed, etc, BUT, I have never put any of my images online. So, yes the meta data exposed on Civitai is probably your best bet vs generating your own files (if .png they will include all the metadata)

[latent sedge]

what are you doing

[atomic mortar]

I goto bed for 5 hrs and i wakeup to 30 pings

Cleaned up 🧹

[swift prairie]

animecouple, holding hands while watching the end of the world together, if the worl was ending i'd wanna be next to you vibe, faraway shot from the train, dynamic shot, red and black colors only, rough drawing, lineart

[atomic mortar]

Doesn't work like this

Please read #artisan-faq , theres no other way of generating images in here but locally

[valid aurora]

"if the worl was ending i'd wanna be next to you vibe" 😂

[atomic mortar]

No idea what the thought process there was

[still glacier]

not necessarily a bad idea, it can be fun to throw vague prompt at the random machine and see what comes out of it.

[atomic mortar]

You didn't see the timeout i did for a minute and the other three pings

I deleted

Last night i got 30 pings from the same stuff lol

[still glacier]

I ve been in the tranches for a while. I ve seen stuff....

People do get confused and start spamming prompt in random channels instead of reading the onboarding / rules for sure

[atomic mortar]

Oh this was directly pinging me 30 times with the same 4 prompts less then 10min or something

I think automod eventually timed them out

Anyways time to order food because work got extended by an hours and my train was delayed so i dont wanna cook

[atomic mortar]

I'm busy for 10min and then they strike

[twilit tendon]

sorry to hear this.so whats going on,what happened?

[unborn hedge]

🤖

[atomic mortar]

gone are the bots

[gentle prairie]

man, everyone is switching to anima... i kinda don't want to start again from scratch with all my prompts, lol

and loras too

[unique ravine]

Idk who everyone is

[gentle prairie]

sorry, i should have precised, like a lot of people who do lora i follow who were doing illustrious switched to anima

[gentle prairie]

tried using it following the model instructions, i don't get great results, dunno what i'm doing wrong... i think i'm gonna stick to illustrous for now

[rotund pilot]

hi

[trail wyvern]

Hi

Are there people who test apps on Google Play before they are released to the public?

[atomic mortar]

Probably not but I'm afraid this is not the place for that

[cloud yacht]

Hi

[buoyant wren]

hi

[atomic mortar]

Hi hi

[twilit tendon]

ni

[mild grove]

hi kkk

[rotund pilot]

hi

[modern pagoda]

Guten tag

[static panther]

i

[velvet lava]

hihi

[wispy moss]

hi

[swift vine]

Hey everyone! 👋 Just wanted to share a new anime/illustration SDXL model I’ve been cooking up: SpaceDiffusion-XL V-PRED 2.0.
It’s built on top of AetherV-XL (a NoobAI & Illustrious V-Pred merge using SVD/Semantic Routing) and then heavily fine-tuned for hours with DoRA using curated high-res datasets (up to 1.7x resolution). It mostly knows anime and pop culture stuff up to late 2025, with a tiny bit of early 2026 data sprinkled in.
I'm not gonna overhype it as the "ultimate" model, but I spent a lot of time ensuring the FP32 to FP16 merging process was flawless, so the anatomy, lighting, and style consistency are pretty solid if prompted right.
⚠️ Quick heads-up before you try it:
This is strictly a V-Prediction model. You absolutely need to set your sampler to v_prediction and use CLIP Skip 2. If you use standard SDXL settings, the outputs will come out completely deep-fried.
Would love to hear your feedback or see what you guys can generate with it! Cheers 🍻
🔗 https://huggingface.co/Bl4ckSpaces/SpaceDiffusion-XL-V-PRED-2.0

[atomic mortar]

Wasnt clipskip completely obsolete after sd1.5

[rotund pilot]

is there anyone that can help me train my lora

[atomic mortar]

Didn't you get a ready made config from someone else already

[rotund pilot]

🙂

[rotund pilot]

I trained my lora but when i type the trigger word "gran" then lora it just generates old women, rn im trying all the epchos via Script -> Prompt S/R "lora:gran-000001:0.7, lora:gran-000002:0.7, lora:gran-000003:0.7, lora:gran-000004:0.7, lora:gran-000005:0.7, lora:gran-000006:0.7, lora:gran-000007:0.7, lora:gran:0.7"

[atomic mortar]

the first few epochs wont work well, and you sorta didnt use a unique trigger tag so some character/concept bleeding is expected

[rotund pilot]

i already trained

is there a way i can change the trigger?

[atomic mortar]

not after the fact no

[rotund pilot]

fuh

look at #🏞｜general-with-images

pls

[atomic mortar]

Ai based Nutrition

I'm assuming a translation error please do elaborate

[swift vine]

It's obsolete for vanilla SDXL, but mandatory for modern Danbooru-tagged architectures like Pony, Illustrious, and NoobAI

[atomic mortar]

Guess ive been using sdxl, illustrious and pony wrong all this time then

[shrewd gorge]

any new models that we are looking forward to?

not much progress it seems lol

[hazy quest]

Wth is ts

[atomic mortar]

Spambot, its gone now, was asleep lol

[atomic mortar]

Hmm theres been plenty of new models lately, wdym

[shrewd gorge]

What's one good one

[atomic mortar]

Z image has been impressive, anima is too. Flux klein has some great editing capabilities

Theres ltx with video & sound in one models now

Frankly im still rocking illustrious because i like how it works but all of them are pretty neat

[shrewd gorge]

Im a big fan of illustrious

its very aesthetic

and the prompt adherence and editing abilities are quite cool with the newer models but none of them replace illustrious

[atomic mortar]

You'll like anima too, the guys in #🍥｜anime have been using it

[shrewd gorge]

isnt anima a tiny model

[atomic mortar]

Hmm it still knows a whole lot and nothing a lora cant fiz

[shrewd gorge]

how is it an improvement over illustrious

[atomic mortar]

Prompt adherence

[atomic mortar]

Like
You cant get this on illustrious
#🍥｜anime message

[shrewd gorge]

looks like its still training or something

i will wait for final version perhaps

[atomic mortar]

Fair enough 😁

[shrewd gorge]

@atomic mortar anima looks intresting i think it has big potential

looks like its less system intensive than illustrious? dont know if thats a good or bad thing

mayeb better if it can use more gpu power

[atomic mortar]

Hmm it will load a text encoder on the side hogging up memory iirc

[shrewd gorge]

anyone using ACES here?

[tacit whale]

hello

gemini tells me SD4 is avaialble, but i cant find it on huggingface

did stable diffusion become proprietary?

[still glacier]

I m pretty sure gemini thinks 2+2=5 too. No, SD4 is not a thing.

[hexed crystal]

hello

[rotund pilot]

is there a good website for searching loras

except civit

[unique ravine]

If you know what u are looking for then huggingface, if not then civit

[iron pagoda]

yo guys, i really wanted to use stable diffusion webui automatic 1111 on my rtx 50 series, is there any way how?

[atomic mortar]

Dont bother with a1111, check for neoforge in #🤝｜tech-support pinned messages

Forge neo*

It looks similar and its actually supported for rtx50 series cards. A1111 is outdated

[iron pagoda]

oh, thank you

[opal coyote]

anyone know what node pack comes with beta_1_1?

[hexed crystal]

moin

[atomic mortar]

The scheduler? Sampler? Just the node name?

[violet juniper]

Hey I’m new here. I’m trying to make some video game ai art but chat gpt although their image generator is amazing and perfect they won’t let me create Mario stuff. Can I get a similar quality to ChatGPT with stable fusion without being restricted?

[atomic mortar]

Hey! You definitely can make some solid stuff with local ai but you'll run into some entry barriers like hardware, what GPU do you have?

[urban zodiac]

hello guys I am new here how do I use the AI tools?

[atomic mortar]

Dont fall for dm scams@urban zodiac
Do you want to run locally or from the cloud

Please dont drag people into dm's out of nowhere, its suspicious behaviour and its what scammers do a lot :-)

[prime hazel]

Hi

https://music.apple.com/ca/album/lost-in-my-own-sound/1525568065?i=1525568067

[warped vigil]

Knicks winning 1000000000000%

[violet juniper]

I have a MacBook Pro 2021, M1 Max 32gb memory

[rotund pilot]

Who is the owner

of this server

[warped vigil]

Idk

[unique ravine]

@still glacier

[rotund pilot]

ru an ai

[unique ravine]

@still glacier

@vapid dove ayo, sry, some major spammage was happening

[rotund pilot]

stop spamminf

finally

[unique ravine]

thanks 😄

[still glacier]

sorry about that, they're gone

[rotund pilot]

💋

[vapid dove]

all set

well not all set, Delete all messages on discord seems to be hitting a hiccup

[unique ravine]

They are evolving 💀 istg

[still glacier]

seems good on my side ?

[rotund pilot]

please add a spam bot

that prevents spamming

[atomic mortar]

I put my phone on mute for two hours
Was it a random bot spamming prompts again?

[unique ravine]

it was some guy which had ai auto reply to every question here, with very..uh idk the word, weird answers

idk the word when someone says a lot but doesnt say much

[atomic mortar]

Oh the average manager then

[still glacier]

I'm not sure it was even that. Some answers were very off topic. I didn't bother looking specifically but I'm gonna guess it just copy pasted messages from the past randomly and answered to random messages.

Probably trying to get rid of the "new account" status

(and underlying limitations, additional discord scrutiny)

[opal coyote]

its a scheduler

i assume its apart of a nodepack but i got a new pc and i have no clue what pack it was from since res4lyf only has beta57 for some reason 😭

[atomic mortar]

Things you could do

1. Set the normal beta scheduler and use both alpha and beta set on 1
   Or
   ComfyUI-ppm pack has beta_1_1

[opal coyote]

alr sick thanks

[atomic mortar]

No idea if the ppm pack is solid or not but i honestly dont see a need for that pack since comfy itself can co it already

[opal coyote]

alr i did have the ppm pack before because i recognize gits and ays thanks man

[violet juniper]

@atomic mortar are these specs ok ?

[atomic mortar]

a hey apologies, yes i remember the Draw Things App for mac in the store is the easiest to use

but those specs will be fine

[atomic mortar]

collab but its not thru discord or github , regardless rule 5.

and to be honest i do think its not the place for this type of stuff

[icy mountain]

What’s good y’all

[atomic mortar]

ai bot replying to a ai bot from MONTHS AGO, thats funny

[main pewter]

Hi guys, has anyone trained a LoRA using Kohya_ss? I have a few questions about it. I’m struggling to generate images that look very similar to the dataset I used for training. Trained lora few times still something wrong, not sure if it’s bad dataset or bad lora settings.

[main pewter]

Not yet mate. I would need someone to talk to and explain w little bit

Another scammer asking about the wallet to connect ?

Leave me alone you scammer, I will talk to someone here later on 😉

[icy mountain]

Why does bro keep spamming your name

[main pewter]

If someone could help

[atomic mortar]

Banned the scammer, sorry bout that. I was sleeping

[atomic mortar]

Personally I'm not too experienced with lora making locally but what i do recommend is joining the trainer discords, onetrainer is pretty similar and might have better advice

[main pewter]

Where I can find the link ?

[atomic mortar]

Theres on the GitHub of Onetrainer theres a discord icon
https://github.com/Nerogar/OneTrainer#onetrainer
(i have it but not yet really learned it well enough)

And as a reminder! We have a #🤝｜tech-support channel, we would never link you to suspicious "support" discords away from here or into dm's
Thats why i dont do direct invites and instead forward you to the GitHub (where there is a invite)

[wild forge]

Are there any YouTubers here who do manhwa/manga/comics/webtoon recap?

[atomic mortar]

Not that i know of

[icy mountain]

Yup

[icy mountain]

You probably have either bad captions or wrong training params. What settings did you use?

[copper salmon]

Hello

[icy mountain]

Yoo Wsg

[winter frigate]

Hey, I’m working on a text-to-image model for Minecraft-style 16x16 textures/items, and I’d like some advice on what architecture makes the most sense.

Right now my setup is a text-conditioned CVAE: CLIP text embeddings + category conditioning, generating small RGBA 16x16 textures. It trains and the loss goes down, but I’m not sure if this is the best architecture for this kind of tiny pixel-art generation.

I found this repo that already seems to work for similar 16x16 asset generation:
https://github.com/OVAWARE/BitRoss

Would you recommend sticking with a CVAE-style model, adapting something like BitRoss, or moving toward another architecture like VQ-VAE, diffusion, transformer/autoregressive tokens, etc.?

The goal is not photorealism, just clean Minecraft-like 16x16 textures with good shape, alpha, and prompt/category consistency. Any architecture advice would be appreciated.

[atomic mortar]

Sir, this is a Wendy's,
are you making a model from the ground up or are you just fine tuning an existing one

[winter frigate]

I would fine tune one if i could

[short holly]

Greetings, friends! 🤝

[atomic mortar]

Well then i recommend checking in with fine tuning communities instead, this is a great place to get started on using models however

[main pewter]

So, I have 61 photos&captions of Women. Using SDXL Juggernaut. LR - 0.00015 UNET LR - 0.00015 Text Encoder 0.000005 Dim 32 Alpha 16. I have RTX 5060 and this is what chatGPT advised me to set. Claude AI just said that those settings are to weak, and should put this higher. I have problem with eyes and some pose, I guess my dataset is not bad so i reckon it is kohyaa setting im doing something wrong

[still glacier]

If you went there, please leave their discord

it was a scammer bot

[main pewter]

I know, once they asked me about metamask wallet a knew its a scam

[still glacier]

glad they can't even be subtle about it

[main pewter]

Could you please look at my settings, you need some more information?

[still glacier]

not gonna lie I'm not good enough at training to help you on that one

Try asking in the onetrainer discord

people here should know more about it

[subtle plank]

does anyone use comfyui

i have this workflow atm but its soooo slow.

wanted to know if any one can help to produce the same images but much much faster

[atomic mortar]

Usually speed is decided by hardware
What GPU do you have and what model are you trying to run?

And how fast is it

[atomic mortar]

@pseudo parrot Not really the place for it so im gonna remove it but for millions of requests a day you might wanna look into setting up a google enterprise environment and not rely on discord strangers.

[steel fable]

stablle diffusion is already working or still need to use forge?

I just wanna know if back work thats all

[atomic mortar]

A1111 is outdated, in #🤝｜tech-support pinned messages theres a install guide for forge or forge neo depending on your hardware

[steel fable]

Can u send me the link for forge please?

[atomic mortar]

Pinned messages in #🤝｜tech-support

[steel fable]

Okay i ll take a look, thanks

[subtle plank]

i have a 4070 super

[atomic mortar]

What's the model you are trying to run and how long does it take?

[subtle plank]

i couldnt tell you tbh

im new to this

[atomic mortar]

How is the workflow called?

[subtle plank]

can i send you the json

[atomic mortar]

Post it in #🤝｜tech-support and ping me i suppose

[wild forge]

Are there any active YouTubers here? let's connect.

[gleaming ocean]

Hi

[swift vine]

Does anyone have a DoRA SDXL training workflow on TPU using Jax/Flax?

[atomic mortar]

@kindred plover rule 5, and tbh not relevant

[brave wigeon]

Hi, there

[atomic mortar]

hi hi

[brave wigeon]

Hi, @atomic mortar

[swift vine]

Awesome! Could you please share it? You can drop it here or DM me, whichever is easier for you.

[still glacier]

Do not get scammed in dms

[brittle slate]

There will be a day, bots trying to scam other bots. Imho this day is not far 🙂

[still glacier]

It happened already. Bots answering messages from bots

[brittle slate]

Interesting times we live in

[swift vine]

Thanks for the heads up, you're totally right. I actually just got baited into a fake 'support' server before this. The 'Admin' there literally asked for my crypt0 wallet address to 'scan and review' a JAX/Flax TPU error. I questioned their logic and they instantly banned me lol. Wild times.

[atomic mortar]

From Larry?

or was it a different user

[swift vine]

No, it wasn't Larry. A different user redirected me to a fake 'Support Desk' server. The 'Admin' there used a ticket system and asked for my decentralized wallet address, claiming they needed to run a 'blockchain integration' scan to debug the JAX/Flax TPU error. I asked them to point me to the official Google Cloud or Hugging Face documentation that mentions this requirement, and they banned me a second later.

[atomic mortar]

if you have a name for me i can ban them

[swift vine]

Unfortunately, I didn't catch their name. The message with the invite link disappeared shortly after they sent it, probably deleted, but I had already clicked to join. The person inside the scam server was named 'ADMIN[ NEVER DM FIRST]', but I'm not sure what the username of the bot who dropped the link here was

[atomic mortar]

ah yeah thats a known group, unfortuneatly they keep making new accounts

[thorny bloom]

What's best nsfw model for realistic?

[atomic mortar]

Cant help with nsfw but i found that z image is pretty good at realism but to get the best kind of realism you want to refine with a different model etc

[dull creek]

Hey everyone

I’m pretty new to the SDXL / AI creator space and I’m trying to learn how people create really realistic and consistent AI characters for social media content and cinematic videos.

Right now I’m experimenting with JuggernautXL, SDXL, Kling, face consistency,  prompting

I’d love to learn more about keeping the same character across generations, realistic skin / lighting, LoRA training, good workflows for beginners, making videos look less “AI generated”

If anyone has tutorials, workflows, YouTube channels or advice, I’d really appreciate it 🙏

[atomic mortar]

Yeah no thats scam behaviour, please dont do that

Nvm only 3 messages and all of them are of you asking people to dm you, bye

[unique ravine]

https://tenor.com/view/squirrel-disintegration-meme-ai-meme-funny-gif-5585270952676380091

fk man i wish gifs were visible 😢

[atomic mortar]

Rule 5, Sorry!@flint sedge

[flint sedge]

How was that rule 5? No advertising. Im looking for someone to work with from this community.

[atomic mortar]

professional work is advertising 👀

and we get such a high volume of job scams (and "DM me for more") i gotta be careful

[flint sedge]

Ok!

[swift vine]

Does anyone have updates on Google's TorchTPU? Is it already available for public use via pip, or is it still in a private repo? I'm looking for the official GitHub repo

[atomic mortar]

not yet. but during the Google cloud next event from what i could gather it was said that its probably next month or the month after that

[swift vine]

Ah got it, guess I'll stick to JAX/Flax for now. torch_xla lazy tensors are just brutal on CPU RAM (getting 47GB OOMs lol). Really looking forward to that eager fused feature

[icy mountain]

Good day to y’all wherever you are

[austere cosmos]

Over the past few weeks, I’ve been writing notes while learning/revising some core concepts behind modern generative AI and transformers.

Topics covered so far:

• Attention Mechanism: https://riteshrm.github.io/posts/attention-mechanism/

• Positional Encoding: https://riteshrm.github.io/posts/position-encoding/

• Rotary Position Embedding (RoPE): https://riteshrm.github.io/posts/rotary-position-embedding/

• Variational Autoencoders (VAEs): https://riteshrm.github.io/posts/variational-autoencoder/

These notes are based on papers, videos from creators such as Umar Jamil and Tushar Kumar, and some personal experiments while learning these concepts.

I initially wrote them for personal learning and revision, but I thought they might be useful to others as well.

Would love feedback and corrections.

[brittle slate]

Without citing, it looks just like a page made from Claude.

[austere cosmos]

thank you for your feedback
i will keep this in mind from next time onwards
although i have mentioned videos or links in current ones except Attention one

[brittle slate]

I am sorry, seeing it now, it sounds a bit rude. Actually i am reading a lot papers in this field and try to understand the things, so i clicked on it.

One original paper https://arxiv.org/pdf/1706.03762 for example.

[austere cosmos]

ok got it

[icy mountain]

What’s good y’all

[pale crypt]

What happened to fooocus? And anyone has alternative to easy inpainting?

[atomic mortar]

Rule 5. @high sparrow

[high sparrow]

yes bro i can help with that

[weary girder]

is Forge( or Neoforge) better than ComFyUI for these specs : 4gb VRAM and 8gb ram ?

[atomic mortar]

hmm generally not better or worse but i recommend checking in with Cs1o in #🤝｜tech-support , with 8gb ram its gonna be a extremely tight fit

[atomic mortar]

brandnew profile responding to a weeks old message that was already resolved. smells like a scam

[rapid osprey]

is there any free image generator server ?

[atomic mortar]

not here but there are some cloud services that offer limited free generations.

i knew a group that pooled money together to rent a 4090 24/7 but those are often closed communities only

[fiery ore]

You can send a dm

[ashen rune]

Post-human choreographic studies continue - https://www.youtube.com/watch?v=s7r-FGSEJ9o

[still glacier]

ok bot

[brittle slate]

[native heart]

guys what model is good one for edit images i been usign Flux 2 klein 9b there anything better than this?

[atomic mortar]

Please stop with the send a dm posts you tend to do it a lot, it looks like scam behaviour if you constantly try to isolate people asking for help

[atomic mortar]

Not at the moment imo, qwen edit is nice but flux klein works the best for me

[unique ravine]

U might as well get him out now, its a fresh discord account, plus he hasnt replied to you lol

'i CaN hElP wItH tHaT'

[vagrant agate]

HEEEYEEAAAAAAAAHH

[atomic mortar]

He actually did in dm, one more strike and he's out

[icy mountain]

What’s good y’all

[weary girder]

can someone recommend good bad hands embedding? there are lots of them and I am struggling to choose

[atomic mortar]

Embeddings arent really necessary to be used anymore since a long time, what model are you using?

[weary girder]

AOM3A3

[atomic mortar]

Oh wow thats ancient

Hmm i think any anime focused will do the trick but with a older model like that it will still do funky business

[weary girder]

what model do u recommend for low end specs?

for reference , pony v6 takes ~2 minutes for one 1024*1024 image

[atomic mortar]

Hmmmmm what specs do you have?

[weary girder]

rtx 2050 4gb , 8gb ram 🥀

[atomic mortar]

Yeah I'm afraid your using the recommended models already, that finetune was decent i remember using it

Sadly hands will be a challenge with that model

[weary girder]

the autismix was good , and the perfectworld was good

do you know the lightning model of illustrous?

[atomic mortar]

Oh yeah i remember those too, good times of chaos

The older models were creative in a way thats unmatched still today

[atomic mortar]

Personally no

[weary girder]

should i stick with these two then?

[atomic mortar]

Yeah

I remember CS1o telling me you can run sdxl/illustrious with 4gb/16gb but 8gb is probably not possible but i recommend asking him when hes around in #🤝｜tech-support

[weary girder]

i ran sdxl, usually one image without loras takes 1.5 -- 2 minutes

[atomic mortar]

He recommended to increase the pagefile but i cant say if it would help you

[rancid blade]

Hi, what is the best way to learn SDXL Illustrious for RunPod on ComfyUI?

Is there like a course or tutorials?

[atomic mortar]

I'm afraid not, but you might have more luck in the comfyUI specific discord for that

No, not really for neural companions? But since i see you are the uploader, #1092446741984444416 might be a better place

[modest island]

Thank you 👍

[steel hull]

👋

[wind sinew]

hey

[icy mountain]

Good day to everyone where ever y’all at

[brittle slate]

[sand plinth]

Hello

[late zealot]

👋

[rigid bridge]

hi

[atomic mortar]

Hi hi

[feral token]

Hey hello

[icy mountain]

What bussiness ?

[atomic mortar]

almost all business posters are bots that sadly join, i wouldnt engage with them 🫡

[subtle current]

Hi, nice to meet you.
I'm very beginner for AI and English but I'd like to see new members and knowledge in this community.

This is my first communication in Discord, so please tell me if I'm wrong

🐈

[icy mountain]

Cool

[native heart]

Anyone know how i can make flux 2 klein 9b have better match color on generation

always change lighting, colors, and contrast

[faint timber]

KJ has a color match node you can throw at the end before you save. It will attempt to realign colors with a reference image.

[pearl monolith]

hello

[icy mountain]

Try locking the lighting/color grading harder in the prompt + lowering CFG a bit. Flux tends to restyle contrast and lighting pretty aggressively.

[icy mountain]

Yoo

[steel ridge]

Hey @native heart, Yes I can help you with that

[atomic mortar]

Please don't try to lure them into a dm in order to sell a service

[native heart]

Looks like alot people do that here another guy did the same in my dm

[native heart]

Ok will try that

Thx man

[icy mountain]

You welcome

Let me know if you’re able to do it

[atomic mortar]

Got a name for me 👀 those are the kind of people we dont want here

[native heart]

Aax gaming the service price was 40US to help with workflow

[atomic mortar]

@tulip comet care to explain

[native heart]

Yeah that one

[atomic mortar]

Did kimmi also sell you something?

Or well tried

[native heart]

No only aax

Can show the print if u want

[atomic mortar]

Thanks for the report 🫡
Sure send in dm

[native heart]

Alright

[winter oxide]

А чо я тут забыл...

А привет всем

Hi everyone

[spare needle]

hi guys