💬｜general-chat

[floral umbra]

I can't seem to wrap my head around coding properly yet due to adhd, but i can grasp some decent basics about code. So i just yeet ideas into GPT, it spits out code, i throw at the wall and see what sticks lol

[atomic mortar]

It works surprisingly well for powershell & scripts

[floral umbra]

So far had gpt alter some comfyui's nodes to be faster, mmap cache models to load way faster, like from 300MB's stock to 4GB's, but nothing i can "sell any skills for", as it was all GPT's handywork lol

[floral umbra]

Lately it's actually been great! (if it was regarding gpt)

[atomic mortar]

Hmm if its selling the improved nodes i dont see a problem Personally

If someone commissions me a custom workflow i charge them by the time i put into it

[floral umbra]

Might need to elaborate, "if it's selling"? As in if it's gained popularity?

[atomic mortar]

Hmm well if you regularly optimize nodes you could post a few for free and some behind a patreon or paywall

Tho it goes against the spirit of opensource but a man gotta eat

[floral umbra]

Indeed. I'm on disability income, so if i gained money from this, it'd be a plus-income

[atomic mortar]

Hmm since you work with comfy you could sell custom workflows if thats something you can do

Sell the skill not the product so to speak

[floral umbra]

Indeed. Though, in my case it's more a "word wizard" for A.I, if that's even a skill, as i still can't wrap my head around large amount of info easily as others can.

You might see a beauitful lush forest, i struggle to even grasp the rest of the tree from the first branch.

[atomic mortar]

Ahh hmm in this case I'd imagine more of a workflow to make such images instead of the image it self

Though you could always try to sell t-shirts online with custom images lol

[floral umbra]

Nah, as A.I generated is not original art. It's existing people, styles and art thrown into a blender. So sadly i find it hard to want to take money for non-original work.

[atomic mortar]

Hmm it depends on the image i feel like

Prompt > sell is bad taste

[floral umbra]

Plus, i'm all for open-source stuff, so since i've no proper coding skills, i have no clue how to even make a job/freelance out of any of it

[atomic mortar]

Prompt > image > photoshop/edit > shirt

[atomic mortar]

Feels wayy better since theres work put into it

[floral umbra]

Where i can find it sellable of generated stuff, is use generated as illustration/inspiration, then make your own on top.

[atomic mortar]

Hmm could also work, or use ai as reference etc

[floral umbra]

Yep. Like the video game i am super slowly starting on, i need inspiration for the "character", so i generate images of illustrations of said character, find a few i like, then create character models in blender from there

[summer sedge]

When I use the Wan image-to-video example workflow in ComfyUI, the output has absolutely zero to do with my input image... Any idea why this might be happening?

Also... the LTX example workflows (https://github.com/Lightricks/ComfyUI-LTXVideo) refuse to load at all. I think I installed all the necessary custom nodes and models, but nothing.

[fast sage]

Anyone know the tech behind Kling AI?

[floral umbra]

I only know as much that it uses 3D Spatiotemporal Joint Attention, which gives generations their natural physics based results.

[fast sage]

yeah $4200/3 Months, I want to make a open source at that point.

[floral umbra]

It actually gave me an idea to "enhance" wan results by implementing a in-between node for comfyui which will use depth models like midas, and bake it into the latents. GPT is scrambling together the code as we speak

[fast sage]

I can help if you're open to sharing 😄

[floral umbra]

Oh, that branched to an idea where i'm gonna have GPT make a node that outputs not only regular video results, but a controlnet stick figure as well

[fast sage]

API prices like this make no sense, what is the difference between professional and standard?

[atomic mortar]

Professionals buy in bulk and probably get a discount etc

[fast sage]

https://app.klingai.com/global/dev/document-api/productBilling/prePaidResourcePackage

[floral umbra]

Of course! As i'm gonna throw the code over to my git Sadly, i can't code for shit yet, just some basic knowledge if there's a red because a syntax error, i can somewhat fix those x)

[fast sage]

Token costs and result are nowhere to be found.

[fast sage]

I'm a SE, 20+ years.

[fast sage]

Ping me, let me know how it goes or if you need python help.

I haven't had good results with WAN yet.

[floral umbra]

I've had quite a few nice results

With basic videos that is.

[floral umbra]

Added ye so we can take it to dm's or something :P

[barren dust]

Hi guys,

I want to create a gif or a video thumbnail that has static picture with the information of my song with moving sound wave across the picture.

I tried videos by Sora, but it doesn't work. it doesn't keep static image on the background. Images also do not work, because sound frequencies waves do not move

I would really appreciate any info that could point me in the right direction. Maybe any other more simple tool then SD?

Thanks a lot

[atomic mortar]

Bro again?

I honestly recommend looking on LinkedIn or fiverr man, this isnt working

Its seriously clogging up gen chat

Im not short of money but i recommend looking on places where businesses look like the one i posted above

Almost all people here are for image generation etc

[main snow]

true

[fast sage]

you can easily do this with python and a frequency wave form.

[barren dust]

You mean python the programming language? 🙂

[fast sage]

What else would there be?

You're just converting frequency to a image, any language can do that easily.

This wouldn't be a task for ai

[upper plinth]

Is anyone using dreambooth with a 50 series GPU? If so, can you tell me how?

[barren dust]

I don't know, like real python for example :). I'm equally scared and inexperienced in handling both of them :)))

Sorry, maybe I didn't make myself clear. I meant to say that I don't need to convert the sound wave frequency into an image.

I need to have static image and running soundwave frequency on top of the static image. I don't know if it's possible at all with AI tools like Sora or filmora or SD

Here's the simple example of what I'm referring to. It is YouTube thumbnail for a song.

https://youtu.be/VqAj8yL2lOA?si=sI71ELV539eKVeY_

Thought I think this thumbnail is a video thumbnail, not .gif

[fast sage]

This has no real application in AI, generating soundwaves is just a series of images (gif) of said soundwave, you would capture and export that using a program or your own written app.

You can generate the background image with SD no problem.

[barren dust]

Thank you, however, at this point I'm not looking to do any custom coding myself.

I don't think all the musicians that post that type of thumbnails on YouTube are coders either .

I think I'm looking for more simple solution than coding.

[upper plinth]

Have ChatGPT code for you G

Humans werent made for coding. That's what AI is for

I made a game in Python with me being the architect and ChatGPT being the translator from English to Python. So far, so good. o4-mini-high is insanely good at coding

[steel prawn]

He gets it

[desert dagger]

@vapid dove fairly well known scam

[median dome]

Hi. My name is Aki and im from Austria. Im pretty new to ai. Nice to meet you.

[rancid crane]

Hello all

Is this group more tech or art

[craggy rune]

general with images is more art this is just general

[desert dagger]

thsi discord is mostly art

[wary tangle]

yeah, thats what im wondering, i just joined, been playing with automatic1111

ive kinda just been playing around with whatever ai i can get my hands on

[wary tangle]

a snake

sorry, couldnt help myself

[atomic mortar]

Mix of both kinda but tech support gets the most traffic

[wary tangle]

ah, cool, i honestly dont know what this server is, i kinda just stumbled on it through the discover tab

[atomic mortar]

Ahh its the official stable diffusion community server

[wary tangle]

do people generate images in this server or do they do them local and then upload them?

[atomic mortar]

Hmm people in #artisan-faq do it here but most do it locally

I prefer local due more control & freedom

[wary tangle]

i prefer local so i know what im doing when the elites cut us off xD

[atomic mortar]

Lmao, that and the only costs i have (after investing initially) is my electricity bill

[wary tangle]

true

biggest pain in my ass is im using an amd gpu...

amd doesnt even try foreal

but they cheap

[atomic mortar]

Ahh rip, hmm amd is close to getting native pytorch support

So once thats around youd get a massive performance boost

[wary tangle]

they been saying that for years tho

[atomic mortar]

Saw screenshots

Of it working but "unstable"

[wary tangle]

whats crazy is, i couldnt get automatic1111 to run on my linux os, i got it to run in windows

[atomic mortar]

I mean fair since a1111 is outdated by a year+

[wary tangle]

what is the better option now and is it free?

[atomic mortar]

Hmm for amd forge is better for sdxl

[wary tangle]

whats great is i can use ai to help me set up ai xD

[atomic mortar]

I 10000% do not recommend that

But yeah in techsupport theres a guide for amd cards forge written by Cs1o

[wary tangle]

yeah, youre probably right

it was walking me in circles

i kinda just figured it out myself

[atomic mortar]

Personally i use SwarmUI but for amd you'd need to change the backend manually to use zluda

[wary tangle]

i ended up just running the bat file normal, uninstalling the torch files, and installing torch-directml

[atomic mortar]

Ohh directml while it works

Its sorta takes up more vram

But if you dont mind slower gen speeds, swarmUI is peak

[wary tangle]

directml? doesnt sd run on top of that?

iono, i feel like it had to have that installed to get it to work

oh well, thanks for the info, im not too serious with it atm, i kind of just wanted to see what it could do

[atomic mortar]

Fair fair

[wary tangle]

the models i was using didnt produce anything good

and even tho i have 16gb of vram, i didnt seem to be enough for more intensive stuff

[atomic mortar]

Yeah with amd you can't really run anything above sdxl

And directml uses a lot of vram on its own

With my 16gb nvidia i have 5s gens in sdxl

[wary tangle]

oof, i would kill for an nvidia card

or for amd to get off their ass

oh well, i guess its just not their market

their aim is more, selling mid range cards and undercuting prices as little as they can get away with xD

they pretty good at gaming tho

[ancient mauve]

does anyone know how to add eta noise into comfyUI?

[warm junco]

Oh rip. Dont use directml

With a 16gb vram GPU and zluda you will have way more fun than on directml.
Checkout the install Guides for Zluda.
First link in the pinned messages of #🤝｜tech-support

[rancid crane]

Hmm you could rent an H100 for $1 an hour

https://app.hyperbolic.xyz/?utm_source=homepage

I almost got Wan 2.1 14b to work on my 5090 but I was too dumb to write my own Cuda kernels

[fervent thunder]

lol I am also struggling to write cuda kernels

H100 for $1 an hour is not gonna be high volume its not profitable price

[atomic mortar]

?? 14b should be easyly possible on a 5090

I run it on my 5080

[silent narwhal]

guys how to generate figure 1 in Midjourney in the style of figure 2，thank uuuuu

[empty pond]

🚀 Seeking a sales & marketing automation specialist!

I need an expert in AI-driven workflows to accelerate lead generation, outreach and campaign analytics. While many of us build multipurpose agents (myself included), if crafting scalable, data-driven specific to sales funnels and marketing optimizations is your passion, let’s connect! (and yes this was written by ChatGPT)

[ancient mauve]

how do I add eta noise seed delta in comfyUI?

[atomic mortar]

gotta enjoy spambots

@ancient mauve have you tried asking in the comfyUI discord or other more specialized discords? Like in L3

[ancient mauve]

Im asking in various servers and no one is answering

im also seaching in google and everywhere and nothing

[atomic mortar]

mind me sending one where you might have any luck?

[ancient mauve]

Im trying to recreate an image and the metadata says it has ENSD 13337

[atomic mortar]

and no its not a spambot this time lmao

hmmm

i remember that value from forge

[ancient mauve]

yes in forge is easy to set up

but not so in comfyUI

[atomic mortar]

ahh i think i found something

https://www.reddit.com/r/StableDiffusion/comments/1381zfe/what_is_ensd_and_why_is_being_used_by_a_lot_of/

it just a seed obfuscation

[ancient mauve]

so its just a number added to the seed?

[atomic mortar]

i assume so from that post

[warm junco]

It just makes the seed more random as it already is

[ancient mauve]

but I know both the seed and the ensd

[warm junco]

Try recreate an other image without an esdn if you only do this to test if you can get the same results then

You won't even get the same image if you use a different webui anyway

Also using different GPUs (nvidia, amd, intel, cpu) makes also different outputs.

[odd matrix]

Does comfy UI work better enough to justify the instalation and the learning curve?

[atomic mortar]

Try swarmUI, you can try comfyUI whenever but its optional as there's a handy GUI

Ui***

It will automatically make a workflow in the background depending on the settings you change so you don't "have to" use comfy but still have all of its benefits

Be sure to not accidentally install stableswarmui though

As its the old old version

[odd matrix]

I have installed Automatic1111 and Forge. they work, but there are things that are "Broken"

[atomic mortar]

Are you on amd by any chance

[frank river]

Can anyone help me understand, i was told to check out Hyperlora for the best consistent character workflow. But it seems to use a reference image for control net, but i dont want that - i only want the face to be needed + a text prompt to describe the scene. Any advice?

[round mirage]

anyone experience with comfyui and wan img to vid? When I try I get rave fest lol :D

[oblique elk]

Works pretty good for me (comfyui) which WAN I2V do you use?

[round mirage]

I downloaded wan2.1flf2v 720p 14B fp16, I have no idea what settings I need. I tried 1024x1024 but I seem to take forever now to generate

[oblique elk]

did you tried the example workflow which you could reach in the menu (Workflow > browse templates > video > wan image to video?

[round mirage]

Yesnt, I downloaded 720p instead of 480p but I essentially didnt change anything besides sampler, and I also put image gen in front of it instead of providing one

[oblique elk]

and the generated image is "clear" not too much noise and not too much objects in it?

[round mirage]

its basically just jittering around and stretches into oblivion with a lot of rave like colors :D

[oblique elk]

just tried it with the basic workflow and it worked flawless.

[round mirage]

I will try that too then

Did you install the default files or "upgraded"?

[oblique elk]

I tried these model to confirm the example workflow is working:
wan2.1 480p and wan2.1 720p both FP8

The umt5_xxl Clip

and the wan.vae

[round mirage]

Okay thank you, will try that in a bit!

[round mirage]

Okay I have now everything, do I need to edit the prompts since I dont really have an image yet? Do I adjust them to fit the image I insert?

[floral umbra]

Oh, that's way too big, as unless you got a 48GB+ gpu, and a latest 5000 series quadro, it'll take an hour xD

Try starting at 480x640 with 41-85 frames

[rancid crane]

I just barely got Wan 720p to work on a 5090. I quantized just the linear layers. And it worked for the t2v, but ran out of memory for i2v

Exactly 1 hour, but down to 40 min when I used sage attention

[odd matrix]

Has anyone managed to get characters with different poses? I keep trying but it never works.

[round mirage]

nah only got 4080 super :D

[odd matrix]

No intel , with nvidia 12 gb VRAM

[atomic mortar]

Ah what are the things that are "broken" then?

[swift fulcrum]

Can someone help me with Stable Audio's input audio upload? I am trying to upload audios but they all just fail. One succeeded and the others are the exact same files as this one, but they fail. It just says "Contact support"

[floral umbra]

Depends on the base model, some further trained base models has been fine tuned/further trained with better pose controls.

Alternatively, you can use the good old stick figure with controlnet :P

[odd matrix]

Layer diffusion and other things

[atomic mortar]

Ah extensions, Makes sense if not maintened

[pine path]

An important paper just dropped
https://arxiv.org/abs/2505.10046

"Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis"

Focus on 6.1 which finds timestep conditioning to be a waste of parameters

[cobalt geyser]

Hi, I’m learning to use Stable Diffusion for the first time. I’m trying to use Controlnet for posing the image, but it seems like most of the checkpoints (even that’s a new word for me) I want to use are too “stylized” to listen to Controlnet’s Openpose.
I don’t actually know for sure what the problem is.
Do you have any advice or workarounds for this?

[oblique elk]

Trying my best to understand your question. One of the controlnet available is an openpose controlnet. To make it work you would need a special open pose image (stickman in different colors for limbs etc.) or use a preprocesser to extract the pose from a source image (again will be convertet to the open pose model). Then just write a prompt incl. a character and the character should take the pose if the strenght of the controlnet is high enough.

[cobalt geyser]

Thank you.
I use a 3D modeling program to make the pose I’m looking for, then I upload that backgroundless png to Controlnet. I use the openpose preprocessor to process the image, and it shows me a preview that the stick figure took the pose and understands what I am trying to do. Then, just to see if it’s working, I set controlnet’s priority to 2, the highest possible strength I see, and then select the box that says “Controlnet is more important” then prompts.
Then I add minimal written prompts.

When I generate the image, it makes a picture without the pose I gave it.
Also, the image I gave Controlnet keeps saying “start drawing” over it despite it recognizing the pose I assigned it.

It seems that some checkpoints are listening to Controlnet, while most of the ones I’ve tried are not responding at all. I hope there is a way to still use Controlnet since I like the art styles of the ones that aren’t working.

[oblique elk]

could you share the pose (output from the preprocessor in general with image?

[round mirage]

So the pretty good animated images or videos on civic are made with banger hardware I assume? Since whenever I try to make it work its so wonky

[oblique elk]

Video minimum requirements (if not endless patience available) would be 16gb vram

[round mirage]

Yes im on 4080 super which has 16gb I believe but still not super sure what Im doing wrong, maybe my prompts are just bad

[oblique elk]

What model are you using.

[round mirage]

loaded the default

or template I mean

[oblique elk]

WAN? Huan? ITX

[atomic mortar]

Are you running wan, hunyuan, ltx

[round mirage]

WAN

[atomic mortar]

Or the ancient tech animatediff

Wan just released a new thing where gens get sped up massively

[round mirage]

480p 14B fp16

[atomic mortar]

Things previously took like 30min on a 4090 is now 3min or something with wan

[round mirage]

generation takes about 5-10 minutes but its just off

[atomic mortar]

Are you trying to animate something real or 2d?

[round mirage]

2d

[atomic mortar]

Thats the problem lol

[round mirage]

Oh

I thought it wouldnt matter, how wrong I was :D

[atomic mortar]

Its not great at it

Technically it works

[round mirage]

I want an animated/cartoon cat walking/bouncing from left to right and its just raving lol :D

[atomic mortar]

Iirc theres another group working on animating anime but its not ready yet

Previews while great were wonky lol

[round mirage]

I wonder how its done on civicai website then, seen some incredible stuff

[atomic mortar]

Trail and error probably, could keep trying but with our hardware it takes a while

You got the original negative prompt?

With all the Chinese in it?

[spiral pilot]

Hey everyone how's it going? I'm trying to do a simple animation in stable diffusion with an existing image using img2img and animate diff but im ending up with the picture basically just tearing apart. Wondering if anyone has some tips or can point me to a basic setup to achieve this. It's basically just a still image and i just want evry minimal animation like camera movement or breathing.

[atomic mortar]

Animate diff is ancient

[spiral pilot]

Great to know, how would you recommend setting this up?

[atomic mortar]

What gpu do you have?

[spiral pilot]

rtx 3060

[round mirage]

So if I wanted to make the cat work, I rather try to animate a real one?

[atomic mortar]

Hmm im not sure how to get animate diff to work since its been a while sorry

[spiral pilot]

thats fine im not married to it

literally just trying to animate my still images. 5 second animations, nothing fancy

[atomic mortar]

Better chances yeah, and people use tricks to get longer videos like extending from the last frame etf

Etc

[round mirage]

ahh neat, thankers!

[atomic mortar]

Your using comfyUI?

[round mirage]

A1111 and comfyui, I have both installed

[spiral pilot]

i don't really care about the tool stack or anything but for example i have a guy standing there i just want some motion, maybe it looks like he's breathing or something

[round mirage]

started with a1111 then switched to comfyui cuz I like the nodes

[atomic mortar]

I'm not sure how to realistically archive that within a reasonable time frame with an 3060, your best bet could be #📝｜prompting-help or #🤝｜tech-support if its not behaving as intended

[atomic mortar]

You could combine them and use swarmUI

Hella discussions going on with wan and its new light speed lora

[round mirage]

I see, I will look into it the coming days. I mostly do this out of curiosity anyway so there is no rush :P

thanks

[atomic mortar]

Bet, if you want i could dm you an invite so you can peek into the server since swarmUI has a comfyui backend

So you can use comfy whenever and have the benefit of most models working out of the box without issue & documentation

[round mirage]

sure sounds cool :D

[atomic mortar]

Just doesn't support comfymanager that well

[spiral pilot]

are you saying a 3060 is not capable of generating a 24 frame animation in a couple minutes or ?

[atomic mortar]

My 5080 takes 10 minutes

[spiral pilot]

for how long

[atomic mortar]

So while im unsure with animate diff

The newer stuff might take a long time

[spiral pilot]

what is the newer stuff?

[round mirage]

which image gen/checkpoint would you recommend that I can use with wan after?

[atomic mortar]

Any really

[round mirage]

Aight, tyty

[atomic mortar]

Traveling home, you could ask in the swarm one since they know a lot more then me immusama

[round mirage]

no worries, thank you!

[spiral pilot]

i downloaded a release of forge but no stable diffusion video https://github.com/lllyasviel/stable-diffusion-webui-forge/releases/tag/latest

[dusty zealot]

Does anyone know what to use to get the following.https://youtube.com/shorts/EpzprZuWEwI?si=6ro8nkmvIvk2CAls

[cobalt geyser]

Sure, if I may tag you. I don't think I can do it here.

[deep narwhal]

Is there a list of stablediffusion benchmarks for the 50 series cards yet?

[rigid warren]

hello

i'm tryng to train a lora style model

in the tags part, should I only focus on the style or do I have to tag the character appearance as well so the AI can know what is going on?

[odd matrix]

Does anyone have a good prompt for generating a solid white background with no shadows? (For creating transparent images)

[oblique stag]

Hello

[rugged pagoda]

pig

flow

[undone bloom]

hello

[daring iris]

i ask ai to give me a promtp and here waht it said

"Generate an image of [your subject] placed against a perfectly solid white background, completely uniform and without any shadows or shading. Ensure the background is pure white with no texture, gradients, or lighting effects."

[native steppe]

Hello Adrian here, aka Sm0ke

[floral umbra]

I tried to fetch civitai's entire library of lora's with a script someone made that made the SQL DB of all the models, tried to fetch all of SDXL lora's specifically for archiviung..

Well.. I ran out of storage on my C drive.. Reached 600GB ish worth of models. And i got this far lol
SDXL 1.0: 4%|█████████▎

Moving them to my server to continue there lol. Though slower because spinning rust.

Had GPT make the script that reads from the model database, and download all relevant images, all versions of a model, sadly it also fetched non relevant ones too, like FLUX and SD1.5 versions of the same lora, so had to manually "post process" them by searching for any .safetensors with pony, flux, sd1.5 and so on in the name

[atomic mortar]

You only need 15tb for sdxl then nice

[floral umbra]

Depends on the remaining models, if it's only 1 model, or 7 versions, as well as if the model was trained with high rank, or low rank :P Low as in rank 16 will leave it at a few 100MB, 128 rank? 1.2GB per model :P

Got 20+TB free on my server atm, and more freeing up as i'm transcoding all my movies and shows to av1 xD

[atomic mortar]

Hmm if sdxl 4% is 600 gb then 100% is 15000gb avg but then again idk how that script tracks progress

[floral umbra]

SDXL 1.0 | Total: 40470

• Refiner : 1 models (0.0%)
• Standard : 805 models (0.17%)
• Unknown : 39664 models (8.51%)
  Other | Total: 6843
• Inpainting : 2 models (0.0%)
• Standard : 4553 models (0.98%)
• Unknown : 2288 models (0.49%)
  NoobAI | Total: 2458
• Standard : 4 models (0.0%)
• Unknown : 2454 models (0.53%)
  Hunyuan Video | Total: 1473
• Standard : 8 models (0.0%)
• Unknown : 1465 models (0.31%)
  Wan Video | Total: 604
• Standard : 4 models (0.0%)
• Unknown : 600 models (0.13%)
  SD 2.1 768 | Total: 567
• Standard : 556 models (0.12%)
• Unknown : 11 models (0.0%)
  Flux.1 S | Total: 504
• Standard : 12 models (0.0%)
• Unknown : 492 models (0.11%)
  SD 1.4 | Total: 300
• Standard : 228 models (0.05%)
• Unknown : 72 models (0.02%)

---

It tracks progress based on model count, not sizes. Wish it did tho.

Here's percentage of models that civitai had as of a week and a half ish ago.

[drifting dirge]

Hey! Would anyone be kind enough to share a working ComfyUI workflow package for cinematic / realistic scenes (like Netflix-style)?
I'm working on viral hook content and struggling to get clean outputs.
A working .json or folder setup would help me a lot. Thanks in advance!

[hallow plaza]

Which version of python is everyone using

[gentle prairie]

if i want to do more than 3 characters, am i forced to use regional prompter? Like when doing 4 characters, instead of picking the 4 of them, he will do 3 and the 4th one will be a repeat of one of the ohters

[drifting dirge]

I'm using Python 3.10.6 (with ComfyUI) — no issues so far.

[floral umbra]

3.12 for comfyui, but any other/older version for other python programs written for older python.

[atomic mortar]

using 3.13 but i 1000% advice against it as only comfy will work but any other nodes will break

unless you patch them

[floral umbra]

Yep. I keep a few different comfyui venvs based on what i use them for. Stableaudiotools iirc needed 3.10.

[quartz siren]

Use ltxv, it can easily gen in a minute on a 3060. Although, wan2.1 is probably still sota quality, ltxv is still decently close and very very fast. Both are far better then svd.

[spiral pilot]

https://github.com/Lightricks/LTX-Video or something else

[stray grotto]

heyyyy

[primal hemlock]

yo new here is there a guide ?

ty

[median dome]

Hi. anybody on?

[atomic mortar]

If you clicked that link and logged in you got scammed

Its a common scam here

You better change passwords

If you gave them anything

[atomic mortar]

Any guide here is found in #🤝｜tech-support pinned messages

[median dome]

is there a guide on prompting? i cant figure out how to link keywords.

[primal hemlock]

I realized it after I joined the server and left immediately so np ig

[primal hemlock]

ty

[atomic mortar]

What model? Xl anime style?

[median dome]

waiNSFWIllustrious its the only model i got to work

i tried making different girls with different hair and skin color. is there a way to link those to one particular girl without it being randomly distributed?

[atomic mortar]

Ah, then you could reference danbooru for tags or do you mean something else in tagging

Ahhh

Thats mostly luck

Or regional prompting

But I'm not an expert on that since i mostly make 1 girl or man women combos

[median dome]

so which are your favourite models to use?

[median dome]

they are the only ones that work for me ^^

[steel prawn]

Illustrious and NoobAI are both solid for anime ive found. But every model has its use case.

[steel prawn]

ive never seen a more obvious scam link in my life lol

[floral umbra]

Oh boi, i'm having a "blast" not attempting to have GPT try to make a node that loads sd 1.5 model lora's, be parsed as if they were sdxl models, and have if any blocks the sd 1.5 lora has to be applied with SDXL So far fuckall results, but i'm still gonna see if it's any way possible lol

[abstract quarry]

that doesn't make any sense

[oblique elk]

Why not use a sd1.5 model with the Lora followed by a sdxl image2image run?

[abstract quarry]

you can load the clip l part for 1.5 loras. Besides that, sdxl and sd 1.5 are incompatible

[atomic mortar]

while it doesnt make much sense to me since i dont know the inner workings of xl and 1.5 but it does sound like your trying to run a diesel car on cola

[floral umbra]

Nope. Idea is to find out if I can make a 1.5 Lora to be loaded with a xl gen

[floral umbra]

More like trying to make diesel work with gasoline x)

Both are combustible, but combusts under different environments.

[autumn onyx]

Testing new A.I chat with Darth Vader in Fortnite https://youtu.be/r37wXl-i6Ik

[abstract quarry]

you don't have to test that. It won't work.

[round siren]

Is anyone here good with photomaker v2 in stable diffusion?

[fervent thunder]

.

info on my profile

[modern basin]

hello

im new here

[desert dagger]

Nvidia GTC Taipia 2025 keynote should be starting in a bit https://www.youtube.com/watch?v=TLzna9__DnI the live pregame show just ended

[oblique agate]

https://github.com/smthemex/ComfyUI_YuE?tab=readme-ov-file I don't understand this. Yue is for making music how is it link to comfy ui workflow

[stray grotto]

Hey, I want to create raw footages like the one in this video https://www.youtube.com/watch?v=QIQYzi7IPGk. I can act out myself but change the background and the face with of someone else's. what all softwares do you suggest me to use

[stray grotto]

what's comfy ui?

like a multipurpose board to plug in different models?

[oblique agate]

comfy ui is just workflow management program mostly for image diffusion

[stray grotto]

Can I make a video using that?

will try it out thanks

[oblique agate]

I think so

[stray grotto]

PLEASEEE

[atomic mortar]

What gpu do you have?

The background is easy as its a single image but the faceswap in video live takes some capacity

Post process too

[dusty zealot]

Anyone have any good models for sd sdnext or even site that can make the talking baby videos

[vagrant basin]

Hi, I’m thinking about saving up and investing into a new computer for making AI art, particularly comics like a Webtoon. Right now I’m working with an NVIDIA RTX 2080 Super and I must have did something wrong because generating one picture took up to an hour a couple times, and the computer froze. Usually it’s 2 or 3 minutes, that’s okay, but I add enough LoRAs and controlnet influences and it’s a cry for help.
And that’s with Xformers and LowVRam.
I newly installed Automatic1111, and I hear good things about ReForge, so I’ll try that soon.

Can I ask what you guys are using, and what you would recommend? This would mean so much to me, seriously.

[still glacier]

Why use lowvram with a lot 2080 super ?

[vagrant basin]

I’m still learning how the program works, so I thought that helped and might have been mistaken.

[lean field]

vram and procesor is key for good image and fast image generation

[still glacier]

by using lowvram the program is splitting the model into multiple parts, loading and unloading them into the gpu whenever they're needed. Thus wasting a load of time when you could simply load the model at once.

2080 super are 8gb so --medvram-sdxl should be the right spot.

aka using --medvram only for sdxl model

[vagrant basin]

I use the checkpoint WaiNSFW with most of my initial attempts since I’m getting a lot of the art style I want from it, the kind that looks like a screenshot of an anime. I hear that model, on top of Noob Controlnet, and hires fix, and using multiple LoRAs if I’m using a style or more than one character, can be very demanding. I figured getting a more advanced computer would help with the generating time.

And considering I’m hoping to make enough images to make a webtoon-style comic, generation time and the computer not freezing is really important.

I just want to get this right, so I really appreciate your insight on this. Gaming specs make sense, but shopping with AI in mind is entirely new territory for me.

[still glacier]

Not gonna discuss anything NSFW in here, but yes you must have done something wrong because it's not supposed to take 1h to generate one image with a 2080 and without any logs / screenshot of your settings / etc I can't properly help you. Also that would probably more of a #🤝｜tech-support kind of discussion

[thorn hare]

so is Auto1111 support dead ?

[pseudo nexus]

Hi folks, maybe it's a question already asked hundred times but can't find a good answer. I want to have AI generated role play heroes. How to generate images using this character? That means I want to retain the silhouette/face/clothes of given character

[stoic onyx]

can anyone help, how can i save the prompts for image generated through flux in forge ui in csv format

[echo ice]

hello

[atomic mortar]

Yes, its been outdated for over a year

But you can always ask questions about its use in older models

[compact lantern]

Alright guys question, my buddy said he would be willing to swap HIS 4070(not to or super) just a standard 4070 for my 7800xt, I do work in stable diffusion and was wondering would it be worth it to swap? What would I gain in terms of it/s and what would I loose (going from 16gb vram to 12gb)

[atomic mortar]

Going from amd to Nvidia you'd get a performance boost by not using to use zluda or directml, you could upscale etc

[atomic mortar]

How long does a gen take you now in sdxl?

[compact lantern]

A 1024^2, 25 sample steps takes about 30s

I'm also wondering if the native cuda support will make it run better as I was just getting into animatediff and it's workability is to be desired or I could just be bad at using animatediff

With AMD that is

[atomic mortar]

Well you could expect faster times thats for sure, and you could probably run flux schnell

[compact lantern]

And I'm assuming the native stable diffusion works better than automatic1111

[atomic mortar]

What

[atomic mortar]

Just use swarmUI, forge or comfy

A1111 is outdated and i have no idea what you mean by native stable diffusion

[compact lantern]

What o meant is you don't have to jump through all the hoops with nivida to get it to work like you do with amd

[wary tangle]

this, the first step for installing a1111 with zluda is update your amd drivers... turns out the newest drivers dont work with zluda...

[hot raptor]

bilibili's open-sourced this anime video gen model called Anisora. The V2 version, which runs on wan2.1, kills it for anime videos—like, it's really good. But right now, it doesn't work with ComfyUI. I'm really hoping someone can make it compatible!
thanks!!
GitHub address:https://github.com/bilibili/Index-anisora

[woeful anchor]

Hello everyone, I'm a newbie actively learning Stable Diffusion (SD). Your guidance and support would be greatly appreciated!

[median dome]

Good morning.

[stray grotto]

i have gt 610 at the moment but can upgrade if that's necessary. Not live actually, dude we actually edited the same video for simon, but what we are trying to achieve is a similar quality video in which he only gives us the content and we make the whole video from scratch. Do you think its possible? Most of the people in our industry are using just elevenlabs and heygen right now but the video is easily spottable as one made with ai

[atomic mortar]

Scam

[atomic mortar]

Theres a github somehwere that was convincing quality but its beyond me

[stray grotto]

check this guy out for example https://www.instagram.com/thevarunmayya/?hl=en

people call him the pioneer lol

[stray grotto]

how do you find such repos?

[atomic mortar]

Eh just Reddit or Google personally

[stray grotto]

bro trust me I have been on it for quite some time

everyone calls themselves the best until you see the results

its like finding a needle in a haystack rn

[vagrant basin]

Sorry, WaiNSFW is just the name of the checkpoint, not that you have to do NSFW with it.

[merry ginkgo]

So with civitai shutting down does anyone know the status of the mirror/torrent tracker?

[lofty tinsel]

سلام عليكم فيد حدا يساعدني بتصميم شخصية بذكاء الصطناعي نفس هاي

[fathom portal]

Hello !

[thorny bloom]

Any lora of the meme tung tung sahur ballerina capucina?

[thorny bloom]

There is only one lora and is not all the characters

[atomic mortar]

If the biggest lora site doesn't have it i suppose your outta luck

[oblique agate]

for generation, my cpu rarely use up more than 15% of the resources

I am rocking an amd ryzen 5600

[atomic mortar]

Some shameless ones that copy the ones from civitai without permission of the lora owners yea

It has a few like tensorart

But compared to civitAI its pretty bad

Once i get onetrainer working properly I won't use public loras anymore

[valid aurora]

anyone good at python. i have a basic python homework and need help

[chrome snow]

Hello!
Could someone please share a working download link for the SD 2.1 motion_adapter.safetensors file (v1.1, ~1.8GB)?
I’m looking for the official AnimateDiff v1.1 adapter (not SDXL or Lightning), SHA256:
5fd196988206cfac90bab1dc56cc59fdd12035492f74ff0662eb6eed5113c547
or
45b71fe98efe5f530b825dce6f5049d738e9c16869f10be4370ab81a9912d4a6

The official Hugging Face repo is down/private. If you have a Google Drive, Dropbox, or other trusted mirror, I’d really appreciate it!

(Please, no SDXL or Lightning adapters, only SD 2.1 v1.1!)

Thanks so much in advance! 🙏

[gilded swallow]

Hello! I'm trying to learn Realistic Vision v6 and tuning. Currently images are still seeing a lot of noise with the following configuration. I'm wondering if anyone had some advice. Also I've seen recommendations for denoising steps/strength configuration, but I haven't seen those in the UI...

Interface: WebUI
model: realisticVisionV60B1_v51HyperVAE.safetensors (models/Stable-diffusion)
Upscaler: 4x-UltraSharp.pth (I added this to models/ESRGAN folder, but not sure if that's right. Also unsure of how to properly configure upscale)
Sampling method: DPM++ SDE
Scheduling Type: Karras
Sampling steps: 50
CFG Scale: 7
dimensions: 896x896

prompt:fashion photography, cover photo, black and white, female model, facing the camera, shot on Sony A7 III, photorealistic, close-up, monochromatic, flawless, ultra-detailed, hyperrealism --no freckles --chaos 75 --ar 1:2 --style raw --profile rsy6uq9 --stylize 750
negative:(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime), text, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck

Thanks!

[bright tinsel]

why with the latest version of the request package, it prints ID not 64 charactersw

I think it the code needs to be upgraded

I have the requests version 2.32.3

[desert dagger]

@vapid dove

[atomic mortar]

We really needs more mods in here ong

[floral umbra]

Making a personal archive of civitai, and it might take a while lol. Seeing as sadly the script GPT made to fetc them, also fetches irellevant models.
SD 1.5: 1%|▉ | 1636/111713 [00:00<00:54, 2028.34model/s]

111k models, but a lot of them are irellevant models like sdxl, flux and the like, so have to use lora model manager to weed out the non SD 1.5 models now and then lol

Wan video has like 3-400GB of loras, same with hunyuan, so it'll be a few TB for the base model lora's i need lol

[past hemlock]

Hello guys, I am p new to the idea of fine tuning diffusion models, is there any resources y’all would recommend for me to look at, to get a basic understanding of how to train a model to get my desired output

Thank you :D

[fervent thunder]

i cant seem to install stable diffusion 😦 problem after problem currently cant seem to install CLIP

[atomic mortar]

What ui are you using?

[fervent thunder]

stable-diffusion-webui

[atomic mortar]

Ahh a1111 outdated stuff if I'm being honest

I recommend forge or swarmUI/comfyui, in the techsupport channel there's some install guides (pinned messages)

[fervent thunder]

thank you i will try that

[atomic mortar]

If your on amd, definitely the forge guide

[fervent thunder]

Nvidia GPU

[atomic mortar]

How much vram?

Or what Nvidia one

[fervent thunder]

12GB, im trying it out with a Tesla P100 I got cheap. I'm new to all this. Got a Xeon 12 core CPU and 32GB RAM

it was supposed to be 16GB but ... typical eBay

[atomic mortar]

Hmmm im not sure how fast tesla is but yous be able to run xl probably just fine

Maybe flux schnell

[fervent thunder]

I heard its a little slower than a RTX 2060 but with more VRAM so we shall see. I'm not expecting miracles from it to be honest. Just want to try all this and see how i get on. Might get an RTX 3090 or something later on.

[fervent thunder]

Is there an install guide for Linux?

[desert dagger]

no. but if you ask on the comy discord, you'll probably find someone that can walk you through it. most custom nodes won't run though, they're written for windows

[weary valve]

i'm upgrading my gpu to a 5070ti in a few days, is my current pc setup good enough to not terribly bottleneck the gpu for stable diffusion? 7800x3d and 32gb ram

[pale cape]

hi

hi

[median jewel]

i need help to be able to run this command in python

## import the libraries(instant)
from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler
import torch

## load the model to cuda(should download the model automatically, time depends on your download speed)
pipe = AutoPipelineForText2Image.from_pretrained('lykon/dreamshaper-xl-lightning', torch_dtype=torch.float16, variant="fp16")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe = pipe.to("cuda")

## inference time(should take a few seconds or so)
prompt = "portrait photo of muscular bearded guy in a worn mech suit, light bokeh, intricate, steel metal, elegant, sharp focus, soft lighting, vibrant colors"

generator = torch.manual_seed(0)
image = pipe(prompt, num_inference_steps=4, guidance_scale=2).images[0]  
image.save("./image.png")

i am getting error after error but it should not be that difficult

i cant find the right information for what version i should have of diffusers, transformer, cuda, pytorch etc

[oblique elk]

Maybe you might share the error you are getting in the #🤝｜tech-support channel

[fervent thunder]

where do .safetensors files go

this sucks

[full lark]

Which program are you using? forge? comfy?

They usually go in the checkpoints folder

[abstract quarry]

it's just a generic format for matrices. It depends on what's in there

[fervent thunder]

its ok i figured it out. It's working now but my god is the text to image a little crazy haha some wild results. This Tesla P100 is working well too.

[minor dragon]

which models do you use for realistic human images?

like real humans

that don't look fake

[inland mortar]

Would it be possible to get a workflow in like comfy that can generate full material assets? Like it can make:
Color Texture

Normal Maps

Metallic Maps

Roughness Maps

Specular Maps

Ambient Occlusion (AO) Maps

Height / Displacement Maps (via tessellation or parallax)

Emissive Maps

Opacity Maps

Anisotropy Direction Maps

Subsurface Maps

Refraction Index Maps

Clear Coat Maps

How much of this would even be possible with all the crazy stuff you can do with comfy UI and could there eventually be a full material generation workflow

[fervent thunder]

who needs vps - 8 gig of ram, 4 cpu and $12 a month

[median dome]

Hi. Anybody online?

[knotty marsh]

@median dome How can we help you ?

[median dome]

Hi. actually im messing around with swarm ui and wan video and it takes a long time to create the video. i was just looking for a way to pass the time^^

are you using wan video?

[floral umbra]

WAtch youtube, movies, shows, anime. Or play games. Ez.

I'm using wan video

[median dome]

well i felt like chatting. i also like to learn more about ai creation. im just a beginner but its a very interesting topic

[fervent thunder]

https://youtu.be/yEdtNRAcMLc?si=5VB4yG69uVmLOCYs

[cyan flint]

Is anyone here familiar with "mimofr"?

I'm trying to figure out what kind of AI he's using for his wizard videos

Trying to create something similar

[cedar girder]

[cyan flint]

[ocean escarp]

Hi Guys!!

[long shore]

hello!

[ocean wraith]

Is stable diffusion dead? Are there any other open source image generators?

[abstract quarry]

flux

but why should it be dead? in particular, sdxl and it's derivatives are still very much alive

[round siren]

Are there any SD Zluda AMD users here? I feel like generations are abnormally slow. A flux 1000x1200 2 image batch literally estimates 1 hour 30 minute completion time

Illustrious is maybe like 2 minutes for 2 images

[lethal zephyr]

I cant generate anything with SDXL since it keeps loading and loading

I don’t know if anyone is having similar issues

[desert dagger]

in the #🤝｜tech-support channel. you want to talk to @warm junco

[desert dagger]

what interface are you using to run it in?

[lethal zephyr]

I’m using Stable’s API to run SDXL in my platform

[desert dagger]

don't suppose you'd DM me your code so i can look at it?

[lethal zephyr]

Hey, sorry just seeing this. I was able to fix it after all, thank you though 🫶🏽

[undone garden]

If I take a picture which has a white background and I img2img it with 1.0 denoise and controlnet, why on earth does it keep generating a light background no matter what? Like, how does it even work? How would it know it has a light background in the first place? I'm trying to understand how it works, it's starting from pure noise and only has lineart contours to go with

[desert dagger]

because it has to generate something and it cant' create transparency

[nimble slate]

heya

i need a lil bit of help

can someone help me 😭

[desert dagger]

what do you need help with?

[nimble slate]

go to general with images

i keep getting a monstrosity 😭

[abstract quarry]

there are two reasons that both contribute to that:

1.) You often do not start at timestep 100% (pure noise), but at some timestep shortly before the 100%
The reason lies in the math of epsilon prediction diffusion (the one most models like SDXL and SD 1.5 are using. Note that SD 3 and Flux are using Flow matching instead)

the reason is that in epsilon prediction the model predicts the noise and not the image. So given an noisy image the model figures out how the noise looks like, then this noise is subtracted from the image

however, if you would give a pure noise image then after subtracting the noise the image would be empty. So denoising from 100% noise would not make any sense

that's why you start from, say, 99.9% noise. You use the underlying image (by default just a blank image, which is somewhat gray) for the remaining 0.01%.

This is the reason why images in SD 1.5 often look "grayish", "unsaturated", "medium brightness", because the model still reconstruct from a gray image

When you do img2img from white images, then the white part often appears in the resulting denoised image, too, although you denoise with 99.9%. The reason is that white has a very high value in the latent and, thus, is still visible through all the noise

[abstract quarry]

2.) In diffusion you mix your signal image with random noise. But what does mixing mean? Basically, you multiply your image with a value called alpha and your random noise with a value called sigma and then you add both together. Alpha and sigma depend on your timestep. Intuitively, you would think that alpha+sigma is always 1 and this is the case but only for rectified flow matching. In Flux or SD 3 its very simple. If you are at timestep 50% then this means that you multiply your image with 0.5 and your noise with 0.5 and add both together.
In SD 1.5 or SDXL it's more complicated. Here, alpha and sigma are determined by your noise schedule and they don't have to sum up to 1, nor do they start at 0 and 1. So you might always have some remaining of your input image in your noised image, even if you would start at timestep 100% (what you are not doing in epsilon models, see 1))

btw. this also means that this effect only occurs in SDXL and SD 1.5

if you use Flux/SD3 and use 100% denoise then you should not see anything from the input image

[undone garden]

Hey thanks for the time for the comprehensive explanation mate

[floral umbra]

Where do model creators go these days to upload their models? As after civitai's latest restrictions due to asshat-visa's demands, appears at least wan and hunyuan model uploads has quite died down there.

[atomic mortar]

Tensor art works but ironically that place has less censorship with many more problems

civitAI has been hit with the no nsfw so the site where borderline (or actually illegal depending on where you live) stuff is hosted gets popular sadly

[abstract quarry]

lol, civitai is far away from "no nsfw"

people are crying cause they can no longer generate deepfake-like porn with their favourite actors and call that censorship

[floral umbra]

Not NSFW entirely, more "against their will", and other stuff like piss and diapers, that sort of stuff.

[atomic mortar]

Oh definitely, love the payment processors deciding everything nowadays

[atomic mortar]

Even sfw gens are now allowed of certain peolle

[floral umbra]

Right? Just become worlds most dominating transaction company and change the world /s

[atomic mortar]

Thank god opensource is a thing so the restrictions arent hurting everyone

[floral umbra]

It's not so much about the open source thing, it's more hurting users wanting to access models, as civitai is the most known, and if they really want their piss, they need to make the models themselves, as civitai will not host them.

[atomic mortar]

Its not about piss but okay

[median dome]

Hi. Anybody on who knows swarm ui?

[desert dagger]

you can host models on huggingface...

[empty pond]

Be cautious—scammers are using curated GitHub repos to impersonate AI developers. Stay sharp out there. 🔍

[empty pond]

anyone in here focused on the api side of ai-video development?

[thorny bloom]

Anyway to create longer videos with wan? If i extract the last frame and use it to i2v it looks like a strange cut in color and saturation

[fervent thunder]

do you mean making an API or using an API?

Docker + Kubernetes + standard python web dev libs works fine

if you need enterprise-grade stuff this won't do but its fine for most uses even up to small startup size

its not rly worth re-inventing the wheel

[floral umbra]

Aye, but i've no clue how i'd find specific models there :P It's not like civit which has tags, base model and so on for search filtering :P

[floral umbra]

Oof, way too late attempt to archive civit's lora's, and so far, like 3-5% of civit's lora's has been wiped.

[desert dagger]

google search for the model - example: this_model.safetensors

they come up

[floral umbra]

And what if i wanna see what's new that people has cooked up? :P Can't know the name of what i've never seen nor used x)

[desert dagger]

you're not gonna see that on civit, either.

[floral umbra]

Nope, so i've been looking for any alternatives to be popping up, so far none

[winter zephyr]

hi

[dusty talon]

Hello

[glossy fulcrum]

what is the best image generation model currently

[willow sand]

best for what?

[steel prawn]

yeah kinda depends on what your use case is. if you want a good all rounder, probably Flux or HiDream. If your aim is anime, cant go wrong with Illustrious or NoobAI. If you want realism, Flux and Pony seem to dominate that area. And if you want creativity, like unique art styles, Flux, HiDream and NoobAI seem to handle different art styles really well.

In actuality, you can use any image generation model to make any type of image, assuming the checkpoint is fine tuned or trained for it. Some behave better and handle prompts for certain types better than others, and the more popular one is in a specific type of image, the more variety you will have of checkpoints

[austere relic]

Dumb question: stable diffusion has no easy-to-use app to generate images, videos, and sound using their models, right?

When is the next upgrade coming? SD is getting far behind...

[atomic mortar]

?? Sd, flux etc don't got their own apps

They got sites maybe

But for local comfy/swarm/forge is as easy as it gets

Swarm being the best out of 3

Imo

And they offer an api for use in said UI's

[austere relic]

Google, openai, midjourney etc developed an official UI.

[atomic mortar]

okay but can you run those locally?

no

developing a UI for only their spesific params and open source mess with it how you want is kinda different and really not nessacary in the current ecosystem of open source models to have a stability UI only

an d correct me of im wrong they do have one ope, just a website https://stability.ai/stable-assistant

[austere relic]

Hence the question to ask when they are going to upgrade the models to achieve something that aligns with current standards?

[atomic mortar]

sd3.5 large is still pretty solid

[floral umbra]

Sadly sd 3.5 stock is terribly tuned vs flux.

[floral umbra]

No one knows. Like they just released a new sound model for creating music with lyrics and all. Ain't as good as suno in sound quality, but it's not bad at all for being free and local.

[floral umbra]

New "next gen" models only arrive as fast as people discover new ways of doing things/accumulate dataset and word things properly to teach the hallucinative models to be more accurate.

[glass spruce]

Hi

[merry coral]

hello

[pastel arch]

Hey everyone! I’m an AI engineer focused on generative image pipelines. I’ve worked with Stable Diffusion (1.5 & SDXL), LoRA fine-tuning for consistent visual styles, and ControlNet for pose/structure control. I’ve also deployed models via Replicate and Hugging Face. Always open to collabs or contract work—especially if it involves stylized character generation or creative AI tools!

[abstract quarry]

you misunderstood this. Official uis are a bad thing. They usually lack features and are technically behind..The great thing on open source is that you have plenty uis to choose from, all competing against each other and come up with new features regularly

[abstract quarry]

Flux was SOTA for long time now. openai's Sora is quite new and, yes, it's prompt understanding is much better than any open source models. But it's also much newer

[fervent thunder]

in 3D and audio they are, they made SOTA releases this year

if you look at their news they are repeatedly hiring 3D people this is probably their new direction

Stepfun released a 30B DiT and the scaling from 14B Wan to 30B Stepfun was not spectacular

I think this means merely scaling DiT has reached its limits

so everyone needs new methods anyway

[lucid bobcat]

Is it me or is civitai barely functional recently?

[desert dagger]

what current standards are you referring to?

[desert dagger]

civit seems to have become a bad taste in a lot of people's mouths in the last 48 hours

[austere relic]

I agree that creating a 3D environment is the future as it allows to keep coherence, and correct shoot angles, lighting etc as well.

Current standards are better in text writing, for example.

[sweet ore]

It's the same thing that happens to every software product

They just become bloated and shitty

Civit exists only because it's the central location to flock to. I'm sure everyone would go with the wind if a new player came to town

[fervent thunder]

https://youtu.be/a6xbUBTlf2s?si=4i6Yr7xIOE8RoIga i finally added my voice to my videos!

[rain sandal]

Hi, I'm a fullstack developer having some fun with image generation for a sideproject, having some issues with Stability API I get this error:
An upstream service tried to fulfill this request, but a required resource was not found,

when using the google collab from main website:
https://platform.stability.ai/docs/api-reference#tag/Generate/paths/~1v2beta~1stable-image~1generate~1ultra/post

How to solve this? I want to use Stability API

How to do this? Invite is invalid it says

[atomic mortar]

Scam

[atomic mortar]

That is a scam link dont join

1326848249579049002, userID @vapid dove another discord scam support link

[ebon torrent]

hey @rain sandal same thing here. I'm developing an app for which Core and Ultra are integral, but both are kicking out those errors. I thought it was my configuration until I checked this link https://stabilityai.instatus.com/

[rain sandal]

I solved my issue by using a different api key , from a different account , and am no longer mentioning Bearer in the authorization which also seems to help
{"authorization": STABILITY_API_KEY, "accept": "image/*"}

Hope that helps!

[pulsar prawn]

Any working Data Scientists here?

[pastel moon]

hi guys im using sd webui replacer, and every time i generate images, the models SAM and grounding dino are loaded each time. is there anyway to prevent this for faster speeds?

[dusty trellis]

About 90% of the time that I generate a set of images from the same prompt (with Flux Dev), the first generated image is the best of the set (of 4, of 8...). Any reason for that? There can't be

[floral umbra]

Attempted to download all of the remaining lora's from civit.. 10794/369564, 3% done, 1.5TB so far lol

[clever yew]

Good morning. Has anyone here already installed Stanle Diffusion on PC? I will need help

[floral umbra]

Best to start with what you actually need help with :P

Also, SD on pc, what else would you have it on? (i guess there's also on phones, but that's so super niche barely anyone here does it lol)

[clever yew]

I want to fix an error message (I can't show you screenshots so it's hard to describe)

[clever yew]

I mean not online version

[floral umbra]

Aye, i prefer screenies as they are easier to explain than words :P

[floral umbra]

Well, technically they are "online" as it's webui after all

[clever yew]

sorry but for some obscure reason I can't send any here ><'

[oblique elk]

You can use channels like #🤝｜tech-support or #🌶｜off-topic to share screenshots.

[floral umbra]

Indeed. Lets hop into TS

[clever yew]

Thanks

[still glacier]

Preshoting the bots and scammers.
There is no "official external support discord" of any sort. Don't get scammed.

[near spear]

.

[hallow hare]

So, uh, with SD 3 being a complete failure, what is there to look forward to beyond SDXL/Pony/Illustrious?

[low tulip]

I'm trying to create a LorA of myself using 15 selfies but the results are pretty bad... I used a paid service using the same 15 pictures and the results were decent, so i know the training material is sufficient, theoretically. Any tips?

Or resources

[atomic mortar]

Its pretty good (sd large) but they are working on mostly 3d rn i heard

And music

[hallow hare]

[atomic mortar]

But there's flux, wan etc that's still being worked on

I just wish there was a more efficient model

Less vram requirements for like Illustrious quality

[pliant moss]

The new illustrious based models have some pretty advanced natural language prompting and non-anime artistic capabilities. Has there been any advancement in the non/anime SDXL world in this area?

[abstract quarry]

which model do you use?

from my experiments so far, Flux is by far the best for photorealistic images of the own face, but SDXL is a bit better with transferring the face into different styles (anime, comic and so on)

[simple trail]

I'm using Google Wisk but I have an image I like but I want to only change characteristics of the image, not the main image is this not possible in Wisk, and if it's not what route must I go ?

[atomic mortar]

Scam

[split narwhal]

is it possible to run LTXV in SD ?

[still glacier]

you re confusing SD with whatever client you were using previously to run SD model.

SD is a type of generation model and so is LTXV

if the question is "Can Automatic1111's stable-diffusion-webui use LTXV model" then the answer is no

[split narwhal]

yes it was the last one yes

darn so pity

[floral umbra]

Need opinions on how people store their loras for a script i'm working on to sort and de-duplicate models by hashes.

How do you guys store them?

sdxl/funny_cat.safetensor for example?
Or animals/sdxl/funny_cat.safetensor?

Or how do you sort your loras?

Speaking of, if there's any mods active right now, what's the mod role to ping for scams?

[atomic mortar]

Nope, no mods just fruit n maxfield

[atomic mortar]

I sort em by archetype > version > type > lora

So illustrious > v0.1/v1/v2 > animals/style/pose/modifier/character> lucy.safetensors

Bit more work but this way I know what i have and don't have for my models

But since your archiving everything on civitai i suppose your method is better

[still glacier]

If none are present, you can try to @ any active "community guides".

[floral umbra]

[floral umbra]

Noted :)

Gonna see if i can make the script have a config.json with the sorting script where they can choose options, easily just replace "1, 2, 3, 4" etc, just simple numbers, with a commented out list below of what each number sorts by

[subtle matrix]

dog ¯_(ツ)_/¯

[jovial pollen]

hello

[ancient mauve]

Does anyone here use txt generation UI?

This one

https://github.com/oobabooga/text-generation-webui

[floral umbra]

Years ago, yes.

[ancient mauve]

@abstract quarry did you used it?

[ancient mauve]

I'm trying to put image support on it

Like, analyze and generate images

[abstract quarry]

I switched to LMStudio

[ancient mauve]

is Gemma3 multimodal ?

[abstract quarry]

yes

[ancient mauve]

Then there is no option for images with the current setup I have

I suppose I would have to add an extension or something

[quaint cape]

we're looking for a comfyUI designer to join our agency full-time, long term position, fully remote🙌🏻 anyone interested?

[atomic mortar]

you might wanna offer more details,
making workflows or making use of custom nodes or even making custom nodes
what is your agency for? what does it sell?
compensation?

hardware/cloudgpu provided?

[low tulip]

I used sdxl, I'll give flux a shot, thx

[low tulip]

Would you guys take interns? And yeah can you give more details

[atomic mortar]

seeing your UK based,
will payment be in Pounds, Euros or Dollars?
What do you envision to be the workload to be? making workflows 40 hrs a week seems hardly realistic

or is it purely marketing (example, sell a product in a AI enviorment like a new bag from a brand)

[quaint cape]

I can give you all the job info on private maybe? don't wanna spam the group... in case they kick me out lol

[quaint cape]

yeah this can be an option too

[glossy elbow]

https://suraural.org/

[hoary pecan]

Hi,

Does anyone know how to set up the image to video in n8n?

I’ve set up the text to image model and want to convert that into a video, though I’m not sure how to convert the image into readable binary format

[paper garnet]

looking to make detailed depth maps from images, to acheive 3d sliced laser engravings... anyone got experience of good models or workflows for this?

[final tusk]

I wanna take a current poll of how many people have failed to watch/read ONE PIECE AKA the greatest fictional story ever told. Give me a percentage/number for the sake of history. We will keep a record. Thanks and goodbye~

[spark mesa]

image

[rare holly]

how well will a 2080ti do in SD ??? iam using a 970 gtx ... will i notice a massive upgrade ?

[keen prism]

Do you need support

[dawn mulch]

So couldn't someone just get an ai to write an ai image generator model that is only 1GB that can generate pretty much anything. I mean stable diffusion for example is just a bunch of what do you call it code or zeros and ones. I think it's possible if you had a strong enough computer. Youll get hit and misses but its doable. And that goes for any other ai generator. Generates in .01 sec and uses less than 1gb of vram/ram. (an AI model generator)

[rare holly]

PMD

[still glacier]

do it, I dare you.

[still glacier]

that s most likely a scammer who s gonna send you some private discord stuff tricking you into installing malware.

there is no "external support discord".

the support place is #🤝｜tech-support and is community driven.

[fringe urchin]

how do I prompt for squinting eyes? I always get a frowning face instead of squint in attempt to see better

[rare holly]

no wonder he has not answerd and still online for hours 😆

[valid river]

Does animate diff for sdxl automatically reduce quality of generated frames?

[signal goblet]

@valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river @valid river

زب طيز كسكس امك

translate

[atomic mortar]

I do know animate diff is autodated and doesn't follow frames very well, what gpu do you have?

You might be able to use newer video models

[valid river]

4090

[atomic mortar]

Try looking into WAN2.1 with the causvid lora, high quality video should take like 2-4min then

Youd need to use swarmUI or comfy ui though

[minor dragon]

listening to this banger
https://open.spotify.com/intl-de/track/4B0AzVyr78p6VK1gbJ0hsh

I don't understand a word
but it sounds aggressive
I think it's dutch

[oblique agate]

https://huggingface.co/Agents-MCP-Hackathon anyone want to join force with me

[quartz siren]

Do you mean the model will just provide training code or create a model from scratch? First is definitely possible right now, but the 2nd is not possible at all right now.

A 1b model is going to have 16,000,000,000 bits(0s and 1s). Even current llms can barely make coherant 100,000 tokens and they suck at anything too out of domain.

[abstract quarry]

it doesn't make sense at all. llms can only perform tasks they learned from humans. They can code, but only stuff they learned and they still do a lot of errors. An experienced programmer can definitely write better code than llms

and creating generative image models requires training data and compute power. Writing code is the smallest issue here

[clever yew]

I have two questions: Do you have techniques for the AI ​​to take into account all the elements of the prompt? And can we insert a reference image into SD?

[abstract quarry]

reference images can be included with ipadapter. For specific use cases there are also control nets

[clever yew]

ipadapter is a software ?

[abstract quarry]

an adapter

it's like an extension that allows image inputs into the diffusion model

there are different adapter models. Some are specialized in transferring style, or faces or content of an image

[fervent thunder]

this exists on some level its called hypernetwork

but they can only create very very small neural networks

[lean ravine]

hi, everyone

[tame vector]

hi

how can i but seperations in my prompt for better navigation? i get lost in the prompt sometime

[atomic mortar]

the usual practice is using a comma and space
,
so youd get:
1girl, retro, goodquality (horrible prompt for non illustrious/pony models)

[tame vector]

sorry, i need to be more clear.
i want to seperate between groups of prompts.
example:
dynamic perspective, good anatomy, looking at viewer [enter seperation] black hair, long hair, brown eyes, [enter seperation] vibrant colors, cel shading, detailed background

my goal is not to influence the prompt with the seperations. I just want to make it easier for my eyes to navigate the prompt when i make changes

[atomic mortar]

hmm it depends on the UI but in SwarmUI i use <break> to split up the prompt (and tokens)

in forge it was BREAK but im unsre

[meager jasper]

Do you guys think ComfyUI tools like LTX/Wan etc can compete with the quality of Kling?
I'm not saying it can't I actually don't know and wanna hear opinions

[drifting stirrup]

what do embeddings do?

[fervent thunder]

on their own they don't do anything they are just some data

[drifting stirrup]

ah

im using comfyui so what do i do with those?

i put it in the embeddings folder but idk if i have to do anything else

[fervent thunder]

Wan and Kling have very similar blind winrates so they are around the same quality

only Veo series are significantly better than Wan

[fervent thunder]

oh you are referring to clip embeddings ok
you need to type something in the prompt box to use them

[drifting stirrup]

ah okay ty

[meager jasper]

Interesting good data to have thx

[tiny estuary]

hello everyone

[signal trellis]

since civitai removed their real people/celebrity loras, does anyone know of an alternative site that hosts them?

[floral umbra]

I've been searching for days, so far heckall sadly.

Only places i can think of is huggingface, but you gotta nearly be a wordsmith to find out the right combinations of keywords to find someone's model repo of celeb loras

[tame vector]

can somebody recommend an Illustrious Lora for expressions?
I want to use it on anime/stylized images

[atomic mortar]

https://civitai.com/models/1204189/best-facial-expression-helper-xtreme-illu

[atomic mortar]

this one is pretty good imo

linked above mb

0.6 works best for me

[tame vector]

@atomic mortar nice, thank you so much 🤙🏼

[floral umbra]

Models: 5%|███████▉ | 19922/369564 [16:14:32<1349:56:34, 13.90s/model
Damn, near 20k models, and think it's closer to 3TB of loras now lol

[floral umbra]

Also, do you guys know of a git that can safely convert a ckpt to safetensor in a contained environment? As even when converting, maliciously injected ckpt's could be executed while converted to safetensors

[atomic mortar]

Ckpt were pickle files right?, im not sure about contained but i suppose you could use a google colab solution

[floral umbra]

Aye, downside, is that as you see right above my comment, i got a few 1000 lora's archived on my server, and loads are the old ckpt, so need to have them converted and pruned.

I'm slowly thinking of just running a VM with windows and converting them by taking the auto1111 extension for converting them, dump into GPT and have it make a script for me to run it standalone, and access the lora location and it's folder structure to auto prune all that can be pruned.

[abstract quarry]

if ckpt is the pickle format, then converting that to safetensors is a single line of python code

[atomic mortar]

Oh no digital currency scams again @vapid dove , deliberate attempts to bypass the filter too

[bright dagger]

does anyone have experience training a lora in kohya? im running one with 50 images and its taking 7 hours per epoch, i cant imagine this is the full performance of 64gb ram with an rtx5080

[abstract quarry]

ram doesn't matter. You have to train on your gpu

7 hours per epoch sounds like you have not enough vram

https://bfl.ai/announcements/flux-1-kontext

🎉🎉🎉🎉🎉🎉🎉

a new flux model!
and there will be an open weight distilled version of it soon!

[high moon]

Hello everyone!
I’m new to the server! I wanted to ask a question — I’m starting to use Stable Diffusion with automatic1111. I have an Asus RogStrix laptop with a GeForce RTX 4060. Which base model would you recommend using?
I’ve been testing with epicrealism_naturalSinRC1VAE.safetensors and generating some styles to add more realism, but whenever I make a prompt without specifying clothes, you know what happens... What do you recommend?
Thanks a lot in advance!

[abstract quarry]

with that gpu you probably use sdxl or sd3.5 medium

yeah, lot of custom models are full of porn. Strong base models like flux, sd3.5 or sdxl turbo don't necessarily need custom models. It depends a bit on what you want to generate

[high moon]

Thank you very much, it’s for generating realistic images

[bright dagger]

i thought 16gb was enough. task manager does show 15.4/16 gb, maybe its throttling?

[abstract quarry]

I think realviz is the typical model on sdxl for realism

[abstract quarry]

I don't know what you train.

[bright dagger]

50 image lora no regularisation

[abstract quarry]

yeah but what rank, what model, what resolution

[bright dagger]

512 squared px, rank idk, no option in my current kohya installation. sdxl 1.0

[abstract quarry]

oh, sdxl? that should not need much memory

in this case I assume your cuda is not working

you probably train on your cpu

you can check by running the python code

import torch
print(torch.cuda.is_available())

in the same console window where you run kohya

but more important: there will be a new flux model 🔥

on benchmark there commercial model performs similar to Chatgpt. I assume that the dev model will be much weaker. But still: no need for ipadapter or redux anymore, we will be able to input conditional images directly in fluz

[bright dagger]

i thought so too, but task manager says <10% cpu load and 99% gpu

[abstract quarry]

I can just say: sdxl should not need more than a minute per epoch when you train om 512px. If it needs an hour then something doesn't work. Mist likely there is a memory peak, but sdxl is relatively small, it shouldn't need that much memory. You can check if it gets better when train on rank 1 lora with batch size 1

[bright dagger]

is it possible that im using a faulty version of sdxl? i downloaded the safetensors from ~~civit ~~huggingface

[abstract quarry]

don't think so

[jolly slate]

Hey everyone. Have you seen voidstomper? on instagram?

[clever yew]

I come back with my question, but do you have any techniques for the AI ​​to take into account as much of the information from the prompt as possible?

[steel prawn]

the AI reads everything in 75 token blocks. it also prioritizes whatever is at the beginning of the prompt more than the end. also, higher cfg = more strict prompt adherence, but its still a bit of a gamble what you get and too high of a cfg will overbake the image.

[clever yew]

My problem is that I have a really precise idea of ​​what I want (a few specific details). So my prompt is very precise and detailed. Except it seems like the AI ​​always forgets a few things

[steel prawn]

thats the nature of AI. its not gonna get the image 100% correct because it doesnt "see" your vision like you do. the things i suggested above will help, but nothing will get it exactly perfect.

[clever yew]

As long as I can strive as hard as possible towards what I want, that’s fine with me. I will try your advice

[steel prawn]

you can always share your prompt with chatgpt or another llm, and ask it to segment it into 75 token segments with BREAK separations, that will help the model read your prompt better. and the llm should use logic to ensure it keeps things together that need to be together

since trying to math out the token count yourself is tedious

as for cfg scale, depending on the model your using to generate the image, i wouldnt go higher than 10 as a general rule. some models will work up to 12 or 13, for anime models they prefer 5-7 with a soft cap of like 9.

[jolly slate]

Does anyone know what voidstomper uses?

[abstract quarry]

what model do you use?
older clip based models like SDXL have only very weak text understanding.
For best prompt understanding you should use Flux.
Also, flux doesn't have a limit if 75 tokens and you don't need break separations and so on

still, for very complex prompts you might have to use inpainting. Flux' text understanding is not on the same level as ChatGPT 's Sora

[steel prawn]

i didnt say there was a limit of 75, that it read the prompt in 75 token blocks. but yeah, flux has much better prompt adherence than most other older models

[abstract quarry]

yes, but that's not true for Flux

[clever yew]

I use Elar-a

[abstract quarry]

never heard

[clever yew]

Ah you mean the checkpoint ?

[steel prawn]

yeah thats a new one to me as well

[clever yew]

Or the sampling method ?

[abstract quarry]

the checkpoint or the base model

[steel prawn]

i think they were asking about the base model, so the checkpoint would tell us

[abstract quarry]

(oh, he meant Euler Ancestral 😬)

[clever yew]

If it's Eular a, yes. Is it bad ?

And so the checkpoint is WAI-NSFW-illustrious-SDXL

[abstract quarry]

Illustrious is a special model

it has no text understanding at all xD

[steel prawn]

so your using illustrious base, and your using Euler A sampler

[abstract quarry]

it only understands these danbooru tags

[steel prawn]

not entirely true, that particular model your correct though

the new illustrious 2.0 base has natural language understanding

and any merges using that base will as well

not to the degree of flux mind you

[clever yew]

I use the version 12.0 if it can help

[abstract quarry]

haven't tested it. I found all illustrious and NoobAI models so far quite limited. They don't understand anything that is outside their domain.. However, they have extreme good understanding of any weird combinations of these danbooru tsgs

[steel prawn]

thats just the version of the checkpoint. i believe their newest version is 14.0, but its still not using a Illustrious 2.0 base

[clever yew]

I don't know what is this danbooru tags :/

[steel prawn]

ive made several merges with 2.0 as a base, and tested both danbooru prompts and natural language prompts, getting relatively similar results.

[abstract quarry]

search for tutorials for how to use illustrious

it's a special model that has to be prompted in a very special way

[steel prawn]

it still has an anime focused training though, and unless you merge loras for crazy concepts into it, some things will not be understood, especially really flavorful language

[abstract quarry]

I don't use these kind of models so I can't help

[steel prawn]

https://danbooru.donmai.us/, its the method of "tagging" artwork used by this site. most older illustrious models (including Wai-NSFW-Illustrious SDXL) are trained on images that are "tagged" or referenced by these tags. The end result is a model that understands these better than spoken language. The prompts will look like "1girl, school uniform, blue hair, red eyes, standing, market, outside, daytime" etc etc

The 2.0 base can understand paragraphs of words written like you would normally write them, but its still not got the same level of inference and understanding of newer architecture models like Sora or Flux based models.

[clever yew]

Ah ok. It's what I do generaly

[steel prawn]

Then the things i suggested to help you will work for you. Still wont be perfect, but it will get your closer. Then you just gotta either inpaint what it missed, or press generate til it gets it right

i will say, sampling methods matter, as do schedulers, to the end result. I find that DDIM sampler or Restart is good for images with crazy abstract backgrounds, whereas DPM++2M with Karras scheduler is good for low steps and high detail. Euler and Euler a pair well with DDIM or SGM Uniform, and are pretty creative samplers if you like a little creativity added in.

Of course, the model will play better with some samplers than others, so usually you should START with the recommended settings, and tweak as needed for your taste.

[hushed hollow]

hello everyone

[fervent thunder]

hihi

[shell tendon]

https://www.reddit.com/r/comfyui/comments/1kywtae/universal_style_transfer_and_blur_suppression/

[pure brook]

https://www.twitch.tv/nekomipii playing my favorite game league of legend lets rank up with 198 ms im i getting to gold today ??? playing NA and JP /maid cosplay

[bright dagger]

i get weird noisy colorful images from my epoch samples, has anybody ran into this issue?

[agile musk]

i keep reading "abliterated" model - what does abliteration mean? no dictionary I know of seems to contain this word

[atomic mortar]

https://huggingface.co/blog/mlabonne/abliteration

assuming yoir doing it for a llm

but if you want a less spesific:

Abliteration involves fine-tuning a language model to bypass built-in refusal mechanisms that prevent the model from generating responses to potentially harmful or sensitive prompts. This is achieved by analyzing and manipulating the model's activations to distinguish between "harmful" and "harmless" prompts

[agile musk]

thanks!

[safe cliff]

Context? Are you training?

[bright dagger]

Hi, thanks for your response. I'm training a 50 image lora on SD1.5 (SDXL gives good results but it takes 50+ hours where SD1.5 only takes 20 minutes) with 1600 steps. I will send you an example of my results in dm because i cant send it here

[safe cliff]

If you started with SDXL and swapped to SD1.5, make sure you aren't still using anything intended for SDXL in the workflow like the VAE. That would be the expected result.

Do you have a screenshot of your workflow, or would that be giving away too much information?

[bright dagger]

I have restarted my entire kohya server before swapping between sdxl and sd1.0
what do you mean by the workflow? just a screenshot of my settings?

[safe cliff]

Yes, what settings?

[bright dagger]

I will look them up, one second

Wait, I think I got it working actually, the time drasitcally went down from 40 hours to only 1:40

[safe cliff]

Oh geeze lol

[bright dagger]

This is awesome

I enabled no half vae, because you mentioned something about it, and it seems to work

i also discovered that i am working in the dreambooth tab of kohya and not lora

thanks for your help man!

[safe cliff]

Glad to help :)

[clever yew]

Is there a way to prevent my image from looking crooked?

[atomic mortar]

Inprove the prompt probably

[spark bone]

Hi everyone

[crystal sail]

Is it possible to use img2img to spoof real photos? The goal is to get 100 unique photos that look the same from 1 photo. Is this method better than ffmpeg spoofers?

[forest delta]

Hi Guys, I am grateful to be here! I have a question: I am trying to morph the face of a photo into a Santa Claus' head. The face of the photo should not be generated but taken as it is. Only the outer borders of the face should morph into Santa's head. I experimented with masks and distinct API methods, but I could not get it right. The nearest I could come it preserving the face of the photo while generating some Santa related features over it. This is great but unstable/unrelyable - the goal is to have a process that brings me the desired result as often as possible. Any ideas?

[wet grotto]

hi guys, Im finishing to record my course of generating images for architetcture with stable diffusion, it's between 25 hours of course. I finhised the practical part and started the theorical
About the models, I will show the size of then but I'm a bit out of the news. Must I show any other model of they that I choosed?

SD1.5
SDXL
SD3.5
FLUX

[oblique elk]

For face "swapping" i would suggest pulid model with flux which works quite good with non extravagant faces. Newest model would be dreamO which is also good to swap faces.

[tame vector]

how can i generate multiple characters on the same image, using different Lora for each?

[steel prawn]

Regional prompting or using a model like Flux is your best bet. Regional prompting is still a tedious process though, and requires some learning.

[wet grotto]

x,y,z script? are you usind a11 or comfy?

[tame vector]

i am using Automatic1111

[wet grotto]

maybe with this, you can generate more than one image with varieting some parameters, idk if it can be used for LoRA, you have to check
https://www.youtube.com/watch?v=1P-6con4xFQ

[tame vector]

oh sorry i meant like multiple characters interracting in one image

[wet grotto]

maybe this, regional prompt

Latent Couple maybe work too

[valid aurora]

Hey guys! Sorry to ask out of nowhere, but I have a favor.

So I've been trying to make a LoRA for like 2 months now and I’ve learned a lot, but I just can’t get any results. Right now, another person is helping me and said the reason I can't make a LoRA is because of captioning problems, and we’ve been back and forth with many possibilities on what could be the problem of my lack of success, not just captioning. Tbh, I’m just a bit demotivated right now because I’ve had no results despite trying a lot.

So my favor is: can you let me borrow a set of training images + captions that you have, that you were able to successfully make a LoRA with?
That way, I’ll know whether it’s a training image/captioning problem, or a parameter settings problem, or something else that’s stopping me from creating a LoRA. If you were able to create a LoRA with the images, and I used those same images but i wasnt able to then i can assume that means its a parameter problem or something else and I can cross out one more possible causes from the list of possible causes of my lack of success in lora creating. I just want a taste of success. I feel like I’ve been heading nowhere despite my efforts, so I just want to know what it’s like to successfully create a LoRA once, even if it’s with someone else’s training images + captioning. Would anyone be able to help me?

(sd 1.5, automatic 1111)

[atomic mortar]

Are you using dreambooth?

if yes i assume thats the problem

try looking into OneTrainer/Koyha

[valid aurora]

ah thanks for replying. no I am using Kohya and the problem seems to be the captioning, after I solved that with the help of someone, my results are finally consistent and better ! just the face is still very distorted when full body and not close up face xD. man I didnt realize how important captioning was

[atomic mortar]

Ah faces, especially in SD1.5 you need to use adetailer

[valid aurora]

ohh i will look into this !

[abstract quarry]

why are you using sd 1.5 and not a newer model?

[lusty prairie]

Can somebody help on how to train Flux 1.D Dreambooth models or Fine-tune not checkpoint merging nor LoRA training on Kohya_SS . I was looking for tutorials and videos but there are only a limited numbers or resourses available online . I was researching in the internet for last 2 weeks but got frustated so I decided to ask here . And don't recommend me this video , when I started with SD and AI image stuff I used to watch this channel but now a days he is putting everything behind a paywall . And I'm already paying for GPU rental services so absolutey cannot pay patreon premium.
If anyone has resourses/tutorial please do share here (at least config.json files which I have to put in Kohya_SS) . If anyone knows other methods also please mention them . (Also it is hard to train any model via Diffusers method and also the result isn't that great thats why I didn't do that.)
Thank You.

[tame vector]

since using regionalprompter my CPU is running at around 95%.
SD freezes often while just writing or copy/pasting a prompt
is this a common issue?

[fervent thunder]

https://youtu.be/nyXeRZJnNLY?si=Rbp95qBQcWofNJmw

[blissful fractal]

Hello Can someone guide as to where I can download and how to install Stable diffusion? Noob here. Ty for your patience.

[steel prawn]

scam dont click

[abstract quarry]

why would you want to finetune flux? There is not much you cannot do with loras

and yes, please don't pay for tutorials xD

[atomic mortar]

another digital currency scam yay

[desert dagger]

@vapid dove scammer alert

[wind tundra]

oof

not a great first message to join on

=o.o=;;;

[wind tundra]

what's a good safetensors model for buildings/environments?

[tame vector]

Can someone recommend me a Lora training guide?

[vernal herald]

anyone got tips against prompt blurring? Like it mixes stuff from previous parts of the prompt with later parts. An example is I want a character to have blue hair and a red shirt, but stable diffusion makes the character have blue hair and a blue shirt with a red logo on it

[steel prawn]

man this week has been wild for reports.

[desert dagger]

@vapid dove

[dusty trellis]

When you use more than one lora in an image, it is normal that either the quality of the image suffers and/or one of the loras loses its effectiveness?

[solemn flume]

hi

[still glacier]

Yes. Think of lora as "patches" on a model. So if they operate on the same "zone" of the model they re gonna fight each other. You can try to tweak their weight in your prompt, use them in multiple pass, use regional prompting, use them along controlnet to constrain them depending of what you re trying to achieve, etc

[atomic mortar]

Mixing style loras can have interesting results though, i useually have 4-5 different ones in use

[still glacier]

It can be done for sure. But it s tricky at first :p

[rocky meteor]

hi guys! do anyone knows a guide on how to generate 2D sprite sheet using AI?

[versed mica]

Hi

[pale wedge]

hiiii, is what you're looking for to generate each frame individually and compile the sprite sheet later, or have AI generate the whole sheet in one go?

[nimble linden]

Is it allowed to search for freelancers here?

[gloomy helm]

Hi. I have a quick question. I've been using stable diffusion with FOOOCUS for a while. But I really love the way Sora works. It gives great results with simple prompts with almost no anomalies or weird outputs. Is there any open source model that works simillar to Sora that I can use on my device?

[abstract quarry]

Flux

[rocky meteor]

more like generate the whole sheet

[violet cloud]

Hi guys, hope everyone's having a great day!
I posted a help request in tech support, if anyone has experience with SD and AMD gpus, please check it out!
I would be really grateful if somebody lent me a hand

[steel prawn]

dont click, spam link

[atomic mortar]

Scam

[violet cloud]

The link seems like a regular invite and doesn't have a mask, but is there a chance that if I already clicked it my discord api token is stolen?

[tame vector]

when i Inpaint one face out of 2 characters, the inpainted face gets features from the other character... how can i fix that?

i am keeping the same seed and prompts when inpainting.
also i use regional prompter and Adetailer

[atomic mortar]

Nope, they want you to install a remote control client or link a digital wallet

[steel prawn]

nobody in this discord that is actually trying to help you will send you a link or a DM to join another discord or private message chain, nor will they ask for any of your personal information outside of terminal logs for debugging. good rule of thumb is just dont click anything lol

[violet cloud]

yeah I just haven't been to large discord servers recently and kinda forgor💀

[atomic mortar]

Unless its a hyper specific issue with comfyUI or similar but those are large servers too compared go the timy scam ones

I've sent people to more suitable servers before as we dont get many lora trainers here

[steel prawn]

thats fair. i havent seen many of those instances. but i also havent been here THAT long lol

[still glacier]

Large audience of any sort inevitably attract scammers. If they see someone vulnerable / asking for help of course they'll latch on them. And if they can do it in DM far from anyone's eyes then it's even better for them.

[steel prawn]

i just immediately report any link shared from someone who has recently shown up, only said "hi" or "hello" and then never interacted with the server in any way beyond that but is suddenly suggesting help lol works for me

case in point lol

[still glacier]

Yeaaaaaah ...

[wind tundra]

anyone know a good model for buildings 'n such? or is this not really a server for that?

all the models I'm tossing into A1111 just keep giving me characters for obvious building prompts

[steel prawn]

put no human in the prompt

[wind tundra]

even with a negative prompt asking for no people, characters, etc.

[steel prawn]

and human or character into the negative prompt

but i dont have a particular model in mind for landscapes or architecture gen

i can look and see if i find one but its not really my thing

[wind tundra]

people, characters, humans, anime, cartoon, low resolution, blurry, bright colors, daytime, modern office, messy composition, jpeg artifacts, bad lighting, extra limbs, distorted anatomy, text, logos, nsfw
is my negative prompt :(

[steel prawn]

i know a lot of people are riding the Flux hype train

[wind tundra]

that's what I'm using currently, and I just got a ton of characters

[steel prawn]

huh

thats interesting, guess its prompt adherence isnt as god tier as they say

[wind tundra]

lol

I saw it on CivitAI and it showed some neat examples of environments and whatnot, so I was like "this looks like the right one!"

[steel prawn]

https://shakersai.com/ai-tools/images/stable-diffusion/best-architecture-stable-diffusion-models/

according to this article, there are 7 that are recommended here for architecture gen with stable diffusion. i havent used any personally. all should be available on civit

[wind tundra]

oh, I think I know what I did wrong... I did the thing I normally do with characters, where I put a "high weight name of thing" at the front, and in this case it was "(The Hollow Veil:1.5)"... Veil

[steel prawn]

i know for a while a lot of people were using Juggernaut XL for realism, and landscapes in the example images looked really cool. but i cant say how much success youll have

[wind tundra]

I removed that and it seems to be working derp

[steel prawn]

could be it

[wind tundra]

buildings don't really have veils xD

[steel prawn]

hahahah

[wind tundra]

god damned it =x.x=

[steel prawn]

not with that attitude

[wind tundra]

hahahahahaha

but yeah, I do that because it creates consistency with characters really well xD

"(<Character's Name>:1.5)", it's not as good as training a LoRA, but man, it works really well with some models

and I'm terrified of training a lora

I looked into how it's done, and it seems extremely tedious and slow

[steel prawn]

uhhh i think its easier than the jargon makes it seem, but i havent done it either. i just got into merging model checkpoints, training a lora is the next step on my journey lol

[wind tundra]

lol

man, flux is fast

[cyan sapphire]

well yeah

[violet cloud]

blender's going crazy with them too btw

[cyan sapphire]

Currently training LoRa with 250 epochs bcs i only can do a batch size of 1

just 2 hrs 30 mins

and idk if the results are good

luckly i have a sample every 5 epoch and Validation after 10

[violet cloud]

what kind of LoRa are you making?

[cyan sapphire]

also not even training with 1024 resolution lol

[cyan sapphire]

character

[violet cloud]

which? if it isn't a secret😄

[cyan sapphire]

u dont wanna know

[violet cloud]

hmmm ok ok😅

[cyan sapphire]

it is my idk 10th time training?

have some Saftetensors now but i want better results

still have some artifacts (tbh the images i use are pretty bad quality (720p lol)

[violet cloud]

hopefully you'll get something satisfactory in the end🫡

[cyan sapphire]

i hope

[sullen hamlet]

How do I use this? I’ve very lost

[steel prawn]

dont click, thats a scam

[sullen hamlet]

Okay. How do I use stable diffusion to generate an image?

[floral umbra]

Very vague question. What tool do you have currently?

And do you have an nvidia gpu? Or intel/amd?

[sullen hamlet]

I have nvidia and I don’t have a tool

[haughty basin]

Can I add a notification sound for sd when I finish img2img?

I have a notification for txt2img

[fervent thunder]

hello

[dense flint]

hello

i'm getting started with stream diffusion and i'm wondering if all SD models are compatible..

[atomic mortar]

I'm not sure with stream diffusion but based on the name it probably can handle sdxl at the minimum

Generally, we use forge/swarm/comfy

[dense flint]

it's meant to be used in real time

[atomic mortar]

Realtime as in generate a image as you type?

[dense flint]

the thing i'm using came with sd-turbo and sdxs.. not sure what those are

it uses tensorrt.. whtaver that is.. also it seems i can use lora too

[atomic mortar]

Ah youd need sd turbo

And if you dont have a 4090 (or better) its not gonna function as well apparently

[dense flint]

chat gpt says this:

Stable Diffusion v1.5 and v2.1: Fully compatible.

Models based on SD 1.5 or SDXL base (e.g., DreamShaper, RealisticVision): Often compatible, especially if they're not heavily modified.

SDXL: Works with StreamDiffusion, though it may be slower and require more memory.

UNet-compatible models: StreamDiffusion needs access to the underlying UNet structure to do patch-wise inference.

[atomic mortar]

Disregard that advice immediately lmao

Chat gpt is 9/10 times wrong with ai models

[dense flint]

hahaha ok

[atomic mortar]

You COULD use a non turbo model but its gonna be lots slower

[dense flint]

i love that my use case doesn't mind low fps and low res..

[atomic mortar]

What's what your trying to do?

[dense flint]

but as i udnerstand.. lower resolution also means lower details up to the point where it affects the content so I might want to stick to 512

a tamagotchi

[atomic mortar]

Exactly, less pixels = less detail

[atomic mortar]

Like a game?

[dense flint]

yeah so i'll probably keep res higher and scale it down..

yeah a game but for now if i can just animate a blob creature using control net that will be fine

[atomic mortar]

Could try looking for animating stuff within comfy if you need a gif but youd need to do research

[dense flint]

oh i got a way to animate the control net itself.. i'm hoping this is enough to roughly control how my creature moves

i'd be curious to do this in comfy eventually but i got another way to host stream diffusion that i'm more comfortable with for everything else

hmm can't share images here otherwise i'd show you how turbo looks so far with default settings

what does cuda 12.1 do better than cuda 11.8?
also somehow a desktop 5090 doesn't perform as well as a laptop 4090 in my case over here..

[hard stirrup]

Hi everyone! I’m looking for someone experienced in training custom models for Stable Diffusion 1.5.
I already have a LoRA trained on ~20 of my own photos, but I need help with either:
– fine-tuning a clean SD 1.5 base model to match my personal visual identity
– or training a fully custom model compatible with Mage.space / RunPod (.safetensors output)

This is a paid commission. DM me if interested — happy to share more details!

[dense flint]

alright so i'm using this atm: https://huggingface.co/IDKiro/sdxs-512-dreamshaper

which lora would be compatible with this?

[rocky meteor]

guys i installed automatic1111 and animediff, generating low quality image to Video takes forever and quality is not that good

any better image or text to video generator?

i heared about this pinochio browser and flux? anyone here using it and is it better?

[bright ember]

Hi guys does anyone know how to use this version of Colab: https://colab.research.google.com/github/Jelosus2/Lora_Easy_Training_Colab

[still glacier]

Yes animatediff is the old way to do videos. Search for stable-video, lxtv, wan models.

[valid aurora]

hey guys does anyone know the difference with sdxl and sd 1.5 is the difference that big ? i heard sd 1.5 distorts face a lot, does that still happen in sdxl ? im curious bout what the biggest upgrade is when generating images in sdxl is compared to sd 1.5

[oblique elk]

First the native Resolution (512x512 vs. 1024x1024). This helps a lot with details as there are simply more pixels available for objects from your prompt. Would say overall faces and hands are better in sdxl for sure

[pale wedge]

for real, SDXL’s jump to 1024x1024 really boosts detail and structure. faces and hands are noticeably more consistent, especially without needing tons of inpainting

[valid aurora]

@oblique elk @pale wedge thanks !

[hard stirrup]

Hi everyone! I’m looking for someone experienced in training custom models for Stable Diffusion 1.5.
I already have a LoRA trained on ~20 of my own photos, but I need help with either:
– fine-tuning a clean SD 1.5 base model to match my personal visual identity
– or training a fully custom model compatible with Mage.space / RunPod (.safetensors output)

This is a paid commission. DM me if interested — happy to share more details!

[gloomy imp]

anyone got some recommendations on checkpoints for fairly good anime generation? I've been playing with Illustrious-based checkpoints but it doesn't play well with openpose.

[vernal herald]

I use anything v5 and AnyLoRA

AnyLoRA is a god at lora compatability and for control net

And anything v5 is just good in general

[gloomy imp]

just Lora designed for SD, or for any?

also curious as to the size difference between checkpoints. Both Illustrious and Pony are around 6gb, while SD is far smaller

[vernal herald]

https://civitai.com/models/23900/anylora-checkpoint

its only 3gb, really lightweight, but really good for the amount of memory it takes up

also REALLY fast

[gloomy imp]

I've noticed some are better than others. MiaoMiao for illustrious is ridiculously fast. I'm working with an 8gb vram limit and its tsill under a minute image gen for 1024x1024 and then upscaling

[vernal herald]

i have only 4gb vram 🥀

and yet anylora is still stupidly fast

under a minute

[gloomy imp]

is there any crossover between SD and SDXL so far as lora use?

[abstract quarry]

no

different model architectures

[dusty trellis]

Is there any good explanation other than luck of the draw that the same prompt will produce 75% good images one day and 10% good images the next?

[abstract quarry]

variety is a good thing. A model that produces always good outputs will probably also always output the same/similar outputs

[warm junco]

There is a specific openpose model for illustrious:
https://civitai.com/models/1359846/illustrious-xl-controlnet-openpose

[gloomy imp]

only thing i'm having issues with on illustrious is that it will ignore a 'dark skin' prompt, even with negative modifiers for pale skin

[warm junco]

Try "tan" as tag

[gloomy imp]

are there any XL models that are trained on danbooru artist tags?

or is that more of an illustrious thing?