💬｜general-chat

[deep horizon]

I'll let you know once I get it setup

[abstract quarry]

no, as said trainers are always separate

you can use this one

https://github.com/bmaltais/kohya_ss

[deep horizon]

can you use LoRA's in InvokeAI?

[abstract quarry]

yes

they have quite a lot of features out of the box (like detailing, regional prompts, regional loras and so on)

[deep horizon]

I crafted a work-around for my problem in comfyui. its ugly but it solves my issue of needing to invalidated the cached prompt on each generation.

• generate random int (example: 1337)
• convert int to string
• prepend it to prompt (results in prompt like: 1337,a {blue|red|green} ford mustang parked in new york city
• concat text using text1: <int to string output>, text2: blank, separator: , to add comma after the generated int
• replace text using text: <positive prompt output>, old: <concat output>, new: blank
• output from replace text node to positive prompt input on clip encode.

put screencap in off-topic

[pure spire]

Hi, is there any decent support for character consistency in Stable? Any reliable resource would be greatly appreciated.

[atomic mortar]

Loras but if its just for the face, reactor face swap / ip adaptor could work

[abstract quarry]

for sdxl and sd1.5 ipadapter works incredible well

[fervent thunder]

InstantX stuff is nice as well

I tried their InstantStyle thing yesterday

[pure spire]

thanks, will try to wrap my head around those

[deep horizon]

InvokeAI didnt support like half of the models I have in my ComfyUI directory. I'll pass for now

[fervent thunder]

yeah I actually recently switched to Diffusers because it has the most model support

[deep horizon]

Link?

[fervent thunder]

https://github.com/huggingface/diffusers

I'm not recommending it though- only if that is the only place you can run your model

ComfyUI is way closer to pure pytorch and so ComfyUI is better for that reason

[deep horizon]

you generating from code or UI?

[fervent thunder]

in code

not sure about diffusers GUIs

[deep horizon]

ah I gotcha. was confused for a moment lol

Might take a stab at seeing what I can do from code with it. Have only used UI's before though, so it could be rough lol

[fervent thunder]

if you can read ComfyUI node code then pure pytorch is super similar

its like the easiest possible way to get an intro to pytorch TBH

[dusky sequoia]

Guys, how hard would it be to make stuff like this in stable diffusion (the comic style pictured here uses)

https://youtube.com/shorts/Bn4RO_wb57E?si=synQlzAB28IWHevr

I currently use Leonardo AI but not a fan of it

[fervent thunder]

just need a lora

[dusky sequoia]

Is this responding to me?

[desert dagger]

yeah - fyi - you can do far more at home than on leo

[dusky sequoia]

Ok that’s promising. Leonardo ai kinda sucks and I hate having to pay monthly

[desert dagger]

you should really join the L3 discord - you'll get far more indepth technical discussions for comfy and all things that go into this technology than on here. this is more for a less technical audience

[dusky sequoia]

Can you invite me please ?

[desert dagger]

can i DM i tto you?

[dusky sequoia]

Sure

Anyone here using a Mac?

[jaunty wedge]

Guys I downloaded a LoRA and put it in the right folder but I can't select it as a checkpoint, only a flux version I downloaded, how do I use LoRA?

[deep horizon]

what UI? it wont show as a checkpoint, because it isn't a checkpoint

[jaunty wedge]

the stable diffusion web ui

[deep horizon]

auto1111?

[jaunty wedge]

where can I select or deselect it?

[jaunty wedge]

where can i see that, i apologize im new to this

[deep horizon]

post a screencap in the off-topic channel

of the UI

[jaunty wedge]

I did, did you see it?

[deep horizon]

ok thats auto1111

responded there

[cursive tinsel]

Land

[dusty trellis]

Any ideas on how this was made? How much is AI and how much is Blender or some other tools? https://www.youtube.com/watch?v=VZfqbSoaGO8

[desert dagger]

with a video camera

[unkempt grove]

One message removed from a suspended account.

One message removed from a suspended account.

[old lichen]

Hey
can you suggest any youtube channel or other resource where they work on games and show all sorts of interesting things in Stable diffusion?

[dusty trellis]

Was hoping for a serious answer

[desert dagger]

You can create those animations with any of the video ai models out there. But you are looking at a LOT of editing of video clips together to make the final video. It wasnt done in one shot

[dusty trellis]

I've never seen that much consistency from shot to shot. Unless I didn't notice repeats or reversing and other tricks

[desert dagger]

Its animated images. Either actual photos, or a 3D location done in something like blender

And its just camera movement

In fact, it could all be blender or unreal engine

[dusty trellis]

Right, that's what I'm getting at. It is not from any of the SOTA video models. The consistency is too good given all the different camera angles

[weary fulcrum]

@still glacier due to weird role names, not sure who has moderator rights here, but the link above probably spreads malware using cracked software

👍

[atomic mortar]

I mean not really, it depends on what your looking for

[atomic mortar]

Made some really really good friends thru discord. One from here even

I recommend trying specific servers on topics your interested in like dnd, anime etc

[fervent thunder]

hmm if you watch carefully the layout of the outside changes each shot

[ebon badge]

Is a rtx 4060 okay for using SDXL? does a 4060 Ti 16gb makes a big difference?

[fervent thunder]

you need only 1.5GB VRAM to fit SDXL in Q4_K_S GGUF format

[pure mango]

Hey, what is the best model that I can use to train with my artwork style to generate new artworks that will come out near identical with my artwork style?

[ebon badge]

idk but I have a 1660 super and SDXL is sooo slow

I don't exactly use SDXL but Illustrious

[fervent thunder]

low step method like hyper, lightning, DMD, TDD, PCM are all good

[peak holly]

Hello! I applied for a community license a couple of hours ago and was curious as to when I should get an email. There is no ETA listed. The form makes it seem like it is already active, but nothing has come through. I've checked spam, etc.

[atomic mortar]

Hi,

this is a community server

i would recommend sending a email to their support

if you require assistance for their products

[opal hedge]

Should be fine but you might struggle generating large images and using things like loras or control nets

And yes 16GB with the 4060Ti will make a massive difference

[wise rune]

I hope I'm not inserting myself here where I don't belong! I can no longer login to Stability Assistant despite having a subscription to the service, despite being able to use it for the first month or so. On login I keep getting redirected to a page that says I need to have "beta" access. The only clickable option is one called "Auth0" which only refreshes the same failed login page. I've filed a couple support requests and receive no replies except a survey that asks me how my support experience went.

[desert dagger]

you need to post this in the #🗣｜artisan-support-feedback channel

[dark hawk]

[reef agate]

hello

[lunar charm]

is there way to see what ai model/tool was used to generate an image.

[still glacier]

check an image's metadata. By default, most AI tools register in there the prompt + some infos and/or workflow.

[lunar charm]

no like a saw a video on YouTube. is there a way i can tell how was that video created (the images generated)

i want to know which tool was used to create those images in the video

[still glacier]

not really then

[lunar charm]

i tried making that but it doesn't geenrate that

[lunar charm]

is there a community for that

[still glacier]

for what ? to guess how someone did some AI stuff ?

[lunar charm]

to get maybe a more precise prompt for that image

or im using the wrong ai tool

[still glacier]

you can try asking in #📝｜prompting-help

[lunar charm]

alright thanks

[odd arrow]

Anyone could share a comfyui wf using Flux1D + ControlNet Depth ? a simple one...

[drowsy robin]

wouldn't you use flux.1 depth lora instead

Honestly I don't know which is more accurate

[green rover]

Hi, I'm used to Automatic1111 but want to use a Flux model which it seems I need Forge for- is Forge still being updated though? Which one is suggested nowadays?

[grizzled bison]

i would try tec h support

[pseudo nacelle]

Good day, guys. I wanna ask, does A1111 generate better results than Forge?

I tried to use Forge due to its performance efficiency, but I found worse image quality than A1111. Is there anything I can improve?

[jolly tartan]

tengo problemas con stable difusion me ejecuta con amd pero siento come muchos recursos para la tarjeta grafica que tengo que es buena, y aparte no me toma bien los modelos y loras alguien puede ayudarme?

[merry ginkgo]

Is auto1111 still relavent? Doing some spring cleaning and trying to tell if im good to get rid of it

That and need to sort 1.5TB of models lol

[pine path]

depends on what you use. I'd say it's outdated, but if you're only using 1.5 and XL models and care about certain extensions then you might still want it

otherwise comfy and stuff that use it have superseded it

[merry ginkgo]

Got it, honestly all I needed to know

[full hound]

hello everyone, new to the group and AI imagery prompting, love the space, glad to be apart

[atomic mortar]

If comfy looks intimidating i recommend swarmUI

Since you get a nice and easy to use interface

But also a optional comfy ui backend

[supple radish]

hello

[long owl]

Info on AMD's 9070 and 9070 XT seems scarce. I'm not really sure where to blab about this, but I have a 9070 and got it working in Automatic 1111. I was getting an error (something about Attribue 'dml'), which I searched around for, and.. The solution was to delete the venv folder.

As per some more searching around, I found someone say to use the "Commandline_args" stuff set as:
--use-directml --skip-ort --medvram --opt-sub-quad-attention --opt-split-attention --no-half-vae --upcast-sampling
This has worked for me. Though, generating images feels a bit slow!

Running the CyberRealisticPony model (one of the top results on civitai that I randomly grabbed), I generate a 512x512 image in about 14 seconds, and a 1024x1024 image in about 1 minute and 15 seconds.
Sampling method: DPM++2m, Schedule type: Automatic, Sampling steps: 30, CFG scale: 7.

If anyone has any tips, I'd love to hear them. I keep wanting to try ComfyUI (mostly so I can try animations), but it seems like the worst time ever with a new graphics card now. Annnd it just feels intimidating to get into 😛

[obsidian coral]

hello c:

[warm junco]

Hey you want to use ZLUDA over Directml. (On Windows)
Its much faster. Unfortunately there is currently no support for such a new GPU.
We only need to wait for the gfx librarys for it then it will work.

[long owl]

Yes! I had heard good things about ZLUDA.. Annnd saw that, at least on the github pages, the newer cards weren't supported yet. I'm looking forward to being able to try it though.

[warm junco]

I will update my Guides as soon as it gets supported

[modern zephyr]

hey cs1o I trying with your settings but is not working now , not sure if is why im using a olg gpu only 2 gb

[warm junco]

2gb ? Thats to less for the high resolution probably

[modern zephyr]

maybe better i find another model?

[warm junco]

Yep you should stay with 1.5 based models.
These are mostly 2gb and should work much better

[modern zephyr]

kk thank you so much

[warm junco]

Np, for example try this:
https://civitai.com/models/84586/aam-anylora-anime-mix-anime-screencap-style-model

[nocturne patio]

配

[visual path]

hello

[honest stirrup]

Hey guys, if I only want to upscale my images, would a GTX 1650 with 32GB of RAM be enough? My CPU is out of the equation since it's a laptop (AMD Ryzen 7 3750H, 2.30 GHz, 4MB L3 Cache, up to 4.00 GHz). Any recommended models?

[warm junco]

Hey, with that GPU you can generate aswell as upscale images

Its not the best card for these tasks but it will work

[honest stirrup]

Hi, thank you, what models or ui would you recommend? I'm using REST API to connect it to my workflow.

[warm junco]

Np, for webuis you should look at Forge / Reforge or Automatic1111

[tropic dock]

I am uploading audio into stable but keep getting an error. It is less than 3 minutes. Any tips?

[desert dagger]

what's the error and how are you uploading it?

[tropic dock]

I’m uploading an AIFF from my iPhone 28mb. Load failed

[desert dagger]

to what, specifically?

[tropic dock]

https://stableaudio.com/generate

[tropic dock]

Directly on the website

[desert dagger]

can you try uploading a wave file instead of an AIFF?

[tropic dock]

I will try to do that thank you

[desert dagger]

let me know if you still get errors

[tropic dock]

I tried it on desktop this time. Exported from logic as a .wav file and uploaded it. This time it went further and started processing. It gave an error upload processing failed contact support message

[desert dagger]

how large is it?

[tropic dock]

I28mb

28*

[desert dagger]

that might be too large.

try 32 mg and see if it goes through

also try mp3 format

[fervent thunder]

it might say on the SAI docs and it will be similar size and file format requirements for this stableaudio site probably

assuming that they keep the requirements similar

[desert dagger]

https://stableaudio.com/user-guide/audio-to-audio i can't find anything that does

[fervent thunder]

hmm ok not sure then

gonna start using their API for SD3 8B Large, I always loved that model

clown made loads of progress with SD3.5 recently though, gonna use that more as well

the main thing is a certain regional sampling method helps it a lot

[desert dagger]

can't wait to see what you create with this

[fervent thunder]

yeah I'll post here and on L3

[desert dagger]

:) looking forward to that

[fervent thunder]

I like Cogview4 as well although its unlikely to get Comfy support
cos Cogview3 was also good and kinda got overlooked
its the same guys who make CogVideoX video models

[desert dagger]

i ran it a couple times when it first came out, my poor little 4060 hated it

[fervent thunder]

ah yeah I rent 4060s a lot its a rough performance level

I'm been trying to port this to comfy its SD 1.5 but with blocks removed and then a bit more training to smooth over the cracks

https://huggingface.co/nota-ai/bk-sdm-tiny

[dense token]

I read the repo and while interesting, I have to wonder who this benefits. TE + VAE + U-Net still comes in at 1060 MB.

You'd also have to make brand new tools to make finetunes of it.

[umbral yacht]

Any recommendations for a model for D&D maps (img2img/sketch2img)?

[fervent thunder]

I found an even smaller one
https://huggingface.co/cqyan/hybrid-sd-224m

and they worked on the VAE as well this time

they made a "tiny" vae https://huggingface.co/cqyan/hybrid-sd-tinyvae

[desert dagger]

just use freeU

[desert dagger]

SD 1.5 does those really well

[umbral yacht]

just straight up rawdogging 1.5 or any particular variant?

[desert dagger]

i just used base sd 1.5 when i did some of those.

[umbral yacht]

alright, I'll try

[fervent thunder]

the base SD 1.5 is very fun but also pretty chaotic

I use realvis mostly for SD 1.5 because that is what PowerpaintV2 and Brushnet were trained on

the idea of these tiny SD models I was linking is just to speed things up really
they are mostly just block removals anyway

[dense token]

I once had a tiny 324mb (iirc) SD 1.5 based model (Full U-Net unlike the stuff you linked) and it worked amazingly well despite the tiny size.

To this day I have no idea how it was done. And it was a full model, not just inpainting.

[tacit fox]

hello

[fervent thunder]

wow nice
yea I wanna explore models around 0.3B mark or so

[pulsar marsh]

ay is there any significant difference between a stock checkpoint with Loras, and a checkpoint specialized in what those Loras would have done?

[faint sparrow]

little off topic but what adblock are you guys using on chrome since ublock got removed?

[dense token]

People still use the Google Spyware app?

[faint sparrow]

ok, just installed brave...fair point

[fervent thunder]

lol am on brave too

but edge for comfyui

[cunning copper]

Hi guys

I have a 3060 12GB vram with 16GB of ram. can i run flux? if yes then exatly what would be the correct way to install and get it working? I tried to follow some tutorials on yt but got it messed up so trying to do a clean install again. need a bit help here

[barren mountain]

16GB RAM is the bigger issue, if you can increase to 32GB RAM then it will be much better, I'd probably start by trying the Q4 GGUF, make sure you're using FP8 T5XXL

[cunning copper]

thank you! i am planning to upgrade my ram couple months later i think

[cunning copper]

youre referring to this one right? https://huggingface.co/city96/FLUX.1-dev-gguf/blob/main/flux1-dev-Q4_0.gguf#:~:text=/-,flux1-dev-Q4_0.gguf,-city96

[barren mountain]

yup, then in Comfy you need to use the GGUF node to load it, I don't really mess with GGUF much though

[fresh ruin]

if i want to train a lora based on my own style from my ai generated images, should I use the upscaled versions of them or the original raw generated ones?

[fervent thunder]

depends on budget

training lower resolutions is mostly to save money

[fresh ruin]

i do it locally

[wide thistle]

hello

[fervent thunder]

local cost is electricity costs but if you dont mind then use higher res maybe

[fresh ruin]

you know higher res images get downscaled with bucketing on kohya, right?

[abstract quarry]

only if you want them to be downscaled

if your question is: should you use downscaled highres images or lowres originals, I would say use the original ones 🤷‍♂️

[fresh ruin]

yeah... grok/chatgpt also said original might be better

cause upscaled might lose details, etc

[dusk carbon]

Hello everyone, I'm new here do anyone know what's the stable diffusion version of Kling AI? I'm looking for Video AI generator😁

[fervent thunder]

IDK I don't use the tools like koyha

been slowly working on a new training script for like a year now lol

its likely a changeable option on koyha thjough

[oblique elk]

Would say for local generations huyuan video or wan

[abstract quarry]

Wan 2.1 is really good

[vital nebula]

yo hows it called picutres that arent portraits?

like a character on screen not that close to the camera doing something

i can't rememeber the word

[ruby lagoon]

for a 5070 ti and 64 gb ram, should I use flux fp8 or fp16? (for generation times less than 10 seconds)

[warm junco]

Fp8

Less then 10 seconds won't work lol

Maybe with flux schnell

[torpid kraken]

is there a way to transfer hair styles from a reference image?

[wicked jungle]

Spoonkid is that you

[latent rain]

Hello everyone! I apologize for a probably stupid question, but why has sd3.5 almost not gained popularity among users? Is it really behind sdxl in many ways? I would like to know the opinion of those who understand this, is sd3.5 still an excellent architecture, which is much better than sdxl or is it another intermediate model like sd2.1 was?

[arctic rock]

hey guys , i have only a 2070 super. what dou you think what model i should use ? I use forgeui , flux dev version is a little bit to high

[hasty badge]

Every time i read about SD3.5 lots of nonsense about it is spread and people don't even try it anymore. It just never gained momentum because many things flux did already better.

But the outputs of 3.5 can be great, much more real looking than flux or any other model, it just follows prompts worse and suffer a lot from weird ai artifacts, malformed elements in generated images and such.

It might or might not be better trainable, but no person can afford to fully train/finetune a model that size, what everyone does with the little lora's is overfitting a style or concept, and flux does that amazingly

[warm junco]

hey, sdxl based models should work fine

[desert dagger]

people get in ruts. 3.5 is extremely trainable - we made sure of that before we released it

But the outputs of 3.5 can be great, much more real looking than flux or any other model, it just follows prompts worse < that's because flux was stuffed full of women, dogs, anime cat girls, and fantasy, then DPO was run on it to ensure that the most common images people are going to prompt for are going to look great. that's also why you have to break flux to get it to do anything else. flux is pretty much just a giant lora

[arctic rock]

okay, thank and i think none of the flux models right ?

[warm junco]

Nope with 8gb vram its no fun.
You could try the nf4 model or q4 ggufs but its not worth it

[hasty badge]

That's a way too rose tinted glasses look comparing flux and sd3.5, oftentimes 3.5 downright breaks and fuses subjects in a prompt together where flux does not, flux picks up style/subject training in small resolution extrapolating it to higher res implicitly which 3.5 does not, picking up from just a few images stupid fast. For practical purposes flux is just better suited, where's the sd3.5 where "the most common images people are going to prompt for are going to look great"

[desert dagger]

that's insider information

you are guessing, i'm telling you facts not guesses

[arctic rock]

okay, is there cloud solution available ? like replicate ? can i do there everythink like i wuld run it local ? because i dont want to buy a new pc ?

Or is the difference from flux.1d to sdxl not so big ?

[desert dagger]

you can rent gpus on places like vast.AI

[hasty badge]

I like sd3.5, if it does right it's better than anything out there, but it's just soooo random, one prompt can give 3 different styles despite clearly referring one style, and out of the 4 interpretations 3.5 gives for slightly hard prompts 3 ignore or fuse key elements, or has the crazy artifacts like malformed hands or repeated little elements. For the few gens that come out right, 3.5 is amazing, but so few do :/ For reliably getting not so generic images, sd3.5 is a chore to get them out of the model sadly, and it's easy to understand why other models are preferred

[desert dagger]

odd - i never have a problem getting exactly what I want out of it

[cedar salmon]

low standards

[desert dagger]

do you try to prompt all three encoders with the same prompt? or do you prompt them to their strengths? are you detailed and specific? or what do your prompts look like

[hasty badge]

lately just all at once that's easier and makes not a whole lot of difference except for style. But let's agree to disagree, you clearly love 3.5 and i can understand why as the good images look so nice.

[desert dagger]

you'll get better results if you prompt each of the encoders - whether you use flux which uses 2, or sd 3.5 which uses 3 - to their specific strengths instead of just using one prompt that they don't understand in the same way

[warm junco]

Yes there are sites like vast.ai and paperspace.com
Flux is good but some good sdxl models can also create very nice images.
It depends on your needs.
For example flux is bad for anime images.

[arctic rock]

its very expensive , 0.5 - 1 $ / h . whats with replicate ?

[fervent thunder]

vast.ai is cheapest so it is what it is really

replicate are what they call serverless

[arctic rock]

okay have some experience with vast.ai ?

is 24h one day or 24h the time i really using the gpu and generate img?

[fervent thunder]

yeah its what I use

not sure how the billing works

its either by second, by minute or by hour
but its for how long you have the server rented

so you pay for idle time

[hasty badge]

if you just want to gen images cheap, look at runware.ai but you have less control than using your own comfyui instance

[arctic rock]

i using midjourney , but i need consistent character so i want to change my strategie

installed forgeui and flux.1 but my 2070 super is to weak

[hasty badge]

there's also comfy-nodes-in-the-cloud by siliconflow which seems mostly free to use for now https://github.com/siliconflow/BizyAir

[arctic rock]

i tought replicate would be the best for me , because you also can lora model training for consistent character. I tought you have some experience with replicate

[fervent thunder]

I forgot runware, they are indeed cheapest in the world for pre-made serverless API

[arctic rock]

thanks

[fervent thunder]

they are cheap cos ran by solar, on custom liquid cooled L40s apparently

[fervent thunder]

replicate or fal would be good for you yeah

I don't use them because I do custom stuff that they won't have

[arctic rock]

can you give me an example?

maybe sdxl is good enough for me. can i also do model training for consistent characters?

[fervent thunder]

yeah you can train loras on there but I am not sure for SDXL

they only host a small number of models

[desert dagger]

if you're going to train models, you might look at what luca taco has on replicate - his trainers are very good

[arctic rock]

sorry for those many questions , i am a beginner. -_-.

when i run forge ui local with sdxl, its the same with lora (character) training with Flux ?

[pine path]

medium is too small to train. Large released after flux did and was seen as inferior. Large is still pretty costly to finetune outside of various loras even though it does learn much better than medium.

and just in case you're wondering why I say medium is too small even though it's the same size as XL, it's due to the VAE. It requires a bigger model to compensate, otherwise training becomes much slower. It may sound backwards, but it's cheaper to train large than medium.

[desert dagger]

medium is not too small to train, but medium really shines if you use it as a refiner rather than your base

[pine path]

It can be finetuned, but it's going to take a ton of compute due to its size

[desert dagger]

sure - but it's designed to be artsy - with the intention really that you'll use it to enhance with. and generate with large

[hardy sedge]

Can you use sdxl/pony with ltx video or hunyuan? If not, what is the best way to create videos with sdxl/pony at the moment? Thanks.

[atomic mortar]

create image with either sdxl/pony > image to video

otherwise not really a thing. maybe animate diff but yike

[abstract quarry]

yes. You can also try ipadapter with sdxl. Sometimes this is sufficient to generate consistent characters without the need of training.
But if you want to train, SDXL is definitely easier to train than Flux or SD3.5 although results are a bit inferior

[atomic mortar]

Spam

[latent rain]

Tanks, anon. Thanks, anon. I've only used it on hf space, since there's no webui, and comfy is too hard to use. Yesterday I went to reddit to read about sd, and there was a thread about sd3.5, in which there was not a single answer in favor of sd3.5. And the model is completely undertrained, and the results are terrible, and they released it only because Flux came out, and that sd3.5 can't even come close to FLUX in quality. And the most important thing for me is that someone there collected a bunch of pluses by writing that sd3.5 is not suitable for training at all, choose sdxl. I was very confused because I remember reading the same thing about sdxl in comparison with sd1.5 six months ago. That's why I decided to ask

[abstract quarry]

fair enough. I also remember how many people said sd1.5 would be much easier to train than sdxl while I found sdxl much better.... I guess one problem is that people compare on their already optimized workflows that are just not adapted and tuned on new models

I also think that Flux generates better images most of the time. However, it's totally possible that sd3.5 just needs more careful prompting

[fervent thunder]

I tried SD 3.5 on the huggingface demo the other day and it was 10x better than the outputs I was getting in comfyui

so there was something up there

[desert dagger]

so that makes me ask, what sampler and scheduler you're using in comfy - what's your cfg set to - and what's shift set to?

yeah, there are discords for that, not this one. not what we're here for

[fervent thunder]

tried the default comfy ones

[atomic mortar]

Another spam post nice.

[slender mango]

Hello good souls, am new on stable diffusion, am originally an artist that like to train my artstyle on an ai, but I couldn't know how things works, if anyone can help I appreciate it, I use stable locally

[atomic mortar]

Hi, if you got time tomorrow i can show you how i make my loras but im off to bed rn

But feel free to send me a reminder or a friend request. I made a few style loras before

@slender mango would help if i actually replied instead of forgetting to tap it lmao, yeah read above

[rich dock]

does anyone happen to know if i can delete files in my huggingface cache folder

its swollen up to 118.6GB now lol

no idea what i can or cannot delete

[warm junco]

Open up a CMD and run
Pip cache purge

Then recheck

[rich dock]

if i do this does it auto-download the cache files it needs again as and when it needs them?

do i run it from that folder or anywhere

also the uhh pip cache is a separate folder (20.4GB) to the huggingface folder

[warm junco]

Yes

[rich dock]

what about the huggingface stuff thats a whole different folder and much bigger

also: thank you!

[wild dock]

Hi everyone, just wondering does anyone sell ai on website like teepublic or etsy ?

[atomic mortar]

People definitely do but i think it's more common for people to do patreons and commissions

[warm junco]

You could delete it and the necessary stuff should be downloaded when needed again.
It stores stuff like clip models etc

[rich dock]

there's a lot of cogvlm models, etc etc
model-00005-of-00008.safetensors, (4.x gb each)
pytorch_model-00001-of-00002.bin (9.3GB)
diffusion_pytorch_model.safetensors (10GB)

etc etc

llama etc

[warm junco]

Yep it stores there stuff for every webui

Do you use comfyui or auto1111?

[rich dock]

i have swarmui, A1111, forge and foocus installed

as well as a ton of other stuff like face fusion and joycaption and supir

[wild dock]

patreon and commission ? how does that work ?

[atomic mortar]

Either people want custom lora's, workflows, tutorials etc

Or people want specific images but in bulk of of a certain character so its less per image then a artist etc

Or you use AI as a tool and edit the image further so you could pawn it off as handwork but its definitely not ethical

[wild dock]

oh ok, I didn't know that it exists, and like... there is website for people selling custom lora ?

[atomic mortar]

Hmm mostly patreon and people posting images/ some loras on civitAI

And then in its post/description "check out my paid loras @ "

[wild dock]

ok gotcha thanks for the answers, I might stick to making my next millions selling ai for t-shirts 😆

[warm junco]

Ah okay, then it makes sense that the huggingface folder got that big

[hasty goblet]

anybody available for paid work please dm me it’s about headswap with style transfer

[atomic mortar]

So a face swap onto a stylized image?

There's like so simple I'd feel bad charging

Kinda busy though

[pine path]

new paper dropped

DiT-Air: Revisiting the Efficiency of Diffusion Model Architecture Design
in Text to Image Generation
https://arxiv.org/pdf/2503.10618v1
We evaluate a range of DiT-based architectures–
including PixArt-style and MMDiT variants–and compare
them with a standard DiT variant which directly processes
concatenated text and noise inputs. Surprisingly, our findings reveal that the performance of standard DiT is comparable with those specialized models, while demonstrating
superior parameter-efficiency, especially when scaled up

MMDiT is garbage

[turbid adder]

hi (:

[abstract quarry]

hm, it reads like they do mmdits, too

they get rid of the dual stream blocks and reduced parameter counts by a lot, though

[fervent thunder]

there's so so many new architectures now anyway

[desert dagger]

a city full of strange virtual bulidings

[dense token]

It sounds like you've found The Thirteenth Floor

[signal field]

how to use

[fervent thunder]

How can you tell if AI will generate your stock images and turn them into your own design?

[pure musk]

Hey all, what is the worst part of your workflow right now?

[floral umbra]

Not fully utilizing multithreaded asynchronous startup and model loading here and there.

I've had GPT write me a multithreaded optimization so it starts up comfyui around 3-5 times faster. but it's not fast enough

[dense token]

Lack of Switches, because no one has released a pack that is only the switches. They lump all sorts of useless crap together.

[onyx fox]

I am looking forward someone 📌

✅ Experience with AI-driven automation & CRM integrations.
✅ Strong Python skills with API development knowledge.
✅ Familiarity with lead scraping, data filtering, and automation workflows.
✅ Ability to work with no-code/low-code tools like Zapier & Make.com.
✅ Previous work or portfolio in AI lead generation or automation projects.

[shrewd lichen]

Hi, i got a request if someone would be avaiable to help me create a img2img stable diffusion inkpunk style for my friends birthday (1 pic).

[desert dagger]

it won't. it's not supposed to. it will use any image you give it as a visual prompt, and usually do something that uses the color pallet, and shapes - to determine what colors to use and where to put the elements you prompt for in your text prompt

[desert dagger]

there is a #1092446741984444416 form for this

[true lichen]

ban it pls

[fervent thunder]

wow the Stable Virtual Camera model is incredible

[crude vector]

anyone knows how safe is Vencord to use?

[fervent thunder]

I was gonna say you could make gaussian splat with that many views
but then later in the paper they did

[fervent thunder]

not sure

is it needed?

[crude vector]

its just extra features on discord using plug in

nice to have

[fervent thunder]

Client modifications are against Discord’s Terms of Service.

However, Discord is pretty indifferent about them and there are no known cases of users getting banned for using client mods! So you should generally be fine as long as you don’t use any plugins that implement abusive behaviour. But no worries, all inbuilt plugins are safe to use!

Regardless, if your account is very important to you and it getting disabled would be a disaster for you, you should probably not use any client mods (not exclusive to Vencord), just to be safe

Additionally, make sure not to post screenshots with Vencord in a server where you might get banned for it```gonna say probably not

[crude vector]

that sounds like a yes

[crude vector]

where did you get that

[fervent thunder]

https://github.com/Vendicated/Vencord

[crude vector]

i see

thanks

[desert dagger]

there are discord bots that provide plenty of extra features for the admins that want them.

[desert dagger]

@vapid dove and you might want to remove this - he spammed it on every channel he could

[signal mulch]

Any experts regarding AI Video in here?

[vapid dove]

more links for the ban pile

[gentle prairie]

can you pause a generation directly from the CMD? I closed the browser and letting only the cmd working in the background, but i need to pause it, is there a command to do that?

[wild dock]

Ctrl+c ?

That will kill the process btw not pause it

[gentle prairie]

well that's not what i need, haha, if i wanted to kill it i would simply close the CMD

[wild dock]

Ctrl+z seems to pause the process

[desert dagger]

you can't pause a generation, sorry

[gentle prairie]

okay, i find this randomly but... clicking into the CMD windows seems to pause it... and pressing enter seems to make it run again... if that's not it, then what does doing that do?

[desert dagger]

you would have to write some sort of manager that could interrupt the process and then restart it

[desert dagger]

you're not seeing the generation in the command window, you're just seeing logging - all you're doing is interrupting the text displaying in the window

which interface are you using to generate in?

[dusty trellis]

When I generate images with Flux, that are in broad daylight, like on the beach or in the desert, they always come out blurry. What can I do to fix that? More steps? Prompting for darker conditions?

[desert dagger]

flux is very artsy in what it does, and you're gonna get the shallow depth of field shots almost all the time. there are some loras people have done that combat that - you might see what you can find on civit

[gentle prairie]

auto1111, but as i said, i closed the browser, so i can't interrupt that way, i launched multiple of them too (queued)... But it's weird cause when clicking in the CMD, my fans stop running, implying the generation stopped

[desert dagger]

personaly, i recommend not using flux and using stable diffusion 3.5 large

[desert dagger]

you might want to ask about this in the #🤝｜tech-support channel - that's where the auto1111 tech users normally are at

[wild dock]

https://www.orionnetworks.net/control-z-and-control-y/#:~:text=Pressing the Ctrl %2B Z shortcut,hit the Ctrl %2B Z again.

When you clixk on the command line the output is blocked not the background process, unless i guess the background process is waiting to write on the console https://superuser.com/questions/459609/what-does-it-do-exactly-if-i-click-in-the-window-of-cmd

[dense token]

The only one I know of that can pause gens (and actually works) is Easy Diffusion by cmdr2, but it's so outdated and slow.

[desert dagger]

i.e. you are interrupting the text displaying in the window

[outer pecan]

Does anybody else have trouble with numpy on stable diffusion deforum ~ collab notebook?

[tacit bronze]

Did anyone get the model realhotspice_sdxl to work? I'm trying to do gens with diffusers but I just get noise.

[snow parrot]

Hello

[void horizon]

hello

[steady atlas]

Yes bro my compiler is not working too..

And thanks guys for giving me a new tws of realme

[steady atlas]

To realme

Actually I am in need of a mobile
Of Nothing Phone 3 A

[devout crest]

generate a really funny picture of a friendly and happy ghost that is sitting in a chair while being interviewed for an executive job

[sudden schooner]

What kind of prompting tends to work well with flux?

[atomic mortar]

👻

#artisan-faq

[atomic mortar]

natural prompting and long winded ones

[atomic mortar]

stuff like
"In this captivating cinematic landscape painting, the canvas is filled with a serene yet dramatic scene. A large and vivid red sun dominates the dark, almost night-like sky, casting warm hues over the darkness below. In the foreground, a gnarled tree with white foliage stands proudly atop a rocky cliff, overlooking a cascading waterfall that plunges into a misty abyss. The striking contrast between the bright red sun and the dark, moody tones of the landscape creates an enigmatic atmosphere, filled with mystery and tranquility. The waterfall adds a dynamic element, suggesting movement and the passage of time amidst the stillness of the tree and the sky. The painting masterfully blends elements of fantasy with a touch of realism, evoking a sense of natural beauty and contemplation. The overall composition is a breathtaking exploration, painting, cinematic"

[sudden schooner]

Great thanks

Any artistic style flux is particularly good at?

[atomic mortar]

imo i dont use flux a lot but i recommend checking civitAI and then on the images tab for inspiration

[sacred osprey]

Hello\

[signal mulch]

Hi, i am new here. Can someone tell me why so many people use ComfyUi instead of Pinokio? Seems so much mire diffucult and complicated... 🤔

[pine path]

Comfy is constantly updated to support newer models and has a lot of community support, including UIs that use it

[upper tide]

Hey what's up guys. I want to create realistic images from my own photo. See I know nothing about image generations I tried some free online tool But the result were low resolution and kinda looked fake. I want some legit thing something pretty cool.
Can you just show me some website also something I should read first to understand the tools
I will do all the work Just jump start me

And yeah Free ones lol... Not paid

[desert dagger]

try mage.space. they have a very good free tier with a lot of options

[upper tide]

Alright. ..Thanx..

[warm junco]

Pinokio is not a image generation webui.
Its only an installation script for tools.

[dense granite]

Hey everone 😄

[barren glade]

Hello :W

ah man messed up the first emoji immediately, this does not bode well 😦

[plain drift]

How long does it take y'all to generate images? Just wondering

Ik it depends on word count and dimensions

[pine path]

@vapid dove scammer bot

[signal mulch]

I know but it seems so much easier to use flux webui or forge and fooocus or SD through pinokio.

[exotic sphinx]

Depends on settings, can be a few seconds, can be minutes

[dense token]

About 30 seconds for a 1080x1920 image (or vice verse) on my 3060TI (8GB VRAM)

[plain drift]

Damn 30 seconds

[ruby oracle]

Hello world

[dense token]

That's actually a lot of time considering I'm asking the model to do stuff it was never designed for.

[plain drift]

With my laptop takes me up to 40 min

Sometimes 3

[dense token]

I've never tried it on any of my laptops, just my desktops.

[plain drift]

3gb vram works but is painfully slow

[dense token]

ComfyUI portable version, and SD1.5 models (good ones like DuchaitenJourney or Aniverse v40 or AImaginationX1024)

[plain drift]

Portable as in its for mobile?

[dense token]

Portable as in you don't have to mess around installing python or messing with venv

[plain drift]

I see

I didn't have to use python for forge

Should I switch to comfyui?

[dense token]

You're already using Forge, but I have noticed speedups even with ComfyUI vs Forge

Not to mention you get so many options for custom nodes / samplers

[plain drift]

Aight

[silver juniper]

Hey all. I have a question about Wildcards. I have some artstyle wildcards, but they are broken down very specifically into sub categories. Is there a way to run the wildcard and have it pick from any of the sub categories?

[dreamy sorrel]

Hey everyone! 👋 I’m looking for a way to add text to AI-generated images where the AI can automatically pick a font, position, and vibe that matches the image’s style. For example, a bold, artistic image might get a serif font in a contrasting color. Does anyone know of a tool or script that can do this? I’ve been using Canva manually, but I’d love something more automated. Thanks! 😊

[umbral yacht]

Is there anything I could feed a D&D map into a model to get a "in the scene" view? I would then like to give that to sora to turn it into a cutscene type video

[arctic patrol]

Hey all! While looking for the best AI engine for producing images for a children's reading app I'm developing I recently stumbled upon stable diffusion. I've been pretty impressed with what I'm reading so far, but to honest I am a bit overwhelmed.

It's seems like there's been a lot of advancement in the last few years. What's the best place to read up on the most up to date best practices? Is an online service just as good, or is running locally with Automatic1111 the way to go?

if locally is better, I typically do development on a m3 Macbook with 36 GB Ram, but I also have a gaming pc with an i7-14700KF CPU/4090 GPU/32GB.

Is my macbook sufficient or would I be better off setting this up on the powerhouse PC?

[snow parrot]

Set up on your power house PC

[mighty monolith]

hi

[river bolt]

I started by downloading the Krita paint program, and then the AI Image Generation plugin for it (it's in the Github repository hosted by Acly, a search brings up a lot of imitators). It will download and install what you need to get started, very straightforward (relatively) and the community is pretty helpful.

[atomic mortar]

5S no loras
13S with 10 loras with a 5080 SDXL

Flux similar times

Embed fail

[dense token]

"Show off!"

[atomic mortar]

Lmao 9mins for text to video though with Wan

9 to 30min for image to video depending on length

[atomic mortar]

I spent a lot of money for bragging rights. Still. 5090 gets better times

[dense token]

If it wasn't for my crappy electric box I'd be able to use my 3090. :(

[fervent thunder]

I use CPU sometimes

[abstract quarry]

a 4090 is perfect for local generation.
automatic111 is a bit out I think, there are better forks like forge or reforge. I always recommend InvokeAI, it's the easiest to use tool and contains a lot of features builtin without any plugins. If you want maximum control, comfyui is the way to go

regarding models I would still recommend Flux, as it has best prompt understanding. It is also capable of doing multiple panel-images (i.e. several images together that show some kind of connected scene like a comic strip)

however, SDXL still has the better tool support regarding ipadapter and many different controlnets

[wild dock]

quick question, on civitai what is difference between merged and trained checkpoint ?

[shut otter]

Can someone help me find a yolo.pt file for detecting full body. I am really struggling to find it 😦

[atomic mortar]

Merged is what the name implies such as wai-nsfw & NTRmix mixed making it something else etc

[wild dock]

Ok so merge is just merging already existing model or lora and trained is additional training kn an existing model or lora, yeah as the name implies, sorry i feel stupi now 🥲

[atomic mortar]

No worries! This ai stuff can be pretty confusing so i get the confusion

[high olive]

Hi all, new to the world of image generation and AI in general. Which is the correct channel where to ask for help about which models to use for specific results, please?

[arctic patrol]

Thanks man, all of this is helpful I'm a noob to this. Going to dig in really today

My main goals are having pixar-like characters that I can reproduce participating in different activities/environments for each "page" of my stories.

I've only tried using midjourney so far, while the it creates high quality characters. It falls short recreating reproducible creaters even with some of their newer features that are supposed to do so. The art style or characters themselves tend to drift off their original prompts

I'll try invokeAI per your recommendation first

[abstract quarry]

making consistence characters it not easy. With SDXL you have IP-Adapter that allows you to transfer a character from one image to another

in Flux you can use multi-panel images (generate multiple images in one) to get character consistency. Of course, this won't work for a large number of images. Inpaint might help. Otherwise you have to train on a character

if you want to generate a lot of images, training a lora for each character might be the best option

[arctic patrol]

Yeah, I figured I would have to do some kind of character training. The stories for youngest age group will be picture rich so I may need 15-20 images of a single character (and may possibly work on series of stories in the future).

[wild dock]

oh god in 1 day I might have a 3090 🙂 🙂 second hand on ebay, bidding ends tomorrow

CAN'T WAIT

[signal mulch]

Same here, for how much do those sell in your country?

[wild dock]

UK asking price was 410£, winning the bid with 476£ 1d3h left 🙂
PNY GeForce RTX3090 XLR8 24Gb

[signal mulch]

Here its impossible to get a used one for under 700 € 😡

[wild dock]

where do you live ?

[onyx sandal]

Hey guys, I am generating 2d img2img pixel sprites for a daggerfall mod and I usually generate them with a white background, so that the AI doesn't get confused by "the void". I usually use websites with a background removal feature, however they always feather the edges a bit and I have to clean the edges of the sprites manually afterwards. is there any way to do this with ai where it doesn't feather the outer edges?

[devout vapor]

https://www.bbc.com/news/articles/c0kgydkr516o

[outer pecan]

Is it still possible to use SD Deforum with Notebook Collab?

[signal mulch]

Germany

[wild dock]

Ok, you know what, that makes me wonder, because if i compare to other cards being sold 3090 (mostly founder edition) they might start at the same bidding price but they go quickly up around 600£, so i got some doubt, i mean he looks like a good seller (100% reputation, he sold computer part before) but why no one is fighting for the bid

[signal mulch]

Usually they wait for the last minute.

[dark sorrel]

A picture showing Two main types of bone include spongy (trabecular or cancellous) and compact (cortical) bone. osteons in compact bone and trabeculae in spongy bone. Figures showing osteons in compact bone and trabeculae in spongy bone, including osteogenic cells, osteoblasts, osteocytes, and osteoclasts and blood vessels.

[river bolt]

You can try combining actors/actresses and adjust their weights to find a look you like, those usually generate fairly consistently.

[river bolt]

Try the free Krita paint program, and install the ai-tools plugin by Acly on Github, click on your contrasting background, ai will figure it out and select it for deleting. Or click on your subject and invert the selection to target the background.
https://github.com/Acly/krita-ai-tools

[onyx sandal]

Thanks! I'll try it out!

[sacred veldt]

Hello there I just joined the server, I was wondering if I could find tips on fine-tuning my models here?

[echo kite]

same

I use CPU all the time

[river bolt]

https://tenor.com/view/austin-powers-dangerously-gif-20939086

[echo kite]

since my graphics card model is 5 years old

curious how long it takes yall gpu users to generate images

it usually takes like ~3 minutes for a sd 1.5 512 image for me

maybe 2, idk would have to do a definitive benchmark

but yeah for cpu users it's not images per minute it's minutes per image

luckily I have a decent 16 thread cpu

I think 6s/it was one of the fastest I have gotten recently

[river bolt]

An SD 1.5 512x512 generates a few every few seconds, if that gives you an idea.

[echo kite]

how much vram do you have?

[river bolt]

I am grateful for the 16GB my 2080 has, it's a beast. But it's an older card and I know the newer ones are probably doing crazy things.

But 6-9 GB gets the job done too.

[dense token]

Here, this should be all you need. https://github.com/Nerogar/OneTrainer

[floral umbra]

I don't recommend using A.I gens for commercial products as it's quite frowned upon. As all the popular image gen models are based off of other people's images/art mass harvested from the web. Plus one can easily tell A.I art from actual art, as none of the models knows anything about anything it generates. Doesn't know human anatomy. All is hallucinated forth.

[sacred veldt]

Thanks!

[floral umbra]

If you want one for FLUX, ai-toolkit works incredibly well for that as well. Nailing my weird obscure loras immediately

[dense token]

OneTrainer works for FLUX

[dense token]

"Doesn't know human anatomy" I beg to differ, as my models are quite capable.

[floral umbra]

Nice. though isn't it all based off of diffusion pipe? Or what makes one better than the other?

[floral umbra]

You'll always get a slight off here and there. Sure, fine tune long enough, and it gets it better each time, but everything is still guessing and hallucination by the model :P

Take GPT. Same shit. Spits out random nonsense and think it's correct.

[dense token]

You should get better models, as there are some good ones out there that need zero loras or other extras

[sacred veldt]

I was using kohya to make SDXL dreambooths, I doubt my machine can handle FLUX. What I'm using it for is generating potential product images but the backgrounds always seemed grey and not the right color, the products look good just the backgrounds are always wrong

Token emphasis helped but still 1/6 generated images have the desired blank background

[floral umbra]

do name them and i'll torture test them lol

Cause so far, all models are hallucinating forth what you tell it to with less than 90% accuracy Always something off somewhere lol

Even GPT4 requiring several 100GB vram is a complete idiot

[river bolt]

I see both sides, but I'd say refinement has become the art.

To make it look "not AI generated".

[floral umbra]

Indeed. What we have yet to get though, is to make a image gen model that doesn't mix what's already has been thrown into it's blender porridge of vectors, but actually create never seen before art and artstyle.

[river bolt]

I've seen some pretty crazy ish from ai lol.

[floral umbra]

but, sadly as all image and textgens so far has been taught from top to bottom, aka you fill a infant's head with worlds knowledge, but it doesn't know how to properly use any of it in a coherent correct manner. Instead teach models from bottom up. What is what, why is what, then give it materials to learn on. That's when we'll get a 100% correct model. When it knows what it makes with scientific accuracy.

[river bolt]

I think your infant analogy is on-point. Give them five years.

[primal current]

Is there a ROPE deepfake based repository that can work in bulk? That tool is incredible, but I have to do everything manually.

[tepid matrix]

Dreams feel fake till you make ‘em real,
Move smart, not just with steel.
Life’s a game, play to kill." 🥷🔥🚬

Active 👋

[river bolt]

That was Elvis, yeah?

[dense token]

Someone wants Skynet...

[arctic patrol]

So is programming using ai assistance unethical because it’s trained on other people’s code?

[floral umbra]

I mean, literally what we got today, just minus the intelligence part of "a.i". Remember naruto when he cast his first shadowclone, and it was literally just a brainless "meatbag"? That's A.I today. "attempted intelligence", when there's practically zero intelligence, 100% guesswork

And it's literally skynet people aim for. Smarter and smarter "dummy's" untill someone eureka's with a 100% emulated/simulated brain that learns just like a human does, but with a few billion times sharper mind.

[river bolt]

But think of the cool pix we could make.

[dense token]

Is that before or after the nuclear armageddon?

[river bolt]

Right before, I'm guessing, I think I will be un alive after that.

[floral umbra]

Yep! where the model doesn't just have vectors of a shape knowledge bank, but know that humans has 5 fingers, and know that bodyparts doesn't float or clip through things straight off the bat xD

[river bolt]

My gauge to when to start worrying is when they understand a hand has only five fingers, then it's Skynet basically.

[dense token]

The models already know this but due to training loss, they can't remember.

[river bolt]

Does that mean it gets too much info, too much reference

[jolly arrow]

imagina un dragon volando

#imagine un robot bailando

#✍🏼｜rules-and-tos

[prisma aurora]

Hey up all...

[sick stream]

Hey

[dense kayak]

Hai- is anyone awake?? I need some help ):

[still glacier]

don t ask to ask, just ask. Otherwise nobody is gonna come forward

[quaint trellis]

Is the 5070 a good card for stable 3.5?

[gloomy ermine]

guys i have a question is a 5 second video really the limit for wan 2.1? I use a 4070S

[tulip shard]

hi everyone, i'm new to SD and i want to generate anime illustrations, are there any tutorials/resources for prompting/settings/models?

yeah that's a bot

aaaand it got bonked

[gloomy ermine]

gpu specs and ram?

and vram?

[tulip shard]

Intel Core i5, 8GB ram, Nvidia Geforce RTX 4050, don't know how to check the vram

[gloomy ermine]

you can try starting with sd 1.5

its faster

but if you want good images

try illustrious

i started with this https://www.youtube.com/watch?v=dMkiOex_cKU&t=1017s back then

if you really want to get the most out of your gpu time wise I'd switch to comfy ui as soon as i get a proficient understanding of stable diffusion

but the cons is you need to watch tons of tutorials cause comfy ui is a build yourself a custom generator kind of thing unlike automatic 111 which is the stable diffusion default you don't need to because all of the tools are combined into one

[ionic wraith]

Hi, does anyone happen to have a good workflow for flux and a lora?
I tried 2 but they require files from rgthree-comfy which doesnt work on my server, "Charmap error"

[tulip shard]

thank you for such a detailed guide! i will look into it

[abstract quarry]

there is no "stable diffusion default" 😅
you can also try swarmui, it's comfyui with a better user interface

8gb ram should be enough for Illustrious

[abstract quarry]

just use a default workflow and add a load lora node ...?

[dense token]

8GB of VRAM is more than enough for almost any SDXL model, I can generate 4K images with it.

[fallow laurel]

I'm looking for a few people to help collaborate on generation for a music video, let me know if you are interested and we can talk specifics! Thank ya🙌

[hollow mulch]

hey

[gloomy ermine]

yeah what i meant is the default ui new users use since automatic 1111 is the most common 1 hence default hahaha

damn my comfy ui broke AttributeError: module 'pkgutil' has no attribute 'ImpImporter'. Did you mean: 'zipimporter'? any one has a fix for this i already tried the pipi upgrade commands still nothing

[hollow mulch]

Hi everyone, can you help me out? Let's say I generated an airplane from the year 1940, and another one from 2025. Which ControlNet model should I use to make sure both of them have the same camera angle and are facing the same direction?

[warm junco]

Either, Depth, normal map, canny or mlsd

All with a lower weight value to not force the exact build style

[hollow mulch]

I tried depth, canny, and mlsd, but instead of preserving just the camera angle, it seems like it tries to adapt the shape of the first plane onto the modern one. What I actually want is just to preserve the angle. Lowering the weights seemed to help a bit, but it also caused some weird and unrealistic results. Thanks for your response.

[river bolt]

Shot in the dark, but try pose, specify it's a plane quite insistently in your prompt.

[gloomy ermine]

damn this sage attention is killing my braincells

[dense token]

Triton issues huh?

[gloomy ermine]

yeah

an i literally cannot install sage attention

everything was good until you get to the point where the guide tells you to run a cmd command which very updated

outdated

[dense token]

Don't feel bad, I've wrote and compiled programs before and sage attention / triton hates me too.

[gloomy ermine]

literally spend 5 hours

troubleshooting this

tried update tools

tried venv

in the end i reinstalled comfy

[shadow smelt]

I created a tiktok profile for my gen AI images and videos, is there anywhere I can share it?

[shadow smelt]

welp, here it is if anyone's interested https://www.tiktok.com/@badjano.ai

[plucky wave]

Yooo

[dense token]

*looks at profile* inb4 we get spammed again on all channels.

[fast sage]

Anyone know a good prompt to keep objects in hun stationary?

[urban warren]

I am new to both stable diffusion and discord. I have installed stable diffusion on my own system in Ubuntu 22.04 and it is reasonably working with a low end rtx3060 8gb Nvidia GPU card. My interests are in nature and will be trying to generate images perhaps using img2img.

[fervent thunder]

I think we need to train embeddings for video stuff

prompt engineering is too tricky

training text embedding is slow but certain to work

[open verge]

Hi

[potent oracle]

Does anyone have good promp for a animal pig tail looks Real and accurate for pic serve ai

[wild dock]

They do, lost the card for 10£, and we were over 600£ in price, feeling not happy right now

[steady parrot]

Anyone have advice on what to use for Generative fill? Other than Photoshop 25.0 +

I'm trying to generate really wide backgrounds, i have some images but they arent wide enough and need to fill them to be wider

[atomic mortar]

well it depends on your hardware

what GPU do you have?

like realistically SD1.5 is possible but its not gonna be pretty lets be real

[steady parrot]

i have a 3060.. and it doesnt need to be realistic, im actually trying to generate cartoon style graphics

[atomic mortar]

hmm well i suppose if you use the install method from CS1o in the #🤝｜tech-support pinned messages to install Forge and follow this tutorial youd come a long way https://www.youtube.com/watch?v=5_dOevJRzEI

though for a specific checkpoint (model) youd need to find one that matches your style a little

[gloomy ermine]

finally i managed to get it woking

dammit

guy i have a problom

tea cache in wan produces pure black generations

how to fix this?

sage attention + tea cache did reduce my gen time but they all produce black images

[cold cliff]

hi

[indigo birch]

was anyone able to run rocm and web ui (any) for stable diffusion on rx 6600?

for some reason rocm uses cpu, I looked for a solution on Google but there are only 3-5 pages with the same problem, and the solution did not help.

(Linux)

[steady parrot]

appreciate the tips

[steady parrot]

Do you know anything about making images seemless and loopable horizontally? so both sides can tile seemlessly?

[atomic mortar]

Like x and y tileable? In forge idk but in swarm its a setting

Lemme look it up

[atomic mortar]

https://www.reddit.com/r/sdforall/comments/1bdv79u/creating_seamless_patterns_and_textures_with/

[steady parrot]

For my instance, it would be just the left and right sides, i get a bit confused on which axis it is, because a few of the programs i use have different axis for some odd reason.. im working in Godot Game engine

So im trying to make seemless backgrounds.

[steady parrot]

Yea i've made seemless textures before, but i think it may be different because they are just repeatable patterns, compared to a long horizontal background image

[atomic mortar]

Hmm iirc Y is tall and X is width

[steady parrot]

for example im trying to generate loopable backgrounds like seen in this video
https://x.com/MarsTouchStudio/status/1691855600521543969

[atomic mortar]

But in swarmUI its a simple button press but idk in forge

[steady parrot]

i guess i'll try to figure out both and see if i can pull it off

[wild dock]

I'm part of the 3090 club now, just waiting for it to arrive but yeah same price as in germany.....

Cant wait to do my own lora locally !!!!

[wild dock]

Sorry guys didnt want to brag off i am just happy right now.

So what are the limit with 3090, i will be able train my own lora sdxl

Hello flux

Hello wan

I might stick around with sdxl, i like it

Run 3.5 ok but lora ?

[dense token]

You can train a lora even on an 8GB card, the requirements just vary depending on the architecture used

[wild dock]

3.5 lora ?

god, i didn't managed to train sdxl on my actual 10gb

but I remember there were a trick with training the text too, sorry it was some times ago I can't remember

[fervent thunder]

you could train on CPU so any GPU works

[wild dock]

I guess to put the hassle on the ram in that case, whatever you don't have in the gpu you put it on the ram

and it is slower at the same time

but tbh I am not sure how I was trying to do it and I don't know if you can put part of sd on the ram and other on the gpu

not gonna lie it is a part I am willing to learn 🙂

[fervent thunder]

yeah you can

[soft ivy]

hello

[shadow smelt]

can you guys point me to a workflow for sketch to image using sd35large?

nevermind, I think I got it

[desert dagger]

usually you just want a lora for that

[shadow smelt]

I found a canny controlnet model for sd35 and used it with a preprocessor

works nicely

now I'm trying to find an ipadapter model for sd35

the one I have only works for sdxl and juggernaut for some reason

[gloomy ermine]

dammi finally got i working

msvc was not in the path and reinstalling comfy was needed

aight im ded

[fervent thunder]

been getting nice SD35 results on their huggingface demo (it uses diffusers)
I think there were problems with my comfy setup before

gonna try to either make something from the diffusers pipe or look around github for SD35 pytorch scripts

[gloomy ermine]

are illustrious trained in 0.1 usable in IL 1.0?

[fervent thunder]

not sure but probably

if you are able to use block-by-block lora loader it helps a lot

[atomic mortar]

Most of them work? I had like two broken loras

Out of 100 tested

[fervent thunder]

you can train a lora more

merge it with base model, then fine tune that a bit
then re-extract lora

its not the best idea but it would work

[feral pike]

where to place the hires files?

like 4xNMKD-Siax_200k.pth

i guess its called the upscaler

[atomic mortar]

Yeah

[feral pike]

where to put it

in forge

like where do i put the file

[atomic mortar]

Hmm there should be a upscalers folder in the same place where the loras, checkpoints etc are IIRC

[feral pike]

none

[atomic mortar]

If not try making it and placing it there. I did something similar when i was still on a1111

Wait

[feral pike]

okay so it seems like i had to place it in models/ESRGAN

[atomic mortar]

No mb its the ESRGAN folder

[feral pike]

and then restart the entire thing

[atomic mortar]

SwarmUI has a upscale folder got confused for a sec mb chief

[gentle prairie]

i need some help with civitai helper, this one :

https://github.com/zixaphir/Stable-Diffusion-Webui-Civitai-Helper

everything works fine, except deleting models

when i try to delete i get this error :

"fail to get mode for Lora"

[atomic mortar]

its a unmaintaned project

[gentle prairie]

i found the issue, when you create more folders inside the lora folder, it doesn't manage to reach it to delete it, it can only delete it in the main folder

[feral pike]

so generating with forge ,sometimes completely freezes my pc , and i have to restart
any idea what might be happening?
i have 32GB ram , i5 13500cpu , amd 7800xt gpu
using ZLUDA backend

[warm junco]

Does it only freeze when you use hires fix or other Upscaling?

If so then enable Never OOM for Tiled VAE

[feral pike]

no , even when not upscaling

[warm junco]

you can show your txt2img settings in #🤝｜tech-support

[jaunty hinge]

Hi, can I run the FLUX model and lora training on RTX 3090 and 8B vram?

[abstract quarry]

3090 has 24gb vram...?

[oblique jay]

They do, yes.

Both 3090s and 3090Ti's

[abstract quarry]

I was referring to sunny

[oblique jay]

Oops I have no idea what I read haha

[unique halo]

anyone using FluxGym?

Does anyone know why it always don't detect the images of the directory I provide the path "__" in the reg_data_dir parameter in the advanced options? it always show WARNIGN 0 reg images...

I have spent weeks asking here and there no know answers!! ... even AI searches aren't familiar.

[dense token]

Why do people insist on using FluxGym when OneTrainer exists?

[abstract quarry]

I mean, it's a good thing that so many training tools exist

[unique halo]

I thought it was rubbish.. couldn’t train flux successfully.

[nova lily]

for A1111, Is there anyway to manually save your entire state completely I mean like all parameters from extenstions? For example if you use that roll back button it only restores your last generation, but what if you wanted to save a particular setup. And also it doesn't fully save extension menus such as AD or USDU. thanks

[dense token]

There's a settings file called "ui-config.json" in there somewhere.

[meager jasper]

Can anybody point me in the right direction for getting started with faceswapping?

I also want to know this...

[gloomy ermine]

guys do you know what resolutions works best for wan 2.1 im using 480p model

[quiet finch]

i dont know

[dense token]

Default ComfyUI workflow is 823x480

[desert dagger]

720p

[chrome adder]

How do you train realistic sdxl

Noticed some really good SDXL lately

[atomic mortar]

The fix is laughably easy you dont need the old one. Remind me in 12 hours when im home again

[desert dagger]

use photos for your data set and just train for sdxl

[dense token]

Gonna be honest with you chief, you get good high res datasets from p*** sites, and caption them with an LLM. That way you can use natural language when people use your model.

Because some of the images from those sites come in at insane resolutions from serious camera hardware

[fervent thunder]

I think these days on Huggingface there are enough good datasets

its interesting cos that really wasn't the case in like 2022

the image quality assessment and captioning is there too now

[dense token]

I haven't looked at a good image scorer in a while, any recommendations?

[fervent thunder]

still trying to work it out rly

its definitely still an area that has issues

they often get small trends in the images that get focused on instead of quality

[dense token]

I can link you to a good one if you need me to.

[fervent thunder]

like with "Aesthetic score" there is this brown colour and wavy texture

its ok I need to read from like 2020 onwards anyway

to catch up

[dense token]

Trust me, I noticed how many issues there are with image scorers. I even wrote a guide on Github on how to add more to wd-park's extension "sd-webui-model-mixer". The one I recommend most is https://huggingface.co/sanali209/imclasif-quality-v001

[fervent thunder]

ah thanks I don't know this one

[dense token]

I recommend using "good" mode, instead of "normal" mode.

Good mode is super strict with scores, while normal mode is very relaxed with them.

[fervent thunder]

okay yeah will try this one out

[dense token]

Are you using comfyui, or a1111 / forge / reforge?

[fervent thunder]

comfy and diffusers at the moment

[dense token]

Then my guide would be of no help to you at all. What scoring node are you using?

[fervent thunder]

oh if I do a big project like this its not gonna be in comfy or diffusers
probably just pytorch

[fervent thunder]

I got this as a starting point https://github.com/chaofengc/IQA-PyTorch

its more on the lightweight side though its missing big VLMs and definitely missing any kind of multi-agent one

still not sure about multi-agent stuff

[upper summit]

Hey everyone – we’re currently testing an early version of our spatial photo frame, and we’re looking for people to try it out and share their feedback.

This frame lets you see 3D and holographic effects with the naked eye—no glasses or headsets needed. It’s still in development, but we’re giving out free units to early testers to learn how people actually use it.

All you need to do is fill out a short survey, and I’ll send you a device for free (you need to be in the US or Canada).

Here’s the link to sign up:
https://shop.cubevi.com/pages/free-cube-box-1-giveaway-form

Video showcasing the product:
https://www.youtube.com/shorts/jqPOBhbVL1s

If you’re into cool visual tech or just curious to see what this looks like in real life, we’d love for you to check it out. Feel free to DM me.

[fervent thunder]

is it different to 3D TV?
I'm in UK so I can't join either way but I was just wondering

I think it would be a good product for certain styles of AI image 🤔

cos they are already 2.5D in style

[mighty tulip]

Hi

[sly tartan]

hello

[lunar solar]

Hey I have a question. What style or prompt structure is used to replicate this specific style? And if it's even possible? Any other AI tools that could help make this style? (I uploaded example images in this drive folder) https://drive.google.com/drive/folders/15sQ2nb4Y11NHuy-XTxZ2upX3CKKEL1Na 🙏

[meager jasper]

Reminding you after 12 hours 😉

[placid hatch]

Anyone here running. 9070 xt for stable diffusion? Wondering how the state of things is with the new amd cards.

[still glacier]

if you're running linux it should work pretty much immediately with rocm, if not .... gotta wait for zluda to work on the new models to get descent performances...

[sacred veldt]

I did an oopsie, left Onetrainer to train and thought it would stop by itself like kohyaa but it did not 😬 checked today and had no more disk space, how is it usually handled?

[supple depot]

[dim gust]

Affordable GPU rentals available now: RTX A4000 for only $1.50/hr. Instant setup for Stable Diffusion, AI training, rendering, and more. DM to rent immediately

[blissful hearth]

hello

[placid hatch]

I am running Linux. I noticed it isn’t in the support matrix for rocm yet. Does that matter at all?

[chrome adder]

Yes. But what if we already have a flux lora

[still glacier]

Oh .... I retact my statement then.... I thought it was at least supported officialy by rocm....

[fervent thunder]

yeah no official rocm support its crazy

[placid hatch]

amd is a serious company

[fervent thunder]

its so strange that AMD are like this

[placid hatch]

mixture of incompetence and lack of resources

[fervent thunder]

if you read about the Tinygrad and SemiAnalysis teams interacting with AMD its like AMD doesn't even want to improve

like people offer to help and they are still slightly reluctant

they hold back much better drivers and software stack for internal use and for enterprise deals as well
which is crazy because they should be trying to grow their popularity

[wispy star]

The answer is simple, 1 person = 1 vote = 1 accumulation account, use heartbeats and or DNA to make sure each user has 1 account, Bitcoin but all users have the same mining rate or stake.

But try and create it and even mention the idea to others and you will be forced into prison to eat your own poop while they deny you food like I was.

[fervent thunder]

this exists already it was done by Sam Altman (the OpenAI guy) using iris scans

I don't agree with it but it exists

iris scans are a better way then heartbeart or DNA probably lol

[placid hatch]

i think there is a culture of contempt for software guys in thier company

[placid hatch]

are flux models supported on amd?

[untold crescent]

hey, I heard that you could use stable diffusion locally using your gpu. Does anyone know where I can download it?

[abstract quarry]

you can install https://github.com/invoke-ai/InvokeAI
the tool is quite self-explaining

when you install, it asks you which models you want and downloads them automatically

depending on your gpu you might want to use stronger models like Flux or weaker but faster models like SDXL. Depending on your needs you might want photorealistic models like Realviz or anime models like Illustrious

there are many other tools you can use, such as comfyui/swarmui or forge

[untold crescent]

thanks for that. I want to generate magic the gathering art that is on the cards. I have a 4080 super and would like the highest resolution possible where everything in the image is very clear, but I don't want the art to be photorealistic. What would you recommend now?

https://preview.redd.it/how-does-one-go-about-acquiring-this-dragon-tempest-card-v0-b6klihsk89ed1.png?auto=webp&s=5a3a87d9a57b7d5838c7ada7ea5ebbcf10c61c5e

I would like to create similar art to this, but then create a borderless variant for example

[abstract quarry]

You can try flux, it can do high resolution cards and has a good prompt understanding. The text should nevertheless be added afterwards as complex text is still challenging for many models

there are also loras for magic - however, if you want best quality then you just do the illustration with a diffusion model and copy&paste the card layout yourself in any graphics editor

but in general every model is capable of that. SDXL might be more artistic than Flux (although there are also very painterly flux models such as PixelWave)

[lapis creek]

Hello

[vast topaz]

Hello im new here ^^

where do i prompt here

[faint timber]

Check out the Artisan FAQ channel.

[supple depot]

@still glacier I don't know who the staff are, but I got an unsolicited spam DM, possibly something more malicious if I were to fall for it

I screenshotted

Basically I can show who they are if they need to be kicked

[still glacier]

sure thing, can you dm me it ?

[supple depot]

Is "Community Guides" staff? ^^

[still glacier]

I ve got some privileges. I am not part of stability.ai staff, but I can handle some things and pass it directly to whom it may concern.

[supple depot]

Okie

[jaunty lichen]

hello

[lunar solar]

ih

hi

[quick plank]

hello

[dim gust]

Anyone need to rent gpu’s?

[atomic mortar]

I know some reputable providers such as vast ai or other run prod services if your looking for some

Or are you advertising but then no thanks, shady ah hell

[dense token]

It's his first and only message, just report as spam, block and move on.

[atomic mortar]

Ah saw some other smucks also advertising similar stuff

[cobalt geyser]

Hi~.

[regal radish]

Hi

Can stable-diffusion-xl be used with MagicAnimate?

[high ruin]

Topaz Video AI which 2 model combo give you the highest output results?? Let say you have a 1080p or vertical 1080p flipped and you want to uspcale the video which 2 model combo would you use to get the best results??

[outer goblet]

hi i see a pic now i want to know what pic's style is can any body help?

[lofty flower]

Guys if anybody done movie recommendation system with collaborative filtering approach please tell me

[fading zodiac]

heyo I have a technical question about stable diffusion models. On civitai there are many models with the ending "xl". Does that mean they are build based on the standard sdxl model which is mentioned in this paper? https://arxiv.org/pdf/2307.01952

[atomic mortar]

Xl indeed implies the sdxl model

[fading zodiac]

alright! But how is the exact relation of sdxl to the specific models like "dreamshaper xl". Is dreamshaper a more finetuned trained version of base sdxl. Or is dreamshaper for example just using something like the base training architecture?

just as an example

[atomic mortar]

Yes, further trained/finetuned

Trained = further trained on

Merge = merge of other models

Training from 0 would takes a long long time

[wicked root]

this lumina model is the only model i found that is good for the result i want

[floral umbra]

@woven panther Can you check if it's doable to make tensorRT models for the wan e5m2 fp8 models? As it's possible to make image gen, as well as some video gen models to tensorrt even if it'll use more than 24GB vram, it'll just dump excess of what it can't fit into ram, making it way way slower to convert, but doable.

[grizzled bison]

what flux model will the 9070xt (having 16gb vram) best handle with a lora or two equipped?
fp8? q6? q8?

[high olive]

Hi all, i'm a newbie and i'd like to ask suggestions on how to upscale an already generated image. Can someone suggest me a way or good tool to run locally, please?

[atomic mortar]

Spam

[dense token]

Look at the model size + Lora size, and pick the biggest model + loras that's smaller than your VRAM

You have 16,384 MB of VRAM to work with after all.

[atomic mortar]

Spam/scam

They be active today

[high olive]

What a luck, i write 1 message and got 2 scammer only for me...

[atomic mortar]

I'm on my pc in a bit. I'll tell you how i do it then

[high olive]

Thank you very much!

[atomic mortar]

He's one of them

[high olive]

I have a 3060ti (8GB VRAM) if the info can help.

[atomic mortar]

Sould be fine, i used to do it on a 3070ti 8gb

[grizzled bison]

where is the flux fp8 download, is it unofficial?

trying to find out its GB size and to test it on my 9070

[fervent thunder]

its not official yeah

[grizzled bison]

do you know the one i should grab?

[fervent thunder]

not sure if fp8 is a good idea on a 9070

[grizzled bison]

ah okay too high

[fervent thunder]

its only official if its from Black Forest Labs

[grizzled bison]

its 12gb

[fervent thunder]

if the card does not support FP8 then it will "cast" it to another format such as FP16, BF16 or FP32 when you start making the image

[grizzled bison]

so for my 16gb 9070xt... what would u say i should gen on for quality without overloading into cpu

[atomic mortar]

Sdxl

[grizzled bison]

q8? q6?

[grizzled bison]

no shot 16gb cant handle flux

[atomic mortar]

Its rather you lose some vram due emulating cuda

[grizzled bison]

been reading folks use it with 16 all over on amd

[atomic mortar]

With zluda/directml

[fervent thunder]

oh I thought you meant an old card like GTX 1070

your card can handle FP8

[grizzled bison]

im on a 9070xt

thats what im saying lol

my card can definitely handle flux euro 🙏

[fervent thunder]

could you sell it and get an nvidia card?

[atomic mortar]

Yeah it can handle nut idk how well specifically

[atomic mortar]

What UI do you use rn?

Cs1o definitely does i think

[fervent thunder]

zluda might go away at any moment

[atomic mortar]

That too yeah

[fervent thunder]

its unlawful so its not certain to stay

this is like the fourth zluda
cos the previous three went away

why not sell and get nvidia?

it won't cost you anything

[atomic mortar]

I mean he'd be stranded without a card for a while

If he uses it for work etc

And second hand = less value then a new card

[fervent thunder]

it doesn't really matter when AMD can't run the vast majority of modern software

which all has AI stuff in it now

[warm junco]

you will have much more fun with sdxl than with flux but you can use flux like Q4-Q6 gguf or fp8

[grizzled bison]

https://huggingface.co/amd/FLUX.1-dev-onnx i found this amd uploaded dev onnx should i try this? i cant find the dl link on the page

[warm junco]

No it won't work with zluda

You could use it in AMDs Amuse software or if you setup directml with onnx

[fervent thunder]

yeah onnx is like an alternative to pytorch or tensorflow
which is the layer above stuff like rocm or cuda

its possible that someone could get onnx runtime working on ZLUDA maybe if they compiled it from source
not sure its worth the effort though

what definitely won't work is the tensorRT thing, but there is a separate generic onnx runtime as well.

[austere holly]

How i start generating a image?

[grizzled bison]

thank you cs1

[long talon]

Anyone has any info on paramters count for Reve Image 1.0 ?

[atomic mortar]

depends, do you want to do it locally or in here?

[deep horizon]

is there a guide on how to make videos with stable diffusion? like how to set that up?

[atomic mortar]

not with stable diffision specifically since theres "better' models out lately. hope they release something from SD again. but it really depends on your GPU

whats your gpu?

[deep horizon]

4090 trx founder editon

[atomic mortar]

ah yeah then you can run WAN2 14b model

[deep horizon]

*rtx

[deep horizon]

i dot even know how to set that stuff up. and havnt foudn guides anywehre.

[atomic mortar]

https://www.youtube.com/watch?v=KcYuWRB1_xI
personally i use swarm

installing swarm is pretty easy

[sacred sigil]

can someone please tell me some completely free ai tools to generate good quality of images

[atomic mortar]

depends on your hardware

cloud based? theres a cost or a few images a day/month

locally? just your electricity bill

[sacred sigil]

electricity is not a problem for me. its problem of the hostel

[atomic mortar]

well what specs does your computer have?

[sacred sigil]

what i need

[atomic mortar]

well

gpu?

[sacred sigil]

just tell me i will buy

[atomic mortar]

well thats not what im asking but if you mean the BEST currently? its gonna cost you over 40k lmao

[sacred sigil]

rtx 3050 with ryzen 5 enough

[atomic mortar]

yes the 3050 can work. though the smaller models

sdxl probably will manage

[sacred sigil]

so you mean i need to use open software well than how can i setup that

[atomic mortar]

i recommend following CS1o's guide:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#nvidia-forge-webui

if you get stuck i recommend hopping over to #🤝｜tech-support

[sacred sigil]

thanks mate

[dawn mulch]

Hey there was a node that loaded multiple images and basically merged it as one image which could be used for ip adapter face id to generste sn image of the merged faces. its different from just batching multiple faces and then generating multiple images based on each. Does anyone know a costume node that does that?

[humble mason]

I just acquired an RTX 4090. What would be the best version of SD for me to run?

[dense token]

3.5 if SD, but you could also generate videos with Wan2.1 or use Flux

[humble mason]

Would you say that Flux is better? I've never tried Flux. I had a tiny 8 GB card and I was running SD 1.0 for a year or so

[dense token]

I haven't tried Flux since they released the first version. I never could get it to run decently, but that was an ancient machine.

[dense token]

I have a tiny 8GB card and I can run SDXL based models and FLUX no issue.

[humble mason]

Wan does sound interesting... although a lot of its results are... preposterous 😅

[earnest jackal]

Meow

[nova siren]

Hello!

[sharp wraith]

how do i cancel the subscription for stable diffusion?

[safe igloo]

Hi everyone, I’m running SD on my 10yr old CPU which is basically a toaster. [proud face]

[atomic mortar]

their helpdesk/support desk email probably

[atomic mortar]

found it on their site:
https://kb.stability.ai/knowledge-base/how-can-i-cancel-my-stable-assistant/artisan-subscription

[sharp wraith]

i found out that i subscribed to a fake version of it and currently trying to email with them thanks tho

[atomic mortar]

if you got a fake version, try doing a block and chargeback thru creditcard (or if you mean its a alternative version with subscribtion then your out of luck)

[sharp wraith]

yeah i’m trying to get a refund threw the bank to get my money back

[sharp wraith]

i’ve also cancelled the subscription so that’s good

[foggy hamlet]

Hi

[ripe atlas]

Hey frens what's the best most advanced cheap service for generating images? (I have only access to my phone 😭)

[ripe atlas]

Looks like the Stability assistant is something I could use

[atomic mortar]

Youd be better off on Fiverr then a community server

And again really?

[honest prawn]

hey there, anyone know why animatediff isn't working for me?

[atomic mortar]

Without the logs its impossible to know

I recommend poppin into #🤝｜tech-support

[pine path]

WAN paper is out(I have no idea where else to post this as research channel was deleted)
https://arxiv.org/abs/2503.20314

[pulsar venture]

Hi, i need help. Forge is the same thing of re:forge?

[honest prawn]

where can i get those? i have everthing installed and all the right files it's just drawing an image like usual and not an animation

[quartz thorn]

Guys can someone help me generate images?

[jaunty hinge]

which help do you need?

[quartz thorn]

Can i dm u?

[jaunty hinge]

sure

[olive fossil]

want to confirm my reading of the api docs: only Stable Diffusion 3 & 3.5 api calls support image-to-image, not stable image core?

[fervent thunder]

are you sure its faster than the 1-step 512x512 version of SDXL turbo

it is actually possible to significantly prune SDXL without it losing details, this is essentially what Segmind Vega is https://huggingface.co/segmind/Segmind-Vega

[cold parcel]

ill remember the day “thegoat9296” brought upon the great ai renaissance. Truly inspiring man

[fervent thunder]

there's almost no demand for fast models BTW
I like this model a lot its one of the smallest SD 1.5 prunes https://huggingface.co/nota-ai/bk-sdm-tiny-2m
but look at the downloads its less than 300 users per month

[mighty tree]

hello

[fervent thunder]

is Euler a same as Euler_ancestral?

[vale ibex]

Really? No demand huh.

A model that could just generate literally anything from a picture to a video game on cell phone has no demand thats crazy

[velvet anchor]

I feel like I might be starting to go crazy... Its been awhile since I reorganized my LORAs but they really need it. I used to be able to delete them and rename them directly from the loras tab in A1111. When you hover over them there were buttons that popped up for that. But they arent there anymore? Looked at a fresh install and it still wasnt there? Is that from a extension and if so does anyone know what extension that is? I tried looking for an extension for that but I didnt find one.

[humble mason]

Hey, there was an open source project a while ago that used (SD?) to take badly blurred photos and restore them. Anyone know what that was called?

[dense token]

Thank you for reminding me this exists, as I've been looking for them again.

[atomic mortar]

Bro again, 3rd time

Yeet it in #🌶｜off-topic since were all users of it here already

[stark goblet]

What's the issue? Please share a screenshot of any error that pops

[fervent thunder]

yeah its crazy
I am very much in the camp that fast models are better

I run the 0.33B BK-SDM-Tiny-2M model even on H100s lol

unlike some other similar "fast SD" things, BK-SDM project also addressed the VAE

Sana team understood this as well that to go fast VAE will be binding if its original SD or SDXL VAE

[zinc adder]

hlo

[manic wraith]

Saw this syntax on a SDXL image

<segment:girl's face,0.2,0.3//cid=1>

I assume it's either A1111/(re)Forge or Comfy syntax or maybe it's some extension, anybody knows?

[cerulean ravine]

anyone who's good at stabble diffusion and could answer me a question, it takes like 3 seconds, i just need help with something... thank you

[atomic mortar]

Hi, why don't you post your question so anyone who sees it could give you a reply?

[cerulean ravine]

it just gets passed on

[atomic mortar]

I mean, this also gets passed on normally lmao

[cerulean ravine]

I just need some help from someone who knows how stabble diffusion really works to tell me if some images were generated using stabble diffusion or not

[atomic mortar]

Like if you posed it, i couldave answered it already

[cerulean ravine]

you're right

[atomic mortar]

Oh bruh just use one of the online checkers

[cerulean ravine]

haven't heard of sucha thing, let me know

i mean i know they are generated

[atomic mortar]

One i found on google just now
https://sightengine.com/ but theres more

[cerulean ravine]

but i don't know what AI was used

[atomic mortar]

Post image in #🏞｜general-with-images

[cerulean ravine]

i want toknow what AI wasused so i can buy/learn it so i can generate them myself

[atomic mortar]

Some have a distinct look

[atomic mortar]

No 4, @vapid dove can we get rid of advertising? Four posts like this in 2 days

[dawn mulch]

Anyone here got their honeymoon 3D two (Hunyuan3D 2) working in comfyui?

[restive flicker]

Is there any way to implement {|} dynamic prompts in ComfyUI

[finite cloak]

that's SwarmUI syntax https://github.com/mcmonkeyprojects/SwarmUI/blob/master/docs/Features/Prompt Syntax.md#automatic-segmentation-and-refining

[finite cloak]

generally the best way there is just ask the person who made it how they did

[autumn gyro]

hi all ❤️

unsure if thats the right place to ask but im a total beginner playing around with sd (auto1111)

how difficult is it to create an image of my pet wearing some sort of a formal suit

[tawny drum]

👋

[humble mason]

It's easy to create a pet wearing a suit. If it specifically needs to be your pet, you'll have to make LoRa to train the AI what your pet looks like, and that takes some doing.

[tawny drum]

So Stable Diffusion 3 is the newest version?

[manic wraith]

Thanks!! Tried searching and couldn't find anything at all.

[manic wraith]

Technically yes, but on Civitai you can see most people still use older versions of SD (fine-tunes that give it better quality), or Flux (usually only good for realism, and pretty good for that)

[covert trellis]

hello, I'm new here.
I just download and installed Easy Diffusion (Successfully.) but every time I boot it up;
the status message says the server is down.
is the server down temporarily or something?

[tawny drum]

Do you know what graphics cards it will work on?

I have a RTX2080

I know I can't use SDXL

[manic wraith]

Yes you can, without issues at all really. I have a friend who's running it on a 1060 6GB! Currently I have a 3090 and I run batches of 6 images usually, so batches of one would be totally doable on a 2080. I recommend you grab reforge webui from panchovix (it's like automatic1111 with a few extra features) and whatever SDXL model floats your boat, like Illustrious fine tunes are very good for booru-style prompting.

[tawny drum]

Would I have to uninstal SD1.5?

to then install sd3?

[manic wraith]

You don't really install SD 1.5 or SD 3, you usually install something that uses a checkpoint for you (the AI that makes the images is just a file that a program runs, you can have many such files)

So you can keep a whole list of models you like and use whichever you want at that time

[tawny drum]

ty for the info 🙂

[manic wraith]

Usually people choose either the ComfyUI route (kinda nerdy but powerful) or more visual web UIs like reForge, Swarm, etc

[tawny drum]

Idk anything about anything other than what Im using 😂

tbh alot of the lingo about SD and stuff really confuses me 😂
I do my best

[manic wraith]

It confused me a lot at first too! Soon you'll get it, don't worry. Also there's some stuff that's really advanced and you're usually safe leaving stuff at defaults

I don't know which program you are using, but for Automatic, forge or reForge, you can grab someone's PNG/JPG from Civitai and copy all the generation info from the "PNG info" tab, this should help you reproduce some result and use it as a reference

[tawny drum]

Pretty sure im running automattic-

I do take a lot of ideas from peoples codes on civit mhm

🤭

[manic wraith]

In that case, just try SDXL! I think reForge has something to manage VRAM better (I never seem to hit an out of memory error on it) so it might be worth taking a look. You can install both to see if you like it; another upside is some illustrious/noobAI models only run on that

(I used automatic1111 before too)

[tawny drum]

Ive tried using it, my PC wont run it D; gives me the memory error

SDXL does