💬｜general-chat

[copper crystal]

no wait, that was neon sorry. you were just there

NOIM WRONG AGAIN itwasYOU!

[trail lion]

I've seen debates on the 'similar images' point on reddit. Some comments have indicated that the 'pro' version is not subject to this issue. I havent looked into this personally, or experienced the issue to the ponit where it has annoyed me. The unaesthetic point I just dont see personally. People on reddit complain about a 'plasticy' look...is that what you mean? Because from any other point of view I see improvement, it does hands better, it does skin details better, I struggle to find something that it can't do better given the right prompt

[quartz siren]

yeah lol, i couldn't test it since I don't have enough vram to merge the lora with fp16 flux. did you try it?

[copper crystal]

oh god. more bad/good news.

i was wrong about that that's the turbo link. so i dug up the one i did mean. IT WAS YOU TOO!! https://huggingface.co/nyanko7/flux-dev-de-distill

#🆕｜sd3 message

i'm a mess. i need to eat or something

[trail lion]

whenever I comment on flux, I only use dev, seldom or never schnell, and so far no finetunes have been declared worthy to keep

[copper crystal]

i dont think i've ever used schnell. if i ever need a totally libre model to experiment for a software solution, i'll dig into it

[trail lion]

XL went through this same phase, until juggernaut, which was nothing short of a miracle. and since then even more have come out. I think flux has that in its future likely

[copper crystal]

they're hard at work on juggerflux surely

[trail lion]

speaking of which, natvision, if you havent looked at it, is a recent XL model, I'm a big fan

it somehow brings almost flux-level prompt adherence to XL

[copper crystal]

went and looked at it on civit. thats a lot of butthole.

[trail lion]

hah, it does nsfw yah

so take that as good or bad

[copper crystal]

with those models i've found that prompt adherance means you must describe porn in the negative, all the porn, so it doesn't go there.

[trail lion]

whoever trained it, gave it a pretty massive dataset, I've done some pretty crazy things to change viewing angles or whatever

[copper crystal]

typically i look at the user gallery, and if that's all people use it for, then that's all it's good for.

I'll stick to juggernaut.

[trail lion]

cant go wrong there

[copper crystal]

lol i dont think i've found one sfw image in natvis gallery

https://civitai.com/images/25279078 wait.. here's one

[trail lion]

give me a prompt, I'll try one

[copper crystal]

that's a good sword pose. long handle but i'll take it

oh geeze. prompt.. on the spot.. rightnow!? Oh geeze... umm a tree on a hill on a sunny day

[trail lion]

sec

[trail lion]

#🏞｜general-with-images message

that was a bit easy btw

[copper crystal]

ew. i sorted the natvis gallery by most reacted. it's even worse. :S

loras for "18 year old woman" all over

[trail lion]

you do you I guess, I tend not to judge by community gallery

[copper crystal]

oh i'm super judgy of a model's capabilities by looking at it's primary use

[trail lion]

come up with something you havent been able to do

[copper crystal]

breakdancing . always break dancing!

[trail lion]

ok, I seriously curious here

[copper crystal]

that's a common case that's not capable

[trail lion]

I'm on it

[copper crystal]

"tree on a hill" is basically what i always draw when i try ot think of something to draw. choice paralysis i guess

[quartz siren]

Yeah I saw the model on hf, but forgot about it bc of the flux hype. It does look nice for a sdxl fine tune.

[trail lion]

dude ,it killed it, standby

#🏞｜general-with-images message lol

[copper crystal]

saturday night kung fu

hot

not quite break dancing but i love it yet.

[trail lion]

I mean, I can throw more terms at it to see

but the point here is, you ask for it, and it doesnt give you a person posing for a portrait

[copper crystal]

the guy in the background has a bit of a bboy stance. it tracks. the model learned something of it def

[trail lion]

I've only had this model for like a week, but I've been trying to push the boundaries

[copper crystal]

have you tried one button prompt? i'm super keen on this script, especially now that i got it working in forge again and have it hooked to flux

[bleak drift]

But can it do photorealism?

[quartz siren]

Both can do it but sd3 will have really bad anatomy and some artifacts. Flux is biased to that mid journey(plastic) look but it’s not too hard to make it more realistic.

[bleak drift]

How would one be able to make it more realistic?

[quartz siren]

Lots of Lora’s for flux you can merge.

[quartz siren]

It’s also really really realistic if you don’t try to talk too much about a person, for example: Steve job, powerpoint presentation, the slide title says “Flux AI has new skills”, three bullet points, “good at text”, “prompt comprehension”, “amazing images”

It will be really realistic then.

[undone garden]

Anyone knows how to load flux models on reforge? Can't seem to find an oprtion for it. Specifically gguf ones

[bleak drift]

Nice. So I could use it to hypothetically take a photo of Shadowheart or Astarion from Baldur’s Gate 3 for a image to image and make them look more realistic and less video game like, correct?

[trail lion]

sure, but you could already do that with XL

[carmine oracle]

Hello everyone.

[bleak drift]

Really?

Just regular sdxl?

[trail lion]

why not? use controlnet with a depth model maybe, and some tags for photorealism

a finetune of XL might be even better, one geared toward realism

[brittle rose]

Do you know why sd3m don't draw country flags correctly? Is there a good variant

[trail lion]

nothing is stopping you from training a lora if you need something specfic. think of these base models as containing the ingredients to bake some kind of cuisine, but in many cases you need to teach it the right recipe to make something specific. reproducing art styles , the likeness of a person, specific building architectures...these are all things that may not be inherent in the base models

[ashen sleet]

sup

[hallow nacelle]

Hi

Anyone use Adetailer?

I have an image of a character in a crowd and I'm trying to make it process my image only the subject instead of all the faces in a crowd

[hasty obsidian]

hey guys, anyone available i wonder?

[quartz siren]

Hmm, what do you mean by correctly? sdxl can do it for some flags, I doubt you require sd3 for that.

[brittle rose]

no even sdxl isn't capable well I wanted to use sd3m for more realism and efficiency and less artifacts

[quartz siren]

Alright what flag is it not doing correctly? And sdxl lightning models are far faster btw, they take like 2 sec to generate on a decent consumer gpu(without any torch compile and tensorrt)

[hasty obsidian]

so far i encounted a problem inferencing

not so sure how to optimize it

my system consists of four NVIDIA H100, each comes with 90GB of VRAM, however disregard my effort, the inferencing process always seem to be heavily utilizing the card labeled: 0

taking up 80GB of VRAM usage

so im just wanna ask, is there a way to full utilize all 360GB of VRAM available? (my personal machine has xeon and 500GB RAM installed as well so there shouldn't be any bottlenecks restricting the performance

and H100 although not the flagship card used in many clusters, i dont think it should be this slow as in my workspace

can anyone help?

CPU: Intel Xeon Platinum 8468 (96) @ 3.800GHz
GPU: NVIDIA H100 SXM5 94GB
GPU: NVIDIA H100 SXM5 94GB
GPU: NVIDIA H100 SXM5 94GB
GPU: NVIDIA H100 SXM5 94GB
Memory: 98624MiB / 515116MiB

The basic spec of mine

or should i resort to the supercomputers available to me on the campus instead? or there are ways for optimization?

[hasty obsidian]

however its gonna be quite a hassle ngl

OutOfMemoryError: CUDA out of memory. Tried to allocate 71.03 GiB. GPU 0 has a total capacty of 93.00 GiB of which 53.07 GiB is free. Including non-PyTorch memory, this process has 39.93 GiB memory in use. Of the allocated memory 24.79 GiB is allocated by PyTorch, and 14.33 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[gray vapor]

Is there any service that allows me to run my comfyui workflows in the cloud (But! Using my local installation for the UI, extensions, etc). I imagine this would be kind of like dynamically creating a docker image, uploading it and running it.

I guess the latency of uploading the models would be bad, but maybe there is such a service that keeps the model on the cloud instance for as long as you dont stop it or smthing

[fervent thunder]

how low latency do you need?

if you need latency of a few seconds then a serverless platform like Fal.ai

[cobalt turret]

how can I say I want a full long shot in flux, always the legs are cut in flux

[fervent thunder]

if you are happy to wait a few minutes for startup then its much cheaper to put all your models in a Backblaze account
and then download fresh each time to the server

[fervent thunder]

I would try control net personally

or generating the characters separately and then pasting them on the background

then refine one more time

[cobalt turret]

ok thx

[brittle rose]

country flags also I wanted to draw some national common symbols of dressing

[fervent thunder]

have you tried IP adapter and control net combo?

its the first thing to try if you can't get an image to work

[brittle rose]

controled net helps somehow but not always compared to dalle3. so i was wondering if there is an sd3 variant that is good in drawing commons

[fervent thunder]

what you can do is go on Civit.ai and search their images

and then sometimes the image will tell you what model they used

[brittle rose]

thanks i didn't know this website exists with this function. I will try

[quartz siren]

If you have enough vram, flux is definitely the best option. Can you tell me an example prompt that sd3 can’t get right.

[floral umbra]

What's the goto img 2 3d nodes/models for comfy these days?

[fervent thunder]

there isn't one really

[cobalt turret]

what is best for upscaling flux photo generations

[fervent thunder]

probably still supir

[cobalt turret]

ok thanks

[brittle rose]

Even Flux turned to be uncommon to commons.

Please check with your setup this prompt:
In the dynamic Anime style with 2D and CGI techniques: A young man with a medium beard, wearing the Palestine Kuffiyeh, pushes against a large metallic wall. His hands glow with raw energy, symbolizing his power to break barriers and reclaim the dignity of his people.

[storm granite]

how to make a prompt to generate an image?

[desert dagger]

it sounds like you just need to train a lora on flags of the world for SDXL

[desert dagger]

read the information here #artisan-faq

[quartz siren]

Is there supposed to be a flag in that image, if so prompt for a flag. What is it not doing correctly?

[fervent thunder]

so I've found that I missed the obvious. Why were all inpaint and erase models absent from IOPaint? Because when you lauch it from the command prompt, you have --model=name to add the inpaint and erase models to use in the WebUI.

[deep horizon]

Hey i was just wondering if anyone knows. I reinstalled stable diffusion recently. before it had a thing wehre when i start typing a promt it game me recomendations to finsish parts. like it would suggest tshirt when id type teeshirt. does nayone know what that is ?

[brittle rose]

Usually with only "palestine" dalle3 draws a flag as well but you can add it

[quartz siren]

Dalle3 has an llm to enhance your prompt, you can usually do whatever dalle3 does but better with flux and llm to help enhance your prompt.

[brittle rose]

Okay I will give yo ua more precise prompt.

In the dynamic Anime style with 2D and CGI techniques: A young man with a medium beard stands on a vast desert plateau, holding a glowing Palestinian flag. The color palette is soft pastel—rose gold skies, white sand, and the flag’s vibrant green stripe glowing amidst the subtle scenery, symbolizing peaceful resistance.

[brittle rose]

just tested flux It turned to be good for this take but It does not draw Palestinian Kuffiyeh for some reason

[desert dagger]

you have to realize that while a model like flux does have billions of parameters of data in its data training set, it doesn't have every single bit of human knowledge. and it's probably trained on what is fairly common and easy to get hold of. for a lot of things that are not overly common to where the programmers would have gotten the information for the data set, you'll need to train a lora with that specific information and use it with the model when you generate

[vagrant raptor]

Do you have a good walkthrough for training a LoRA for Flux?

[desert dagger]

start with this https://x.com/LikeToasters/status/1836632745075736913

[vagrant raptor]

Thanks!

[unborn hedge]

i wonder if its possible to run FLUX locally yet, i got a 4080 with 16GB of ram, i heard there was lots of advances in cutting the vram requirements

[low moon]

You can run some flux versions on a potato.

[unborn hedge]

really?

im just trying to run flux locally so i dont have to sell my liver and lungs on civitai to use it

[marsh crag]

"I'm looking for some ComfyUI flows that can quickly and accurately paste product mockups onto white templates (well, 80% accuracy is fine, I can do post-processing later). If anyone has this information, could you share it with me?"
some kind like pasting exactly the brand of left photo into the bare bottle on the right

[hallow nacelle]

You can try, I have 4080S and running flux dev has a little struggle with not enough VRAM

It will not spill out cuda error though, and utilizing more system memory

No issues with flux schnell and nf4

[warm junco]

Hey, the extension is called boorutag autocompletion

[pearl berry]

hi!

[ashen sleet]

gm

[unborn hedge]

i see, thanks! i only use civitai's standard version of flux, im not sure what the difference is between standard and pro.

[brittle rose]

I try to be as native as I can but It seems I need to learn a tutorial about lora

[fervent thunder]

for 1 character its easy, for more than 1 or other stuff like nsfw you need experimentation and choosing right tool

[remote vector]

help)

[desert dagger]

i understand, i do the same thing. but that's why loras were invented in the first place - so that models can have their information updated as needed for specific things without having to retrain the entire model

[hallow nacelle]

Yeah pro version only available as api

As they say https://blackforestlabs.ai/

Pro version is ther best version

[unborn hedge]

ahhh alright

[fallen basalt]

Hello to all 🙂

[swift juniper]

hello frens. does anyone know if theres a way to upscale only part of an image?

[bleak matrix]

Hello, hello!

How is everyone?

[gray vapor]

What is your fav platform for using comfyui in the cloud? That allows for storing extensions, models etc. (Or at least allows to get them from url and keep them loaded for the session). Better if it is paid by the hour

Right now I rent a full desktop computer in paperspace but I think it might be an overkill

[gray vapor]

You could just crop it and upscale it

[lime marsh]

has anyone installed RVC on google colab pro lately ?> Seems impossible, just never ending depedency hell.

[fervent thunder]

its very situational

different platforms are better for different individual GPU models

and some are better in terms of download speed and disk space etc

another thing is the ratio between the costs of disk space, download amount and server rental time can vary

[gray vapor]

I was trying rundiffusion, but they dont seem to offer the possibility of installing custom extensions

[fervent thunder]

you can install comfy nodes on rundiffusion

[gray vapor]

I guess i did something wrong then

But I have to install them again for each session, right?

[fervent thunder]

the way rundiffusion works is that you have permanent storage but it only lasts for 3 days

so you have to install them twice per week

[gray vapor]

that is cool actually

[fervent thunder]

if you subscribe to their creator's club then you get true permanent storage

and a discount on servers

[gray vapor]

The reason after those questions is that i am going to give a course on ComfyUI to architecture students, and obviously most of them wont have powerfull desktop pcs

And I don't think I should make them mess around with docker and stuff

Thats why I need a convenient, paid by hour or credits way to run it

[fervent thunder]

main choice is serverless or not serverless then

ComfyICU is an example of serverless

[gray vapor]

The best for me i think would be to have comfyui running locally, but run the workflow in the cloud via api

[fervent thunder]

well you can get comfy nodes that

do the ksampling via API, for example

[gray vapor]

but you need the vram locally for loading the models right?

[fervent thunder]

no the models are loaded on the API host's server

[gray vapor]

Oh that is nice then, how are those nodes called?

[fervent thunder]

there's a set by black forest labs

https://github.com/black-forest-labs/bfl-comfy-nodes

there are ones by the big API providers- Fal.ai and Replicate
but I don't want to recommend a node pack for them as I haven't used them

[copper crystal]

run diffusion have always done the SaaS thing right. minimal shady practices

[fervent thunder]

they make Juggernaut too, which is nice

[gray vapor]

Yeah, for now rundiffusion looks like the best one

[fervent thunder]

Fal contributes by far the most out of these
cos they made both AuraSR and Auraflow

[copper crystal]

true fal puts money back in too. i was counting that effort by rundiffusion.

[gray vapor]

But whenever I install a custom node in rundiffusion and I stop the session and open a new one, the node is not there anymore

[copper crystal]

community orientated right? like, i support businesses that sponsor the local youth ball teams and stuff like that. they're putting money back into the local space. thats what a healthy economy needs

[swift juniper]

is that the only way? i thought about that but i thought it be annoying to put it back perfectly to the image

[fervent thunder]

yeah rundiffusion did good by making Jugger
I just meant the scale is different
Jugger cost around $5,000 and Auraflow cost over $250,000

[copper crystal]

it is a good point

[fervent thunder]

you could inpaint

[swift juniper]

fair enough

[fervent thunder]

although most modern inpainting workflows do use the crop and stitch method

you don't have to, you can inpaint in place

[swift juniper]

youre talking about comfy ui right

im still an automatic1111 scrub

[fervent thunder]

I don't know A1111 stuff sorry

it doesn't look good by the way, to upscale only part of an image

I see people do it a lot, especially with inpainting

but you can see that different bits of the image have different detail levels, its kinda weird

[swift juniper]

i see

[fervent thunder]

it might work ok if parts of the image are dark/shadowy or blurred

and then you could save time by not upscaling those

but even then, there are actually benefits to upscaling a blurred area, it still looks better

[quartz siren]

You are probably better asking in a rvc server, this is for stable diffusion stuff.

[fervent thunder]

https://github.com/WadRex/RVCompact

[tacit lantern]

Is there a good program for locally managing AI generated images, prompts, metadata, tagging etc? Breadbox is clunky, and McAfee thinks the latest version of Diffusion Toolkit is virus infested.

[fervent thunder]

SQL database?

[tacit lantern]

I don't have the requisite skills.

[fervent thunder]

I think there is software out there that does what you want

I don't know though as I don't really like to use GUIs

maybe someone will know

[vagrant raptor]

Thank you very much for that. It pointed me to this: https://github.com/cocktailpeanut/fluxgym And once I changed .safetensor to .sft, it worked like a charm and I'm using my new LoRA already. I really appreciate the help!

I learned that loading 2 LoRAs for a Flux model is not something I recommend, tho. 32gb memory and 16gb vram and it's all maxed out.

[desert dagger]

yeah - you want to make sure it's using your Vram too, especialy if you have an AMD CPU, which likes to take over and try to be the GPU

[vagrant raptor]

I got Nvidia this time, just for this. It's a 4070 Ti Super, I'm watching the Task Manager chugging away at it.

It's going to be a 10 minute 1024x1024 image. Yikes.

[fervent thunder]

sounds like your VRAM is full so its using your DRAM

you fit Flux Dev onto very small GPUs if you keep the text encoders on DRAM, and then use small quants for the Flux model e.g. Q4 or NF4

[vagrant raptor]

Do you have ideas on improving performance?

[fervent thunder]

yeah what I said in the previous comment will fix it

[vagrant raptor]

Well, I don't know how to implement that. How does one keep text encoders on DRAM though I could look for NF4 at least.

[fervent thunder]

https://github.com/SeanScripts/ComfyUI-Unload-Modelthere are some nodes like this that can force the text encoders to go to DRAM once their prompt is finished

[vagrant raptor]

Ah, clever

[fervent thunder]

I don't keep up with the state of comfy memory management cos I force GPU only
but its possible default comfy memory management can handle your situation anyway and unload the encoders for you by default

[unborn hedge]

remember when we thought FLUX was impossible to have Loras for iirc? or impossible to train? AI development goes fast these days

[quartz siren]

That was more because of misinformation, people said the license did not permit Lora’s.

[desert dagger]

what'll you'll find is that almost all Loras for flux do not do anything different than a correctly constructed prompt does

[copper crystal]

lol wrong

[quartz siren]

There are many styles that flux has no idea about, and loras cant do much except add styles, or accelerate inference or add a bit of knowledge.

[desert dagger]

think what you like

[desert dagger]

sure. but most of what you'll find out there do nothing, dont' even do that. they're just there because the community demanded loras and someone made them some

[proud comet]

hey guys i just saw a picture and i want to know if it is ai or not can i send it here

[desert dagger]

use #🏞｜general-with-images

[sour imp]

hello

[craggy orchid]

hello

[hardy sedge]

Dumb question. When I train a loras, are the images stored in the lora? Or are the images only used for training? Thanks.

[fringe shoal]

How do you guys upscale in img2img using Forge UI? Lets say you have used hires fix in txt2img and sent it to img2img. How do you upscale? In stable diffusion automatic 1111, I used latent upscaler in img2img which I cant find anymore in Forge UI

[trail lion]

they are not stored in the lora model in a way that can be extracted later, of course your model will start to reproduce images closer to what your training dataset had in it

[hallow portal]

Hi

[gray vapor]

What would be the best workflow to place a design into a billboard?

Ive seen CatVTON for clothing, is there something similar and more general

The key is that i need the original design/img to be respected 100%

Or is the best approach just to generate a "mockup" and replace the design with photoshop

[fervent thunder]

you don't actually want the original design to be respected fully cos then there would be no change

its more about having smallish changes that look pleasing

there's a bunch of VTON models but you could also get there with regular lora, IP adapter and control net probably

[ivory sierra]

hello

[fervent thunder]

i added my cartoon character as a human version to my dataset for lora training. does this a ctually help distinguish my character token from a human or will this just mix both?

[rare hazel]

Hey, I have a question. I have 2 images with the same character that I want to extend. In the first picture I want to add simple pajama bottoms with the same style as her top (the picture doesn't have bottom, is cut below upper thighs area). In the second picture there is only hoodie and i want to again extend it to full body. However as I try to inpant the image the result is so bad that I can't accept it. How can i be generating better results? I'm using stable-diffusion-webui-forge. The character is drawn in anime style and I'm using anime styled model.

[copper crystal]

consider using fooocus or ruinedfooocus UI. ruined is a fork and has flux support.

they're great UI's that simplify things and for sdxl models it has a custom inpainting system that works veyr well.

[rare hazel]

okay thanks i will try that, do you have other useful tips?

[copper crystal]

drink water. wear sunscreen.

[dawn mulch]

is there a way to figure out why and how SD determines if an image or parts of an image should be blurry or sharp and then hijack it and allow the user to specify the bliurry and sharpness? I tihnk IPAdapter uses this idea to do what it does.

[copper crystal]

it's all prompt vectors. ip adapter turns an image into a prompt style embedding. those give the system a vector to aim for.

you can use a lora too, which adapts the network to respond to vectors in new ays

[dawn mulch]

are 'prompt style embedding' just text?

[quartz siren]

ip adapter embeddings are different from text embeddings. They are both embeddings that guide the unet/dit model but they are different.

[quartz siren]

prompt is different from text, an image can be a prompt for example.

[wicked monolith]

Anyone know how to add openpose models to controlnet when using the forge ui?

[copper crystal]

tried putting them in your Models\ControlNet folder?

[wicked monolith]

not working

[copper crystal]

you've changed soemthing about your ui then. that's the default location where it looks for them

make sure you're trying to use an sdxl openpose model on an sdxl model. instead of a 1.5 model

up to date forge doesn't check for versions though

[wicked monolith]

it worked and then they disappeared from my model folder? this is wild

[junior notch]

hello!

[fervent thunder]

anyone familiar with comfy ?

i need someone asap on my team (paid)

[hollow gorge]

If anyone is interested in trying a new bot with inpaint and other features theres an invite link in my bio. The bot is not live until the first, but It's running rn if anyone wants to try it

[waxen lark]

How can i made ai pictures ?😅

[tired tree]

hi

[pulsar venture]

Someone does know the difference between the nodes Freeu and Freeu_v2 for comfyui?

[opaque oasis]

hello

[wicked monolith]

can someone tell me what "Value error: failed to recognize model type!" means?

[subtle edge]

hey, will stability ai launch a native coin?

[dreamy turtle]

anyone know how to make consistent generations?

[desert dagger]

Super specific promots. Or train a lora

[jolly field]

Any other NSFW artists here? i would love to discuss some concerns

[hard swift]

yo gm

[fading delta]

hey ppl

[mystic galleon]

where is svbot python dev

[wise dawn]

Will we ever get Midjourney serf like thing for Comfy ui?

[fervent thunder]

yeah its called IP adapter

[winter ridge]

#artisan-1

[unborn hedge]

I'm thinking of making a comic with ai art and I might struggle with that too, and idk how much a LoRA could help

and even use a ai voice thing to voice some characters

I was thinking of learning to draw then making a LoRA of my characters out of my own art to make it easier on myself

[quartz siren]

Lots of ip adapters are here for flux, I think the best is pulid. You can use that to generate a consistent character.

[fervent thunder]

https://stable-diffusion-art.com/consistent-style/

it injects into self attention and groupnorm

[copper crystal]

tehre's only one ipadapter for flux from xlabs. pulid isn't an ip adapter, it's an identity clone system.

my real world suggestion for consistent generations is to use the same seed, settings, and prompt. tha'tll get you the same or sameish image everytime

[fervent thunder]

the method I just posted should work 100% with flux 🤔
no one has ported it yet but it could be good

[copper crystal]

i don't think comfyui ever ported controlnet reference to comfyui even

[fervent thunder]

ye quite a lot of cool stuff isn't ported

Diffuse High allows SDXL to generate at 4k resolution, no upscale

but its only in Diffusers

[copper crystal]

this one? https://github.com/yhyun225/DiffuseHigh

looks just like the kohya hires fix that's been out for a long time.

[fervent thunder]

kohya hires fix involves upscale though

[copper crystal]

not really. not like hires fix would. it's generating at the target resolution the entire time, but it samples things differently in earlier stages as if it's making a lower resoluiton image. i dont think there's anything novel here. diffusehigh seems to be interpolating lower resolution structure all the same.

kohya's deep shrink was never ported to diffusers though. so they got the method now. that's good

from the diffusehigh paper "In this work, we probe the generative ability of diffusion models at higher resolution beyond their original capability and propose a novel progressive approach that fully utilizes generated lowresolution images to guide the generation of higher-resolution images"

So their approach requires low resolution source or it will generate one to use. Same idea as deep shrink

[fervent thunder]

just to clarify, kohya deep shrink and kohya hires fix are not the same thing

kohya hires fix is an upscale and then a second pass with another sampler

[copper crystal]

deep shrink is synonymous with hires fix. the comfyui node is called deep shrink. the extension on webui is called hires fix. it's both the same code.

if i'm mistaken i'd love a source, since that does not clarify things

[fervent thunder]

I thought hires fix on A1111 was either a latent upscale or a pixel upscale, and then a second sampler pass
and then kohya deep shrink is a bicubic downscale at block 3, for about 30-35% of the steps

[copper crystal]

https://gist.github.com/kohya-ss/3f774da220df102548093a7abc8538ed this is the only code kohya ever published for it

[leaden cargo]

hi

[fervent thunder]

oh I looked at the A1111 repo and I think I see what it is
original A1111 hires fix is two passes with an upscale
but A1111 kohya hires fix (also known as deep shrink) is just one pass

with the downscale at one block

[civic cradle]

any fnaf fans here? i need some people who wana help me create some fnaf loras

[native nest]

Hello!

[final rose]

Grrrr. Finally pulled the trigger on a 3090 EGPU so I can play train SDXL with decent settings. DOA new in box. Apparently the Aurus gaming boxes haven't improved since the 1070 days. (Had 2 of those years ago, also both DOA. One the two started smoking when first plugged in).

[analog grove]

Hi all!

[hallow sun]

hmm

[earnest plaza]

lll

[minor herald]

111

[plain raptor]

boop

m, bored

[covert kestrel]

Is A1111 Forge supports LCM? I attempted to use it but it always said LoRA version mismatch for KModel

[hard swift]

yo

[sage burrow]

can anyone tell me how to genreate images

in this server

[edgy wasp]

111

[vivid warren]

Every workflow I see online uses the same inputs to text_g and text_l on SDXL images. Now that it's been out a while, is the prevailing concensus that its not worth it to separate those inputs? How is it possible that identical inputs is superior?

[fervent thunder]

its probably not superior

[trail lion]

but it's also probably not worth writing separate prompts, ymmv

[fervent thunder]

ye

some people really enjoy prompt engineering but I just write "Photo of a thing with a thing and a thing. There is a thing and a thing"

I only do nouns after a paper showed that nouns have much more effect than adjectives

[fresh ruin]

is it just me or the reactor extension makes the face very "artifical" and loses all imperfections/pores/etc... ? (I can send examples if you want)

[tawny parrot]

https://tenor.com/view/insideout-joy-hi-hey-hello-gif-11341448685299692120

[fervent thunder]

try turning gan down, or off entirely

[fresh ruin]

gan?

[fervent thunder]

usually in faceswaps its either codeformer, restoreformer, GFPGAN or a GAN made by a company called Insightface
there are a few others though
GANs are a model type

if you have heard of ERSGAN, that's also a GAN

[wicked monolith]

Has anyone tried to use OpenPose editor with Forge ? Ive installed the extension but its not popping up

[unborn hedge]

just curious, what's your guys favorite checkpoint/model? My favorite one for humans is DuchaitenXL or whatever it's called

[fervent thunder]

realvisxl 5

[unborn hedge]

I never heard of that one tbh

[fading harness]

silhueta de pessoa seguindo um caminho em direção ao sol e com a cabeça erguida aquarela

[tidal wraith]

Can anyone tell me what ControlNet is used for?

[trail lion]

A way to guide the generation, where things are poses etc

There's only so much you can do with language if you need something specific

[desert meteor]

is there a way to set saturation in forge UI?

[warm junco]

It's bugged. You need to run pip install basicsr inside the venv to make it work

[wicked monolith]

i have no clue what that means haha, thank you though!

[warm junco]

Np, maybe I have time tomorrow to show you how

[dreamy turtle]

I'm planning to make manga with it too

I also have this crazy idea of making anime using frame by frame animations of controlnet

although idk if its possible

[pure stone]

Hiii!

[plain raptor]

so bored

wut yall doin

wuts uh..

wuts the deal with airplane food, amiright xd..

[keen willow]

hi to everyone

[ashen sleet]

gm

[stone dawn]

Hello all !

[hard swift]

gm

[oblique pine]

James Cameron wow🤩

[little ruin]

hi everyone

[low moon]

hi

[wicked monolith]

I would love that, thank you so much

So what do we prefer, comfy ui or Forge?

[desert meteor]

cant get comfyui working in forge anyone know of a instruction video?

just get *** Error running postprocess_batch_list: I:\forge\webui\extensions\sd-webui-comfyui\scripts\comfyui.py
lib_comfyui.ipc.callback.RemoteError: 'bed1b90b-6a59-478d-ad5f-51914f413106'

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "I:\forge\webui\modules\scripts.py", line 916, in postprocess_batch_list
    script.postprocess_batch_list(p, pp, *script_args, **kwargs)
  File "I:\forge\webui\extensions\sd-webui-comfyui\scripts\comfyui.py", line 57, in postprocess_batch_list
    iframe_requests.extend_infotext_with_comfyui_workflows(p, self.get_tab())
  File "I:\forge\webui\extensions\sd-webui-comfyui\lib_comfyui\comfyui\iframe_requests.py", line 121, in extend_infotext_with_comfyui_workflows
    ComfyuiIFrameRequests.validate_amount_of_nodes_or_throw(
  File "I:\forge\webui\extensions\sd-webui-comfyui\lib_comfyui\ipc\__init__.py", line 41, in wrapper
    return function(*args, **kwargs)
  File "I:\forge\webui\extensions\sd-webui-comfyui\lib_comfyui\comfyui\iframe_requests.py", line 78, in validate_amount_of_nodes_or_throw
    workflow_graph = get_workflow_graph(workflow_type_id)
  File "I:\forge\webui\extensions\sd-webui-comfyui\lib_comfyui\comfyui\iframe_requests.py", line 145, in get_workflow_graph
    return ComfyuiIFrameRequests.send(request='webui_serialize_graph', workflow_type=workflow_type_id)
  File "I:\forge\webui\extensions\sd-webui-comfyui\lib_comfyui\ipc\__init__.py", line 20, in wrapper
    res = current_callback_proxies[process_id].get(args=(function.__module__, function.__qualname__, args, kwargs))
  File "I:\forge\webui\extensions\sd-webui-comfyui\lib_comfyui\ipc\callback.py", line 73, in get
    raise res.error from res
KeyError: 'bed1b90b-6a59-478d-ad5f-51914f413106'

[tulip yarrow]

Has anyone trained a flux character pack yet

[jaunty osprey]

yo guys do any of you have a 3D generator running locally? i had tripoSR with pinokio but it broke. On comfyUI i tried several but they're all broken it seems too... any clues?

[unborn hedge]

3D generator?? like making 3D Models??

[jaunty osprey]

yes. Haven't you heard about luma genie, tripoSR, hyperhuman.deemos...? there's a lot around already, but the open source ones are all broken at the moment it seems.

[plain raptor]

tripo is pretty neat

wish there was a stable diffusion equivalent tho

im sure is to some extent, but im also sure it isnt as good as tripo

buh meh '

tripo is pretty gud, at least in my experience, of making a 3d outta stable diff image

[final rose]

I've tried both, I find Invoke much easier to use

[wicked monolith]

ive never even seen that! ill check it out.

[final rose]

Using Onetrainer for training, Invoke for generation. I don't have the hardware, but both are working on Flux support.

[jaunty osprey]

yeah sadly the first model is not working anywhere anymore and their second one is not open source

[unborn hedge]

i never heard of ai 3D models before tbh

[thin lagoon]

Hello

[copper crystal]

So invoke is saying that you need a pro license to invoke or professional flux license in order to use your outputs commercially. Apparantly they got direct confirmation from BFL that you don't own your images unless you pay for a license.

[jaunty osprey]

ur missing out man it's getting pretty good and useful

[quartz siren]

lol idk about nsfw(I doubt), it could probably do individual furry characters though.

[unborn hedge]

huh, I do remember trying it out once and it was literally doodoo water, I'd be interested in something like a blender addon though,maybe one that helps you make a good mesh idk

[final rose]

I have no intention of using it commercially, so I never pay much attention to such notices. My main uses are creative, surreal art for my own enjoyment, and making rude pictures of family and friends for my own amusement.

[jaunty osprey]

jesus christ, TMI indeed

[final rose]

A lot of technical advancements happen because of peoples' pervert wants, but that advancement benefits everyone. For example, PonyDiffusion was made by wierdo Bronies but led to an explosion in Western cartoon/comic styles that didn't exist before.

[jaunty osprey]

i know, that doesn't make it any less gross tbh. I'd agree to disagree because i don't wanna be insulting or anything, but honestly furry stuff for me is nearly as gross as incest and dan schneider's preferences

[copper crystal]

porn exists yup. doesn't mean it's acceptable to watch it in public

[final rose]

You don't have to look at it. Civitai has filters if that's your wish.

[warm junco]

go into the forge folder where the webui-user.bat is
then click into the explorer bar and type cmd and hit enter.
then run the following commands one by one:
venv\Scripts\activate.bat
pip install basicsr
Then relaunch

[copper crystal]

that extension isn't really useful. you're better off just loading comfyui and using it in it's own tab. all it does is put comfyui into it's own gradio tab, and it doesn't interact with anything else in the webui system. AND it's made for webui, not forge. Forge has moved to gradio 4, so many extensions will have problems due to that

Also, the last update to that extension was 8months ago

[trail lion]

what is the point of it? just use comfy if you need comfy

[inner zealot]

Guys, I'm a newbie, I want to learn how to use stable diffusion, but my notebook doesn't have a GPU, I saw that stable diffusion no longer works on Google Collab for free, but if I pay can I use it (I think it's obvious that I can but I want to confirm)? I'm afraid to pay and get screwed.

And which plan do you recommend?

[quartz siren]

Stable Diffusion does work on colab. Also, kaggle is probably a better option(its very similar and free). It's also by google and gives you 2 t4 gpu's instead of just 1 and provides much more ram so its faster and you can run bigger models.

[inner zealot]

Oh, thanks a lot man!

[copper crystal]

i first started with stable diffusion using a collab script. i think they're okay with that. it's the webui's and the people using it for waifu's instead of learning that they frown upon

[unborn hedge]

fair enough, and its not like the furry fandom is exactly doing a good job of making it clean, which kinda ruins all anthro characters because of the assumptions made with it, dont let anyone tell you that a fandom cant ruin something

similar to the concept of ai art, we're hated because of the very nature of our interests and assumptions that people make about it

[wicked monolith]

i think i did it wrong im trying again, will it show up as a tab?

[warm junco]

yep should show up

but the extension is old so it could still not work

[wicked monolith]

is there a better version?

[warm junco]

if you load an image into controlnet and select the openpose preporcessor you can then click on the Explosion icon.
then you get a preview image. There is a button to Edit it. that opens the openpose editor of controlnet

[devout prairie]

Hi there! 👋

I'm seeking a job as a Full Stack Web Developer and UI/UX Designer. 🧑‍💻 I have 8 years of experience in front-end and back-end development, and I specialize in building websites, web services, microservices, and apps from scratch. 😉

If you have any job openings or new ventures arising, please DM me.

Thanks!

[fresh ruin]

is it just me or it seems like Lanczos is not upscaling?

[fresh ruin]

and nearest aswell

like, the upscaled looks the same as the original

[fervent thunder]

nearest is the base line that you compare things to

because nearest is just turning every pixel into 4

for a 4x upscale

compared to nearest, Lanczos does upscale a bit

[robust needle]

Could anyone hook me up with a tutorial for downloading Stable Diffusion that's compatible with AMD? I know I want to create realistic images, and think that also might require a specific download, but as might be evident I'm not super hip to this so an assist would be killer

[peak tendon]

any img2vid for automatic1111? all i find is for comfyUI or is standalone

[molten quest]

Does anyone know how I could set up a cofyui layout to do the following:

I want to input a string that looks like this [Beach, Hotel, School, etc] and then I want to make a different render using each of the elements in the string. Any ideas?

[cosmic cape]

Can someone tell me what is the AI Model that is use to convert a painting to realisitc version? Also, is there any free online website of it?

[plain raptor]

TanTan xD

respect the openess

[steel fable]

Anyone can help me to fix the face from My Lora please?

[cosmic cape]

CAN YOU TELL ME MORE ABOUT IT>

[plain raptor]

bout wut

u wanna see wut they said or..

i got it archived

i keep everything archived

[plain raptor]

to answer ur question

Yes

it can

tripo 3d can do, suggestive material, but, u need to remove the "parts" for tripo to accept it, once done, u can add them back in blender after the fact

[plain raptor]

if ur doing it by image that is

i think tripo has a word verification algorithm that prevents the use of vulgar language.

so, for best results, make an image of OC, or draw ur OC, tpose style, remove the "parts", let tripo make into 3d, load it into blender, make any corrections, add bone and animation after, done =3

[hard swift]

sup

[plain raptor]

well golly gee

ig this is the only server im in where ppl be talkin in

[cosmic cape]

I mean what does it exactly do?

[plain raptor]

ah u mean tripo

its like stable diff

but, u give it a prompt, negative prompt, and based on what u tell iut, it make a 3d model

u get like 700 monthly coins

and u have to pay for mor

u can also give it an image, but model quality varies based on image quality

it also doesnt like the colour black, so if u make a model from an image that has the colour black, it prolly wont render that surface

that really about it

u can find the website at tripo3d

[cosmic cape]

I m talking about TanTan

[plain raptor]

wha

tantan is a person tho, not an it

im confused by what u mean "what does it do"

i was initially replying to tantan's Tmi message '

i thought it was funi how they was like, so unconcerned by wut they said in pub chat xd

[cosmic cape]

And intrestingly, TanTan Ai actually exists which is a complete intregration of OpenAI ChatGPT, Dall-E Image creation and Ai...
lmao

@plain raptor

Conversational Chatbot + Speedy, (built-in) Reddit Media-Fetching

What it is and what it does:

There are many separate tools, the chatbot, the reddit fetcher, the image generator, and the downloader, all in the progress of smoothly integrating into one simple "Siri on roids" that does what you need it to do, and is always a click away.
Chatbot.py incorporates OpenAI's API, allowing access to different models, utilizing the combined power of ChatGPT's knowledge, DALL-E's image creation, reddit's existing content, and more, without the need for a browser, which in today's world, is increasingly vital for privacy from snooping third parties.
The capabilities of Chatbot.py are blossoming fast, and presets for different use cases, as well as better integration with new tools are continuing to be worked on. I would love feedback on presets that are still in their infancy.
This work in progress is just for friends, but as it progresses, it's intended to be a versatile tool to harness the power of AI and design neat browserless strategies to complete all sorts of tasks! Whether you use it to engage in rich conversations shaping your life perspective, to goof off with fun presets, or even to task it with niche developer work that's historically human-only, the bounds truly are limited only by your imagination.

[plain raptor]

oh dang xd

funi misunderstanding xd

yas

[cosmic cape]

A next level misunderstanding that actually turn out to be a knowlegeable..

xD

[steel fable]

There is any face restoration extension or something to install for better faces?

[flint night]

Hi Guys! Could you please tell me if you have any recommendations or collected, working setups for rotoscoping?

[coral ginkgo]

hello everyone, im new to stable Diffusion, can someone help me how to make the ai image generator things?

[wicked monolith]

Anyone can recommend good inpaint models for changing outfits and things in anime/manga style art?

[covert cedar]

Hello everyone!

[tulip yarrow]

Has anyone trained a character pack Lora with flux yet

I searched on civit (there weren't any)

Is it even possible

[iron swallow]

Yo i want to create hyperrealistic images that 99 percent of ppl will fall for and not think its AI and also be able to create multiple pictures of the same Person i generated

locally on my PC as online stuff has Filters in place not allowing you to do certain stuff

can anyone send me a guide to set everything up

[tulip yarrow]

Do you have a GPU?

And even then you need a good one

[iron swallow]

rtx 3060 12gb vram

[tulip yarrow]

Well good you have a gpu

You need a gpu to run stable diffusion

Don't know how to set it up though (mainly because i don't have one)

[iron swallow]

oh alr

#🏞｜general-with-images i posted a example pic

of what level of realism i want

[tulip yarrow]

Just don't use it for anything heinous ok?

[lime marsh]

guys can someone help me ? I want to install a copy of RVC on google colab pro for voice conversion, i already have RVC disconnected that i used to train a model, what one should i use on google colab for voice coversion.

[fervent thunder]

there's nearly 4,000 RVC models on Huggingface

[warm junco]

yes, there is Adetailer

[warm junco]

checkout the pinned messages of #🤝｜tech-support for an local install guide

[fervent thunder]

tiled upscale is also an alternative to adetailer

its easier to setup and gives the same effect

the downside is its like 50 times slower than adetailer

cos its upscaling the whole image instead of a small face

[sharp grove]

Hello guys

[warm junco]

inpainting is also an alternative, in fact its all adetailer does

[chilly path]

Hey all, is there a channel here where you can create stuff like midjourney ?

[fervent thunder]

ah yeah I forgot manual inpainting

I never do it, but it works well

[desert dagger]

read the information here #artisan-faq

[chilly path]

Yeah ty I figured it out

[fading violet]

why is it when i generate an image it ussually takes about 10 min but sometimes it takes like 1

with the same parameters

i dont understand it

[chilly path]

Probably traffic

[fervent thunder]

sounds like full VRAM

[fading violet]

yeah but is the exect same parameters same prompt and all

and its run local

[desert dagger]

you might get better help in the #🤝｜tech-support channel

[rigid harbor]

hey bro can you help me test a lora from hugginface? i've tried using it but the images i'm getting suggests the lora is having zero effect on final output

[desert dagger]

are you making sure to use the lora's keyword as part of your prompt?

[rigid harbor]

yep

[rigid harbor]

this is the lora https://huggingface.co/bingbangboom/flux_dreamscape

[desert dagger]

and you are trying to use it with flux, correct?

[desert dagger]

yeah, doesn't appear to have any effect for me, either

[rigid harbor]

ohh finally so it aint just me. i thought i was crazy

[desert dagger]

I'll correct what I said, slightly, it gives my images a cute feel, but doesn't change the style or do anything else.

[rigid harbor]

can you show me please

[desert dagger]

sure. just a second, i'll post in #🏞｜general-with-images

[rigid harbor]

ok hanks

[desert dagger]

they're posted

[rigid harbor]

it works for you then lool

[desert dagger]

doesn't give me the look of the images on the huggingface space. it DOES do something. it does NOT do what he's showing

[rigid harbor]

well what your getting is wayyy better than mine.

[desert dagger]

mine have the workflow in them if you're using comfy. click the image, then click open in browser, then right click and save as. maybe i'm using a different lora loader than you are

[rigid harbor]

ok thanks i'll try your flow

[celest bear]

Is there like an faq

[quartz siren]

just ask here if its simple or tech-support.

[wise ginkgo]

3060 is a good gpu for SD?

12gb is enough - for most work (projects)?

[umbral vapor]

hi 🙂

[trail lion]

take a look at this page to get a rough idea of where it stands in the hierarchy. If the question is "can I use it to generate images?" probably, but it'll be slow. "Can I train with it?" Not really. Not with any model since 1.5 probably. https://www.tomshardware.com/pc-components/gpus/stable-diffusion-benchmarks

[wise ginkgo]

Oh yeah, I see...thanks....I've been saving up for a 3090 - I think either a 3090 or 4070 Ti Super - are 2 good choices? The 3090 has more vram but if used - either gpu is around the same $$ in my region/area. The 4070 Ti S is probably a bit more expensive, actually - however, it's newer - better power consumption but with the trade-off of less vram (16gb vs 24)... What would u pick? 4070 Ti S (used, though) - which is about $100-$200 more or the used 3090?

[trail lion]

fwiw - I just upgraded to a refurbished 3090 (to save money), best decision I ever made. the 24G vram is absolutely essential. The 4090 will give you the best performance, but for me the cost benefit wasnt there. Good luck on your decision

[wise ginkgo]

interesting... which 3090 model? Actually, I just looked at used prices- some ppl are asking for around the same as some 3090 sellers.... maybe a bit more but pretty close.... either gpu is an option unless I find the 3090 at the lowest price (I've seen).... I suspect they go quickly though...I'm not there yet...

I want the gpu for SD, DR and maybe some Blender work...SD and DR - main programs, though

yes, the vram is pretty beneficial to the programs I will use - at the very least, it allows for some extra vram just in case

the 3090s I often come across - FE models, Gigabytes, sometimes Zotac and occasionally Asus Tuf - for some reason, it's usually those

[trail lion]

in my case I got a 3090 TI for $850 US and beat the absolute crap out of it with training, no issues, I got lucky

[wise ginkgo]

wow, nice score

[trail lion]

it was a gigabyte geforce rtx 3090 ti

[wise ginkgo]

3090 ti here - rare and often overpriced imho

it's often the same price as a used 4080 / 4080 super - sometimes, even more $

[plain raptor]

mmm, amd 5600G with radeon vega7 integrated

[wise ginkgo]

I'll probably pick a 3090, ultimately - since, the extra vram is nice to have and gaming is way down the list of priority ...the extra vram is nice to have

the only thing is that the 3090 is a bit of a power hog? Do you use Windows?

I was gonna dual boot Windows and Linux

[trail lion]

what I would say here is, the vram will definitely bottleneck you on text 2 image type activities, with just gaming, you can get away with far less

[wise ginkgo]

that doesn't sound like it'd be very powerful for SD? 🙂

[wise ginkgo]

16gb?

or do you mean, you need workstation cards?

[trail lion]

I had a 16g amd card before I upgraded, there was nothing I couldnt do, but....it was slow

[wise ginkgo]

ah

which amd gpu?

I wouldn't go with anything less than a 7900 xtx if I went with amd

[trail lion]

5700 XT, and it's still on my shelf

[wise ginkgo]

ah

[trail lion]

it probably isnt even on that chart I linked

[wise ginkgo]

that chart u gave me - the 7900 xtx is only so-so?

[trail lion]

yah, amd doenst do well with SD

it's better in linux

[wise ginkgo]

even with a 3070...pretty disappointing?

I wonder by how much though

I considered AMD gpus but they suck in Blender too...they suck at too many software programs (for productivity)

[trail lion]

even in linux, it would be at the bottom of that chart

[wise ginkgo]

hmmmm

pretty bad 😕

[trail lion]

I'll put it this way, I'm a patient person, so I got by with generating images on it, but training, I used cloud services

since I got the 3090 I can do everything local

[wise ginkgo]

Yeah, I hear u....but, Ithink ...the 7900 xtx isn't a cheap card - performance similar to a 3070....that's awful and not acceptable

interesting to know

[trail lion]

agreed

I used amd for years, this is my first nvidia card, I think ever

[wise ginkgo]

I was going for a 3090 anyway - I was just curious about 3060 performance because it's a cheaper card and I actually owned one once upon a time 🙂

oh really? Do you use Windows?

[trail lion]

linux

[wise ginkgo]

is SD better in Linux?

Really? Which distro? Supposeldy, using Wayland is a better experience

[trail lion]

it works great in linux, I think it can work great in windows, but not on amd

[wise ginkgo]

ppl tell me I am crazy to use Linux but I have used it before.... so, it shouldn't be too bad

right, right

using AMD's rocm and other software requirements - is supposedly a headache or ? 🙂

anyway, I have only used nvidia gpus in Linux myself....so, I don't see a reason to switch

[trail lion]

um, I used docker images, ones that were built for rocm, so my experience was pretty easy with it

[wise ginkgo]

foss is nice but not if the performance is subpar and you might even run into problems and stuff that is not supported

[wise ginkgo]

ah right...docker - some ppl use that for DR too

[trail lion]

https://hub.docker.com/r/rocm/pytorch

[wise ginkgo]

that's when you used the 5700?

[trail lion]

nice thing about it, it doesnt conflict with your main OS

yah

I still use docker even with nvidia

[wise ginkgo]

gotcha

interesting

wish I could get a 3090 now 😦 I sold my previous card.... now, I wish I didn't.... I spent that money and saving again

I had a 3080

however, it was 10gb - I thought it *vram - would be too low

3090s here go for $800 and up

[trail lion]

800 is a good price, I've watched it for a long time

[wise ginkgo]

I probably would have enough if I saved the $$ I got from the 3080 😮

ya, it's coming down in price - but, some ppl seem to think they can get most of their $$ back they spent on it new? lol....some ppl want $1k

[trail lion]

I suffered on the 5700 for years, because I couldnt justify the upgrade because I spent that money on the PC first (because why put a nice GPU into a crappy PC)

[wise ginkgo]

yeah, agreed

good to get a nice cpu and enough ram - the gpu can wait - also, u can get a better one the longer you can wait - even used

[trail lion]

yah, everyone else is chasing the newest new thing...if you're one generation back and refurbished or something, you can get a fraction of the cost

[wise ginkgo]

I do worry about the hardware though - some of them have been mined and there is the memory pads on the back? the 3090 ti doesn't have that concern? But, I worry about it with the 3090

I really would prefer it if I don't need to do a re-paste job 😉

[wise ginkgo]

that's the idea 😉

[trail lion]

agree on this point, that's why I didnt do used. refurbished is different, someone returned it due to a defect and they fixed it and are reselling it

meaning it's basically still new

it doesnt mean you wont have an issue, but then again even with new you can have an issue. a reburbished card has been double-verified, so in some ways you're almost better off than with new

[pure orbit]

what's a trustworthy site for refurbished rtx cards for europe?

[trail lion]

I'm usa, so I just used amazon for my purchase

[desert dagger]

amazon has an EU site as well

[robust vault]

RDNA or ROCM? there is a difference

[unborn hedge]

yeah unfortunately NVidia has a monopoly on these things

[wise ginkgo]

rdna 2 or 3 - both suck - use rocm - and that doesn't help

[wise ginkgo]

right...but, it seems like amd hasn't invested in their gpus - at least, for productivity use.... they're fine for gaming cards

[plain raptor]

ummmm

issei u might wanna see if u have anyting hoggiing ur ram xd

u can maybe set ur args to ignore ram size and jus keep truckin despite no memory, that sometimes helps, but can render ur pc slow, or cause it to hang, however, if too much ram is being used for a long amount of time. and ur pc freezes, more often than not, it'll bugcheck

with stop code VIDEO_MEMORY_MANAGMENT_INTERNAL

unless ur using a card, in which case, itll bugcheck, and youll be on fallback, or integrated

[low moon]

Someone quick make a Friday Lora.

[brave hemlock]

hey

[low moon]

hoy

[shell ravine]

Hi, does this work like Midjourney, same kind of prompts etc?

[low moon]

yes

use the Midjourney LORA to get the same effect.

[raw forge]

mainly for amd

windows directML is like water compared to rocm

i mean it's still lagging far behind cuda, but serviceable

[obsidian wagon]

Hello friends!

[bleak matrix]

Hullo!

How are you all doing today?

[wicked monolith]

Should i get rid of forge and just use comfy? Im very new to all this and comfy looks extremely complicated. None of the tutorials for things online seem to work with forge, should i go back to automatic 1111?

[atomic spindle]

Any room on this disc to ask for a bit of help with lora training? hehe

[trail lion]

sure

if you mean advice or technical questions, maybe ask in the finetune room I'll reply there

[atomic spindle]

Ah, thanks a bunch, just found out the finetune channel was hidden by default here.

[solid spindle]

what prompts i can use for describe sizes of objects? like i want to use cm/inch or "object has the size of something"? "the Bottle has double the heights of the glass" somethig like that?

[trail lion]

yuge, bigly, a wee mink, a nip

that was a jest, dont use those

[sharp fulcrum]

does running stable diffusion on a SSD decrease model load times?

[trail lion]

oh yah

by like 10x or more

so on my setup I have a NAS device where I write all my generated images to, it has like 6TB or something, and is spinning disk. the models are definitely on M.2

[solid spindle]

SSD is recomended

but, well there are some faster HDD outside with 7200U/min...but its always slower

[trail lion]

FYI writes to SSD and to SAS are nearly on par

and you probably wouldnt notice a difference with an image gen

[fallow estuary]

how to create images?

[copper crystal]

the big constraint with HDDs is they have a little magnetic tip on a robot arm that has to scrub over the spinning platter and find the bit it needs. the seek time. SSD's have zero seek.

you can install the software to any drive, and keep your models on an SSD. loading them from storage into active memory is faster when the storage is fast

[fiery wing]

Anyone know the right words to pop into a prompt for SD1.5 or others to get a good "wolf cut" for a character's hair?

[kind oracle]

Hey does anyone have good setting tips on stable diffusion xl for images?

i run it on mac pro with m2 chip

this is my actual tweeking commands export COMMANDLINE_ARGS="--medvram --skip-torch-cuda-test --opt-sub-quad-attention"

[quartz siren]

Mlx is a very good option for macs, Not sure how much ram you have but you can use this: https://github.com/filipstrand/mflux

[kind oracle]

is there any youtube tutorial for that maybe ? @quartz siren

[still glacier]

it's literally one line to install it https://github.com/filipstrand/mflux?tab=readme-ov-file#-installation

(ok maybe two lines if you don't have uv)

[kind oracle]

okay i will try hahah thanks

wich model would you recommend ?

[still glacier]

that's for running flux tho, not sdxl

check out pinned guides in #🤝｜tech-support otherwise

[kind oracle]

okay

[low moon]

Use both, forge for just messing aroudn and Comfy when you want to go deep.

[wicked monolith]

Whats better illustration type stuff?

[quartz siren]

This is for stable diffusion: https://github.com/ml-explore/mlx-examples/tree/main/stable_diffusion

If you have 16gb ram or more, I would recommend flux but it's your choice.

[low moon]

that depends more on the model u pick but probably comfy for illustration as ud have more control

[wicked monolith]

Is it as complicated as it looks?

[quartz siren]

That probably just depends on the person, some people find it complicated, some don't. Basic workflows are pretty simple and won't be that complicated. You will also learn more about how these models work.

[wicked monolith]

Is there a place where people suggest work flows for specific end results, like illustration? It seems most of the youtube tutorials are focused on photo realism stuff and surrealist concepts, thats not really what im trying to do.

[quartz siren]

Most of the times, the style usually matters on the model/lora.

You just require a prompt that doesn’t involve realism and need a good illustration model. You can check some here https://civitai.com/search/models?sortBy=models_v9&query=Illustration

[wicked monolith]

I have so many loras and im still not getting the results id like

[quartz siren]

Prompts will have some effect too, what prompt are you using?

[wicked monolith]

Im using whatever prompts the model description suggests and then some general stuff.

[kind oracle]

did someone used stabilitymatrix?

[plain raptor]

mmm, so many deleted messages in here

red, red everywhere

YOOOOOOOOOOOOO

that moment u tel windows for the 2nd time to stop FECKIN turning ur screen off after 20mins, only to find, it reverts the setting on its own, and goes to sleep >:(.

[low moon]

Initially.

[solid spindle]

ForgeUI - instead of using wildcards how do i create random choice prompts - i got the Jinja2 templates option but the promting seems bit to complicated ['red', 'blue', 'green']...can i just use [red,green,blue] without spaces and without jinja2
Also how to random choice of short Phrases.

[plain raptor]

invest in dogge coin

[austere relic]

Guys I need help. I've been trying to get stable diffusion working on my PC but im using a 6900XT from AMD

and everytime I attempt to generate an image it just tells me that no NVIDA drivers were detected 😦

[desert dagger]

go into the #🤝｜tech-support channel and read through the guides that are pinned in it

[keen minnow]

Got AI Art? We invite you to share your artworks for a chance to get printed in The AI Art Magazine! 👉 We are looking forward to your submissions https://art-magazine.ai

[tawdry ore]

hi yall i need some help on some confusing ai legalities that are too complex for my smol brain and i just need advice not another argument

if anyone wants to join just ping me

[lusty beacon]

Is there a workflow for https://huggingface.co/spaces/jasperai/Flux.1-dev-Controlnet-Upscaler

[brittle rose]

NSFW is not safe for life for me but I still want to have some uncensored political prompts they call it as PG-13 noncompliance

[hollow oxide]

It's possible create epic 3D text, similar to

https://youtu.be/8pcJ-NgmK1g?si=yIIKsbxwzKLbx1fq

[azure nebula]

which flux model is really good and has lot of variety?

mainly want to generate game assets like props and environments based on my sketches or images

[ashen arch]

Flux1d

SD is just as good

ngl

SD1.5 is my fav. Such high quality

[azure nebula]

Yeah been using SD 1.5 long time

It's really good

But lacking new features

[azure nebula]

anything from civtia?

or just install the base one? worth using flux?

[warped heart]

guys why my regional prompter works so bad

it ignores my prompt

only 2 characters

[vast thorn]

Actually I'm artist and do all kind of art work
I didn't want to disturb you, but my friend is in very bad health and both of his parents have died, and he is only 16 years old, so I am collecting money for his treatment. That's why I need a commission. Would you like to help me?
Should I make something for you??
like Emotes, Overlays, PFP 2D/3
models, 2D video animation, YouTube channel banners art, or any other art done for your OC etc

[ashen arch]

the base model

[ashen arch]

it comes down to your gear

Flux1d doesn't even work on my computer

and if it were to work.. itd break it

multitasking would go out the window thats fs

it works on my m1 macbook w 32g ram

but i found SD1.5 to be the most reliable

[warped heart]

guys

who is here

selfattentionguidance integrated which settings is the best for anime?

[fervent thunder]

maybe scale 0.2, blur sigma 2

[vestal dew]

Is Flux the latest model for Stablediffusion?:

Im still using SDXL...SD3 was a big disappointment

[fervent thunder]

So one second I'm generating a cute girl or something and then I add an innocent seeming word and suddenly stable diffusion feels like generating an eldritch abomination
Why is this?

[desert dagger]

the computer grimlins have decided to troll you apparently

[pastel cypress]

hello

[long pelican]

ai art > human art

[young ivy]

you mad bruh? 😄

[tribal hedge]

hello

[tender pilot]

hello

[final loom]

hi

[final loom]

no

[final loom]

human's art has more quality

[frank halo]

Hey

Can someone help me with a problem I have?

I have a problem with Hand Refiner preprocesor in Inpaint section, anything I do gives me an error

If anyone can give me an idea on how to fix this I would appreciate it.

[slender remnant]

Is the server dead

[rugged dirge]

yeah it looks like

[warped heart]

are there better samplers for anime than euler a

[long pelican]

what is there to be mad at? it's my preference

[young ivy]

mad as in crazy

[long pelican]

quality as in how

by what metric

[unborn hedge]

running local llm's is kinda cool tbh

[manic nebula]

how do i generate images here ?

[trail lion]

I was just reading kohya_ss notes, he managed to get training working on 12G vram. so impressed with what they've been able to do

on flux

[warped heart]

guys how do u create good arts with regional prompter my characters always mix

[glossy elbow]

hello

[warped heart]

hi

[trail lion]

what model and gui are you using?

[warped heart]

forge, prefectponyxl

[trail lion]

I havent used regional in a while, but I used to generate one first without any regional prompting, things will mix but that's ok, get the characters where you want them, then use that as a canvas to guide your regions and fix your prompt

[warped heart]

through img2img?

[trail lion]

yah, or you can use txt2image + controlnet depth

[warped heart]

ok i will try

but in most cases it doesnt generate even third character

third and second character mixes

[trail lion]

well 3 is difficult, but are you stating there are 3 subjects iny our prompt?

so like, an image containing 3 persons standing in a room, a woman in a dress, a man in a green shirt, a man in a black shirt.... the colors will mix, but then do it with regional

[warped heart]

i type 2boys and 1 girl

[trail lion]

I cant really give any pony guidance, I've never used a pony model

[warped heart]

i think pony and xl have bad support for regional

[keen birch]

is there any method of auto subtitle generation for local or is that model excessively resource intensive

I've seen the talking head addon for a111 but I know that's a bit different

[full lark]

there is no ai art, only ai image generators

[long pelican]

which is ai art

[trail lion]

now we just need an LLM that can create pointless arguments

[quartz siren]

prompting is different with regional prompts, you usually need to start with a general prompt like "3 characters"

[royal forge]

Hi all, does anyone here understand why Stable Diffusion doesn't understand prompts and makes siamese and many inconsistencies?

[trail lion]

you need to learn how to work with the models, so part of it is understanding the nuances of the text encoder, part of it may be that the model doesnt understand the concept. your question was vague, so the answer is appropriately vague

[crystal spade]

alguem br ae pra me da uma mao

[neat turret]

Where I can créate images ?

[spiral fern]

hi

[crystal spade]

Guys, I want to have a mixed base of photos for Stable Diffusion. I’ve downloaded almost 50GB of files separately, along with some pre-made datasets. I would like to know if there’s a way to mix these ready-made datasets and then add my own images to this base.

[trail lion]

how about on your pc?

[kind oracle]

xd

[desert dagger]

sure. unzip the data set into a folder, add your images to the folder

[crystal spade]

but the file is . safetensor

[desert dagger]

that's not a data set, that's a model

[crystal spade]

oh, and have a mode for join then?

[desert dagger]

you can train a model on your images, and then try doing a merge - but you need to make sure that you train your model for the same base as the safetensors files you have were trained for.

[hot sigil]

hello! I am currently trying to generate my original character, and using txt2img, I found a good starting image. My character has a bandaged torso, like a Sarashi wrap, but in my txt2img the starting image I chose has an all-black torso like she is wearing a skin-tight shirt. I am trying to use inpainting to change this into her wearing a Sarashi wrap, but it is not working with the prompt I'm using, just "bandage wraps, sarashi wraps" and people are generating onto the torso. What can I do to fix this?

[pallid quest]

hi

[vagrant carbon]

hi

[young bronze]

Hi, trying out stable diffusion to fast track ideas on images and stuff

[gleaming flax]

hello

[dusty trellis]

Is this the channel I can discuss Stability Matrix and Flux?

[final ice]

Furry background.

[runic ginkgo]

helloo :))

[fervent thunder]

hi

[steel fable]

How to upscale and adjust pic for 16:9 so they dont get cut like wallpaper?

[fervent thunder]

hi

[deft vault]

Hi.

I have question about Training Lora. Is this the place to ask? Or there are specific subforum?

[ashen sleet]

yo

[rancid forge]

hi

[worn aspen]

I need a new desktop computer, but I don't have the desire to build another one. Anyone have good results with a prebuilt? Any recommendations?

I recently bought a zephyrus with a 4060, and I'm almost tempted to plug that into my monitor, but I don't think it will do for everything.

[brisk mirage]

which is the best performing ui you guys use for sdxl?

[serene olive]

hello I'm new

[bleak matrix]

Hullo, hullo! Welcome!

How's everyone this afternoon?

[true turret]

so guys what is this about nVidia essentially blocking 8GB VRAM cards from generating SDXL quickly with the new drivers?

is this a thing they're doing on purpose or just a glitch in the drivers?

on my 2070 Super I can generate SDXL without issues within seconds

and now I'm too scared to update the drivers

https://www.reddit.com/r/StableDiffusion/comments/1fslym2/if_you_have_a_gpu_with_low_vram_like_3060_ti_8gb/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button

this is what I'm referring to

it's completely and utterly unacceptable that my GPU which works flawlessly with SDXL suddently after a SOFTWARE update no less would become incapable of rendering the most basic things it could render within seconds before.

rolling back the drivers works for some but not everyone I guess

my current drivers are 555.85 and work flawlessly

[fervent thunder]

are you sure it was incapable of rendering?
seems like it was still able to render, it was just slower

[true turret]

incapabe of doing what it could do before the update

20 seconds to 2 minutes pretty much makes it incapable in my book

[fervent thunder]

that's not the same as actually failing the render though
what has happened here is that some of the layers of the Unet have been kicked out of VRAM and on to DRAM

[true turret]

so what if it isn't the same, you're getting hung up on the wrong thing

the point is it causes problems for AI generation

[rich torrent]

Hi

[warped heart]

guys if i want to steal pose(not style or any thing) from another art which models in controlnet should i use

[trail lion]

depth map probably

or canny or both

[warped heart]

yeah canny

th

thx

[heady parrot]

hey guys is there someone who can help me with prompts ? stable diffusion Foocus

[true turret]

The way I prompt is by keeping it simple
first I put the type of image I'm looking for and color grading - this can also go last depending on the model/checkpoint.

then the subjects and base action if needed
description of subjects and their actions

environment

I will never understand poetic prompting, I keep it simple and clean and always get the desired results

I've saw some idiot on youtube once prompt "girl wears a hoodie which she loves very much" why does the AI care if the girl loves her hoodie very much? If you want the girl to have a loving expression just type that "girl wearing pink hoodie" "girl has loving expression" something like that

[robust pivot]

is there any huggingface bots that are currently working that run devilishphotorealism sdxl?

if not that, is stable horde worth getting?

[fervent thunder]

its just a case of memory management
unet layers getting kicked out of VRAM is optional

[true turret]

and yet when people disabled it, the problem wasn't fixed

I think it's something to be wary of if the fix isn't as simple as disabling it so it doesn't rely on System RAM

[fervent thunder]

it needs to be disabled in the software as well as Nvidia settings
would recommend Comfy UI with --highvram flag

[true turret]

as long as it doesn't affect my InvokeAI and fooocus it's whatever, if I can disable it

but if it does we'll have a problem for sure

hope the few new games I want to get by the end of the year like Visions of Mana and Methaphor don't require new drivers, the demos work fine on mine

[fervent thunder]

hopefully it will be okay
when you are this close to the VRAM limit (within a few GB or so) its tricky
cos the tiniest thing can fill up your VRAM

I've used 8GB VRAM cards on cloud a fair bit for SDXL and I got out of memory issues a lot on comfy

[brazen scarab]

Just started using Reforge and it's generating XL images very blotchy, why is that? On Forge I select XL option and it generates nicely.

[steady fjord]

hello

[true turret]

never had issues with my 2070 super 8GB with fooocus or invoke and I'd generate and inpaint and what not for hours on end. I use t2i adapters to turn my own drawn outlines into photos and it has always been fine. Sure superior control net models do take a lot of time due to the lack of VRAM but t2i adapters don't and they're more than enough

[bright wren]

will pulid works on image to image?

[idle trench]

Can I create a picture with words here?

[bleak matrix]

Good morning, everyone! How are we all today?

[analog delta]

Hello

[analog delta]

Please are you the owner of this community

[bleak matrix]

👋 Hello! And no; I'm just a moderator who volunteers her time here! 🙂

[analog delta]

Then who's the owner

Cause I will love to volunteer as a admin or a moderator

[weak smelt]

how many steps does everyone here use for their images?

[peak tendon]

aloha, I'm trying to switch from Automatic1111 to ComfiUI, and I'm doing the thing where I edit "extra_model_paths.yaml" to use the same models I use on Automatic1111, but I'm confused on when it asks me where my Clip folder is
Where the hell is that?

[weak smelt]

u using stability maxtrix?

[peak tendon]

Same for the Controlnet path

[peak tendon]

I have no idea

[weak smelt]

logo looks like this

https://discordapp.com/channels/1002292111942635562/1004159122335354970/1290714806315389106

[peak tendon]

nope, I am not using that

just searched a tutorial on how to make comfyUI use the same models as Automatic1111 and it told me to do this

[arctic coral]

Hi 😃 is there a guide somewhere about different Stable Diffusion generators? Like I'm using Automatic1111 so far but I saw people saying forge UI was much better because it gets updated more frequently or so..

So I wonder what programs are out there and what the differences are, before I try and learn Automatic1111, because except for the very basics which are simple, it does seem rather complicated to me personally

[weak smelt]

try stability matrix

[arctic coral]

What is it about or what's it "better" at than others?

[weak smelt]

easier for beginners, has better ui

[arctic coral]

Can it use the same models/Lora and such as A1111?

[weak smelt]

yes, i use mine from civitai

[arctic coral]

Me too okay, I'll look for a guide to set it up and try out .. thank you:)

[cinder vessel]

Hi, im looking for an api to image mask extractor with prompt like CLIPSEG2

[hollow current]

hmm... pop_os, mint, or debian.... can't decide which one to play with next for SD + Ollama + OpenWebUI

[still glacier]

doesn't really matter, they're all debian based, or debian.

[woven rapids]

Hi!

[clever lichen]

Hi

[warm junco]

you dont need stability matrix, as its not a webui, its just an installer

for auto1111, forge, comfyui etc

Best is to start and learn with Automatic1111
Then you can also use Forge as its a fork of Auto1111 with advanced features but less extension support

[arctic coral]

Learned that too, thanks for confirming it. With a1111 I fail with the simplest tasks already however. I can't seem to be able to follow any tutorials when I try things like swapping my face into the generated person or changing out the subject, fix hands by in paint all that sorts of stuff.

Actually getting consistent characters is what I mainly want to achieve. I can also ask for help in #📝｜prompting-help again, I get that this is getting a bit off-topic for #💬｜general-chat

[peak tendon]

so what do I do? I asked how to solve my issue with the paths of automatic1111 and idk why canttel keeps telling everyone to just use stability matrix

all I want to do is use comfy ui because nothing is compatible with Automatic1111, but it won't work because it has a bunch of troubles

[warm junco]

The clip folder is in comfyui models/clip

Does your extra model path .yaml file works ?

[hollow current]

true, but they also all have their own quirks... I've played with Deb and Mint in the past, installed Pop once in the past... so gonna give Pop another try just for kicks and grins

[gaunt egret]

hello

[minor iron]

does any one has experience, uisng comfyui on google colab? i am trying to install the comfyui manager but it is not successful

[true turret]

I am not advertising or anything, I have nothing to do with this, but I wanted to highlight a flux lora that accidentally fixes hands and feet even tho its primary use is not for that

https://civitai.com/models/684810/flux1-dev-cctv-mania

Just don't add the CCTV Footage trigger and the lora does its magic

it might lower the saturation of your images but I don't mind given how over saturated flux can be, it makes everything look realistic

while giving you fantasic poses, hands and feet 90% of the time

I've been using it with Flux Fusion
https://civitai.com/models/630820?modelVersionId=705611
Which is already good at this stuff and at making more interesting people
but adding this lora to it does wonders IMO

[fervent thunder]

could probably find one that fixes hands that isn't blurry CCTV style

I've had this happen before where I liked one part of a lora, but the actual overall lora style was not helpful 😄

[true turret]

I wouldn't say it's blurry at all

for the type of images I like making

[pearl oyster]

If I gen something sfw with a model that is capable of nsfw, is it permitted to post here? I can guarantee that there's nothing nsfw about the image. People's brains might crash from the trippy-ness, but nothing that's along the lines of what is mentioned as not being allowed.

[fervent thunder]

okay that's good
if the blur doesn't come through

I use some movie loras on SDXL sometimes that bring some blur with them

[sly olive]

Hi everyone, does stable diffusion 3 medium support IP adapters?

[floral umbra]

Hoi, does forge have a page with it's environment variables? As i need to add custom folders to forge to point at folders in comfyui for controlnet models

Or does it use the same as auto1111?

[open escarp]

Hi everyone, just a little promo, if you like the comic style, come check out my latest LoRA. I hope you'll like it. Have fun!

https://civitai.com/models/813172/marvelmixx

[lofty kite]

I'm 𝐋𝐮𝐜𝐞𝐧𝐭 an 🏆 AI SPECIALIST ON UPWORK 🏆 and am 💎You are Free to DM if you have any questions concerning this💎
An 𝑨𝑰 Phone 𝑽𝒐𝒊𝒄𝒆 Calling 𝑨𝒈𝒆𝒏𝒕 Developer for 𝑶𝒖𝒕𝒃𝒐𝒖𝒏𝒅 & 𝑰𝒏𝒃𝒐𝒖𝒏𝒅 call, 𝑨𝒑𝒑𝒐𝒊𝒏𝒕𝒎𝒆𝒏𝒕 𝒃𝒐𝒐𝒌𝒊𝒏𝒈, 𝑳𝒆𝒂𝒅 𝑸𝒖𝒂𝒍𝒊𝒇𝒊𝒄𝒂𝒕𝒊𝒐𝒏𝒔, 𝑴𝒖𝒍𝒕𝒊𝒍𝒊𝒏𝒈𝒖𝒂𝒍, having experience on 𝐕𝐀𝐏𝐈, 𝐁𝐥𝐚𝐧𝐝 𝐀𝐈, 𝐒𝐲𝐧𝐭𝐡𝐟𝐥𝐨𝐰, 𝐀𝐢𝐫 𝐀𝐈. I also help with 𝑴𝒂𝒌𝒆.𝒄𝒐𝒎, 𝒁𝒂𝒑𝒊𝒆𝒓, 𝑷𝒂𝒃𝒃𝒍𝒚, and 𝐧𝟖𝐧 for API integration & Automation and different AI chatbot development using Manychat, Uchat, Botpress, Wati, Voiceflow, Sendpulse and Tidio for website bots

[low moon]

flux is so good - u cna make loras for it easy

[iron swallow]

can u help me with flux

[low moon]

download it and use it

[iron swallow]

how do i get it to comfy ui

[low moon]

i use GGUF there

works like a charm

[carmine kestrel]

anyone interested in this? just got pushed to beta recently, I can't use it on my hardware but maybe someone else will find interest https://huggingface.co/ostris/OpenFLUX.1

[vivid warren]

Are there any comfyui nodes that can feed a different seed to multiple KSamplers, so you don't need to control 4x different seed inputs?

[earnest axle]

Hello all

[exotic flax]

Anyone got a good gui that i can use with the latest python

AUTOMATIC1111 fails to install with latest python

[fervent thunder]

could you use an older python version?

using the latest python is rarely adviseable

[exotic flax]

I cant as i have another script that is using it

[fervent thunder]

but can you run that separately on a different python version?

[exotic flax]

I can try

[fervent thunder]

virtual environments with venv or conda, or otherwise, containers like Docker

[exotic flax]

this is what i get Windows PowerShell
Copyright (C) Microsoft Corporation. All rights reserved.

Install the latest PowerShell for new features and improvements! https://aka.ms/PSWindows

Loading personal and system profiles took 680ms.
(base) PS D:\sd.webui> python --version
Python 3.10.14
(base) PS D:\sd.webui> run
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.10.1
Commit hash: 82a973c04367123ae98bd9abdf80d9eda9b910e2
Launching Web UI with arguments:
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
Traceback (most recent call last):
File "D:\sd.webui\webui\launch.py", line 48, in <module>
main()
File "D:\sd.webui\webui\launch.py", line 44, in main
start()
File "D:\sd.webui\webui\modules\launch_utils.py", line 465, in start
import webui
File "D:\sd.webui\webui\webui.py", line 13, in <module>
initialize.imports()
File "D:\sd.webui\webui\modules\initialize.py", line 39, in imports
from modules import processing, gradio_extensons, ui # noqa: F401
File "D:\sd.webui\webui\modules\processing.py", line 18, in <module>
import modules.sd_hijack
File "D:\sd.webui\webui\modules\sd_hijack.py", line 5, in <module>
from modules import devices, sd_hijack_optimizations, shared, script_callbacks, errors, sd_unet, patches
File "D:\sd.webui\webui\modules\sd_hijack_optimizations.py", line 13, in <module>
from modules.hypernetworks import hypernetwork
File "D:\sd.webui\webui\modules\hypernetworks\hypernetwork.py", line 8, in <module>
import modules.textual_inversion.dataset
File "D:\sd.webui\webui\modules\textual_inversion\dataset.py", line 12, in <module>
from modules import devices, shared, images
File "D:\sd.webui\webui\modules\images.py", line 22, in <module>
from modules import sd_samplers, shared, script_callbacks, errors
File "D:\sd.webui\webui\modules\sd_samplers.py", line 5, in <module>
from modules import sd_samplers_kdiffusion, sd_samplers_timesteps, sd_samplers_lcm, shared, sd_samplers_common, sd_schedulers
File "D:\sd.webui\webui\modules\sd_samplers_kdiffusion.py", line 3, in <module>
import k_diffusion.sampling
File "D:\sd.webui\webui\repositories\k-diffusion\k_diffusion_init_.py", line 1, in <module>
from . import augmentation, config, evaluation, external, gns, layers, models, sampling, utils
File "D:\sd.webui\webui\repositories\k-diffusion\k_diffusion\config.py", line 6, in <module>
from jsonmerge import merge
ModuleNotFoundError: No module named 'jsonmerge'
Press any key to continue . . .

[fiery wing]

Yeah you just can't use the latest Python with A1111 nor would you want to

Creating a virtual environment for a version specific for A1111 would be best.

[cobalt reef]

Hello,

I bought a package 30 minutes ago and have received access data by e-mail. However, I cannot log in with this data. Who can I contact or how can I write a ticket?

and I keep getting “Thank you for subscribing” emails from stable... It's already the fourth one in 20 minutes... Is this normal?

[rocky marsh]

Hello 👋
Good day Who can i discuss with regarding partnership & marketing proposal?

[peak tendon]

So, now I installed ComfyUI trough stability matrix, but for some reason no matter what prompt I pull Up, it stays at 0%, any idea why?

[peak tendon]

with me

[rocky marsh]

Okay sir

[burnt kite]

Anyone know what's the best Stable Diffusion I can run with an AMD RX 570 (8GB VRam)? I can run SD1.5, but last I checked the WebUI can't run anything higher than SD1.5 on AMD.

[warm junco]

If you go with ZLUDA you can run sdxl too probably

[warm junco]

Try install comfyui without stability matrix.
If that doesn't work come to #🤝｜tech-support

[echo kite]

anyone know some good models that come close to NAI's latest model's quality?

since rn NAI is kind of the "arcade perfect" in terms of anime models from what I have seen

combining novelAI imagegen quality with the ability to use extensions and Adetailer would be great

[fervent thunder]

SD3, Flux and CogView-3Plus are all stronger

the new NAI model is SDXL, trained to handle V-Pred and ZTSNR like CosXL can

but the VAE is the issue

[wary belfry]

maybe it's downloading the model in the background. So it stays at 0% until the model is downloaded, which can take some time

[robust helm]

Images#3d

[pearl oyster]

i'm not sure zluda would work with such an old gpu. it is a 570, not 5700

[echo kite]

I would also be curious to know, I have a 5600 XT

I can run sdxl but it is SLOOOOOOOOOOOOOW

[ashen sleet]

sup

[hard swift]

hi

[burnt kite]

What is ZLUDA?

[peak tendon]

I'm not sure, I left it there for a big while and it won't do anything

[warm junco]

@pearl oyster Zluda can work with a 580 so 570 could work too. But its tricky

[warm junco]

5600xt can work with zluda

[echo kite]

wait, so this means that even with an older non cuda supporting GPU I don't have to wait like a minute every image?

(I generate using cpu rn)

[warm junco]

Yep should work

But idk how fast it is. But for 1.5 I guess much faster than 1 minute

[warm junco]

Something that makes cuda work with amd.
Enabling faster generation speeds and less vram usage

[burnt kite]

Can you please link me a guide to it?

[warm junco]

https://github.com/CS1o/Stable-Diffusion-Info

[burnt kite]

Thank you very much!

[pearl oyster]

how many sampling steps do you have it set to, that cpu genning only takes a minute?

[echo kite]

20

I think

also I have 32 GB of ram

how long does it take for you???

oh

also, I use SD 1.5 instead of SDXL 95% of the time

[burnt kite]

It say this:

For RX 6700 or 6700XT download Optimised_ROCmLibs_gfx1031.7z
For RX 6600 or 6600XT download Optimised_ROCmLibs_gfx1032.7z
For RX 580, 5600/XT, 5700/XT, VEGA 56, VEGA 64 and Radeon VII or RX7700S you need to download the ROCmLibs.7z
For AMD 780M-APU you need to follow the steps here: <github link>
```But I have an RX 570, which one do I use?

[pearl oyster]

using sdxl(pony, specifically) at 20 sampling steps a single 896x1152 image takes 45 seconds. but i usually run several batches at once and pick the best ones

[echo kite]

oh

I usually do 512x

[pearl oyster]

512x512 isn't optimal for sdxl. the exact resolution that's optimal varies by model. for pony 896x1152 and 1152x896 are among the optimal resolutions. there are others, but they're all larger than 512x512

[echo kite]

^

SDXL takes waaay too long an a CPU

[dusk carbon]

how do people create their own model AI ? i want to train my images but locally is that possible i never did that

[jovial igloo]

guys im new and need a lil help

so i have some ai generated sketches from midjourney that all look like pencil line art. What would be the best way to color them in stable diffusion?

[pearl oyster]

you're not training your own full model from scratch. that requires a ton of processing power. what you can do is train LoRA's and merge models. those are much more achievable for the average person

[warm junco]

Same as rx580

[warm junco]

Controlnet canny or lineart anime

[dusk carbon]

yeah i want to do it locally but im confused where to download and how to run it

[pearl oyster]

your own from scratch?

cuz that's most likely not happening

[dusk carbon]

i dont mind any opensource

since im not that advanced at coding

[pearl oyster]

i'm not talking about where you get the images. i'm talking about using images obtained by whatever means to train a full model from scratch. that requires serious computational power that is out of reach for most of us, unless we're flush with cash

[dusk carbon]

oh okay then i don't want from scratch if that takes a lot power bec the bill gonna beat new record

[steel fable]

I often try to create an image in a prone position (belly) or something like that, but the AI ​​doesn't recognize the commands most of the time and out of 10 images that I create only one comes out right, is there any way to avoid this and be more precise?

[pearl oyster]

are you using a vae? that can possibly help. also, another option is to use a lora to explain the concept of being prone to the model

[steel fable]

Im trying to use Lora and still the same, maybe some extension or something to help?

[pearl oyster]

is the lora for the same type of model? you can't use xl loras with 1.5 and you can't use pony loras with xl. you have to use a 1.5 lora with a 1.5 model, xl lora with a xl model, pony lora with a pony model, etc

[steel fable]

Both Lora and model are 1.5

[pearl oyster]

you can adjust the lora's weight to nudge the ai to pay more attention to it

[steel fable]

nudge u mean let 0?

[pearl oyster]

you're not in control like you think. think of it as telling a person behind a glass barrier what you want and them trying to do the right thing. nudge as in how you'd use words to get that person to pay more attention to a specific detail

[steel fable]

Yeah but even with lora is not working

[wary belfry]

you can use ControlNet and get the exact poses you want

[steel fable]

I never use controlnet, but i already install here bc i want to learn, specially to make animated videos

[wary belfry]

it's super powerful to get the exact image compositions you want 😉

[jovial igloo]

guys which user interface should i be using? new here

[fervent thunder]

comfy ui

[jovial igloo]

thanks.. seeing most tutorials on how to install show automatic1111.. Also see forge ui which is supposedly faster. Most tutorials are from like a year ago tho. Can you or anyone else chime in on these other user interfaces?

[fervent thunder]

Forge is a fork of automatic1111, its been upgraded a bit though

[wary belfry]

Use Comfy UI. It gets the new stuff first most of the time and is super powerful. Because of the nodes you have ultimate freedom to do stuff. In A1111 you always have to hope there are predefined buttons for the workflow you want to do. And in ComfyUI you can do other stuff like using CogVideoX to generate videos

[dusk carbon]

i trained poop images

and got banned

[dusk carbon]

i accidently gave the realistic image

it was supposed to be a plushie or cartoon thing

i rushed up to save as many images as much