ComfyUI (Stable Diffusion UI), Quick Start with local installation on Linux
HTML-код
- Опубликовано: 3 окт 2024
- ComfyUI is a web-interface for Stable Diffusion using a nodes system for generating complex workflows; this is an intro video for starting using it on Linux, although it requires a really powerful hardware
ComfyUI on Gituhub
github.com/com...
Civit AI
civitai.com/
Stable Diffusion Samples artticle
stable-diffusi... 
Наука
i really cant estimate my video timings anymore ..it was supposed to be so super short :( but i'll make up to you, will do one video for creating UIs and one on creating consistent images; with renting GPUs obviously coz sooo poor also :p
Clicked for the tutorial, found an emanation of wisdom delivered by the voice of an angel.
never seen anything like you on yt before but it totally worked on wsl 2 using nvidia 3080 - thanks for the tutorial! my image took 10 seconds to generate btw :)
Well, I downloaded it and then spent full night tinkering with it. Never thought AI image generation could be so fun! Thank you for the video, will wait for the next tutorials from you.
try out the lcm sampler, you will need the lcm lora and set the cfg to 1; then you will only need 4 steps
That was philosophical bro but exactly what I was thinking the other day. If our perception of time can slow down does that mean we can have a relatively infinite last thought?
Me who uses Rocky Linux had to install python3.11 stand-alone anyway and mess with environments. 😢
Also CFG controls how closely it follows positive prompts assuming the negative prompt being like electrical ground. That's why if you run with cfg=0.0 and something in the negative prompt you will get roughly what the negative prompt described. Positive Prompt is completely ignored at cfg=0.0 but negative prompt still matters.
Also what intro?
first time i hear someone using Rocky on this channel!
( gonna check that cfg trick!!)
Basically the negative prompt wasn't intended to be used. It's use and spread is entirely a cargo cult.
The algorithm was intended to be used like:
Take noisy picture. Denoise with prompt.
Take copy of noisy picture. Denoise without prompt.
Compare both results weighted by cfg using the sampler.
What negatives prompts do is don't use an empty prompt to compare to but the noise guess made from the "negative prompt".
So negative prompts are not negative in embedding space, they are subtracted in image denoising latent space. What you are actually doing negative prompts is mess with the GPS of the denoiser based on something you copy pasted from someone on the internet instead of using the AI as it was trained for.
Yes it gives you more illusion of control but not if you just copy paste stuff from the internet. 😂
Something far more interesting interesting IMHO is working with changing the positive prompt. Certain things like hair color, eye color and so on settle in different order usually. Like you can take the seed and prompt for a green haired girl, swap out the prompt for red haired girl mid diffusion and end up with a red haired girl which has exactly the same pose and look as the green haired girl did.
Experimenting with that requires more compute though. Although there is a trick with advanced generation nodes so you can set it up to divide your sampling into multiple steps, and reuse old result (if seed and prompt stays the same) and end up with identical image to single sampling performed by a single node. Let me see if i find my notes on that.
@@modernsolutions6631just share your workflows (if you have them on Comfy)
I'd rather explain principles than workflows. (Don't have workflkw at hand either)
Basically a KSampler(add_noise=disable, noise_seed=1, control_after_generate=fixed, steps=20, start_at_step=0, end_at_step=10000, return with left over noise
=disable) returns a pixel perfect identical image to these two chained (assuming your scheduler isn't bugged).
KSampler(add_noise=disable, noise_seed=1, control_after_generate=fixed, steps=20, start_at_step=0, end_at_step=10, return with left over noise
=enable)
KSampler(add_noise=disable, noise_seed=1, control_after_generate=fixed, steps=20, start_at_step=10, end_at_step=20, return with left over noise
=disable)
You can now keep refusing the latent of the first 10 steps with you just tweak around with the later parts of the pipeline. 🎉
Rules to get pixel perfect match: steps describes the total amount of steps in an equivalent non advanced pipelines. Start at steps of the first element is zero. End_at_step is shared with the Start_at_step of the last one. end_at_step and steps must match at the end of the last thing in your pipeline.
The return_with_leftover_noise must be set enabled for all but the last advanced KSampler in your pipeline. Returning with leftover noise makes image a bit worse but that noise is needed for the in between steps.
You now have the best possible control over the generation process. You can also divert from that in a bunch of fun ways, if you do that the jump of the value in the scheduler can cause unexpected behavior. sgm_uniform seemed robust to that.
Operating with fixed control_after_generate is a good idea. As you are actually trying to improve an image, not just make superstitious changes and pull the lever on a slot machine and imagine you are doing stuff if you are just doing ineffective/harmful/superstitious things otherwise.
what file explorer did you use?
currently using webui, 6900 xt on Fedora... Ill check this UI later... tho it seems not that intuitive
thx for the video!!
xoxo lurker
Using rx5700xt. Torch just wouldnt recognize the gpu.
use rocm
Maybe some day I will learn how to run this god damn program. No help from this tutorial.
Won't it cause feature problems with pacman if you just install things to the system with pip?
0:13 on fedora 39 I just did `dnf install python-3.10`, it works and nothing exploded (`which python` is still pointing to 3.12, but `which python3.10` is pointing to 3.10)
umm, completely missed we could install diff versions of Python from repos
No match for argument: python-3.10
Error: Unable to find a match: python-3.10,
Eh? How come my Fedora 39 can't find it? XD
Imagine getting Mii's GPU and her ghost pops up on every ai-art you generate on it. 💀
What distro is this?
Only a 1060? Are you poor? Lol.
Some of us live in shitland.
when will Mii unlock sex?
when you super thanks her for $500 ..multiple times!!
@@mii_beta deal O⍵O