This model is going to be insane

#37
by Ainonake - opened

The prompt understanding and quality in v31 is already very impressive (hands and body stuff is still nightmare though).

What's the expected generation time? I'm currently getting 1 minute per image on 3090 with 26 steps.

My current goto model is illustrious (wai) but it's overcooked and prompt adherence is bad.

Yep, it's getting more awesome with each epoch. It's still sometimes leaning too much towards animation/3d render style, even when the prompt seems to be strictly real-world visuals, some words are just too powerful, even LORA sometimes fail to dictate the correct realistic style.
1min. for 1024x1024 sounds about right, 1:05 on my 3090 with Euler.

Hands can be helped a bit with LORA, there are some for Flux and they do work in most cases. For Forge there's some code that needs fixing, otherwise it blocks the LORA:
https://github.com/croquelois/forgeChroma/issues/4

For a noticeable speed boost, use ComfyUI + torch.compile + sage attention.

Using FP8 (a speed advantage of RTX4000+ series because of fast matrix ops) and 30 steps (Euler/OptimalStepsScheduler), this takes 36 seconds on RTX 4070 (12GB VRAM):

image.png

And yes, this model is incredible.

I also recommend using the following node as an image post-processor, with setting close to these (but of course it depends on your artistic preference):
image.png

I agree with you @HMRMike , I'm into photo realism / rendering and still working on it.
Maybe I can share my humble / non-expert feedback, hoping this maybe help:

  1. A simple euler / beta seems fine now (wasn't necessarily the case earlier). But I must admit, I tend to like euler_a, and even ddim_u, if I get wild.
  2. In my recent adventures I traveled back into the 70's: I wanted a vintage look and got into this page. It's about SD, but it's interesting for realism.
    2.1. f i add This photo was captured with Bolex H16 at the end, it brings an incredibly lively scene and realism (and a unique "I am a Rolling Stone" feeling).
    2.2. Also checked with Leica M10, pretty good, but not tested with others lenses sorry.
    2.3. If you have a better ref for Flux for cameras / lenses, I'd be grateful.
  3. I dislike tags, and I get illustrations most of the time, but the following seems to improve the rendering:
    3.1. Just add photo taken with Bolex H16 (or other) at the end
    3.2. fun fact: you may get watermarks with Bolex H16 written sometimes, don't know how to get rid of them...

I hope you'll find this useful, and I'd love to get your feedback.
Take care @HMRMike and friends.

PS: in Chroma we trust.

I hope you'll find this useful, and I'd love to get your feedback.
Take care @HMRMike and friends.

PS: in Chroma we trust.

Many thanks!
Seems like staying away from tags or terms that are too typical of booru style prompts is the safest route. Sometimes describing some unrealistic body proportions or features, or keywords like "detailed" instantly switch to 3d/drawings.
I don't like too verbose prompts and it seems that strong LORA like GrainScape_UltraReal (and others from Danrisi) can do the trick in general, along with strongly denoised (0.6-0.7) hires upscale- since Chroma itself isn't familiar with high resolution yet, LORA that's trained on better content seems to fill this gap very well.

For a noticeable speed boost, use ComfyUI + torch.compile + sage attention.

Using FP8 (a speed advantage of RTX4000+ series because of fast matrix ops) and 30 steps (Euler/OptimalStepsScheduler), this takes 36 seconds on RTX 4070 (12GB VRAM):

image.png

And yes, this model is incredible.

I also recommend using the following node as an image post-processor, with setting close to these (but of course it depends on your artistic preference):
image.png
o
Can you share an example of your workflow with sageattn and the optimized speed options?

I am encountering an issue, i get noise Image when using torch.compile with torch2.8+cu12.8+triton-windows 3.1.3post19 , which is quite troubling. However, I am already deeply immersed in chroma, and I love this model. Thanks Bro!

the guy behind this model is like kind a of god for me...

i really like to know whats the background of this AI creator : studies, how he learn informatics, informatic languages he knows etc.... and as soons as i know i will start learning informatics arcanes through same training programs he followed....

@tuolaku please try to remove --fast from comfyui launch args if it is there
Also try to use a stock 'torch.compile' node (as advanced ones often do not work well with Chroma for some reason)

@tuolaku please try to remove --fast from comfyui launch args if it is there
Also try to use a stock 'torch.compile' node (as advanced ones often do not work well with Chroma for some reason)

NICE,its working now。Thank you so so so so much for your guidance.

Sign up or log in to comment