SG161222

/

SPARK.Chroma_preview

like 62

I'd say this is a definite improvement over base Chroma. I haven't seen any anatomy issues, if anything I've seen improved anatomy.
The only issue I've had with it is if you put in the adult caption "woman riding a man in a chair having sex cowgirl style" or something like that there's horizontal line errors throughout the result. Every other generation prompt seems to work fine.

this model + chroma-flash-heun_r64-fp32 lora + lenovo-chroma lora , merged and turned into q8_0 gguf ; @ 8 steps. using sa_solve_pece & sgm_uniform , cfg1.

First tests (both without Lora and with lightning Chroma Lora) are globally quite positive for me with more realism overall, better skin texture... everything what I would expect from you :)

I do encounter some issues where SPARK produce SOMETIMES lower quality image compared to ChromaHD with hands (deformations, placement) and faces then they are small in the image (lack of details). But overall 80% of the images are an improvement over Chroma if you want realism (did not test illustration and other non realistic styles yet)

At this stage, I would need to create an image with Chroma and with SPARK and then choose the best one (~80% of the time it would be SPARK) while I guess the goal you have is to be able to replace Chroma and be confident that almost every image will be better with SPARK.

I did not change the settings for the images compared to Chroma though (CFG/steps/sampler/scheduler): maybe SPARK needs slight adjustments?

But still a very promising "preview" I would say. Good work and thanks for the time you put in this work.

@SG161222 Great Job! Here are some examples! FP8 version used, rest of the files used in the workflow are from Qwen-Image Edit suited for the main FP8 version. I will do more testing and customisation this week.

I wanted to mention something else as well: I've trained a few character loras on both Chroma and Chroma-Spark and I'm finding the ones trained on Spark turn out much better with more likeness to the person and less "blockiness".

Thank you all for your feedback. You are helping me a lot with it.

Like the others said: much better realism compared to base-Chroma, with less plastic skin, deformities, etc. This checkpoint is what convinced me to switch to Chroma. It avoids the "AI Slop" look that I abhor. Some tests I did: one was in cinematic screencap style, the other ukiyo-e and oil painting. As you can see, it also does those styles quite well. I haven't had much success with western animation styles (i.e Disney), nor have I tested it with anime-style prompts just yet

Also works with Chroma Forge: https://github.com/maybleMyers/chromaforge?tab=readme-ov-file

Sampler Settings: res_2s / beta57 / 15 Steps / Chroma-Flash-Heun - Rank 256 Lora (S: 1.0) / CFG 1.2
(Post Processing / Grain Filter)