>>3082121

>my problem is more that it's not consistent sometimes where the streams themselves go

You won't get anything better without more work. "Not always consistent" is basically image gen's subheading. Cherry picking mass produced gens is a legitimate approach, just skip any hi-res and ADetailing steps until later. Forge makes it easy, I am sure Comfy has a way to activate those detail steps after the fact as as well.

>I haven't tried anything to do with Chroma or Flux yet, so I'd be curious where to start for that

https://huggingface.co/black-forest-labs/FLUX.1-dev

The good news is Flux (and Chroma, which is really just a retrained version of Flux) operates pretty much the same as SDXL (which Illustrious is a retrain of). Flux doesn't have a negative prompt (Chroma does have negative, at the cost of computation time), and requires you to make a detailed paragraph in English about what you want to see. This paragraph should included poses, styles, colors, shadows, etc. etc. as well as the basic characters and actions SDXL requires as well.

To put it more basically: SDXL needs you to be short, concise, and on topic. Flux needs you to be almost verbose, but you still need to avoid "purple prose".

The bad news is Flux basically doubles your system requirements. 24 gigs Vram, 64 gigs Ram, 100's more gigs of space on and SSD. You also need to download a couple of more required components. Those are "ae.safetenser and "t5xxl_fp16.safetensor". They act like the "VAE" layer in SDXL, but have to do with how the language model interacts with the image generator. The 2 I listed are the basic once that released with the dev version of Flux, but there are others out there. From what I read, they really aren't any better than the basic.

Good luck!

Flux/chroma will work if you have a solid 16 Vram and 32 gigs of ram, but it will be twice as slow as it could be, making the whole endeavor more frustrating.