Qwen/Qwen-Image-Layered

[nesmonmehdydeal]

https://huggingface.co/Qwen/Qwen-Image-Layered

[agoulddesign]

Came here to post the same thing. EDIT: looks like quants are ~12gb for Q4. So it might be possible 🤞 Not sure even Nunchaku can quant this to be useable. Transformers folder is 40+GB...?

If not, I will be exploring API integration to delayer images directly in Krita. Maybe a custom workflow with API, not sure.

I've been waiting for this for years now!

[Acly]

It's probably not much effort for basic implementation.

But I'm not sure it's worth to integrate it any further. It's a huge model and I'm testing it on RTX PRO 6000 Blackwell - even there it's very slow. I don't even want to try to run it on consumer GPU. Yes you can use quant and block swapping to somehow get it loaded and run, but it will make it even slower.

Results are okay..? After 1-2 minute wait sometimes the result is good for some layers, but it's kinda random what ends up where, and doesn't always fill the background correctly, and sometimes doesn't do anything useful at all.

[nesmonmehdydeal]

Ok I give a try.

https://huggingface.co/Comfy-Org/Qwen-Image-Layered_ComfyUI/tree/main/split_files

[Acly]

Takes 11min for 1 image with 4 layers on 4070 12GB with Q4_K_M - the result is quite bad though compared to same image/seed with BF16 (not entirely sure it's working correctly with GGUFs)