3-Stage WF to get 1.5K -Krea 1- quality
Tested on RTX 4060Ti 16Gb
1 stage takes about 48 sec | Utilizes about 12Gb RAM and 13Gb VRAM
2 stage - 87 sec | Utilizes about 13.5Gb RAM and 13.5Gb VRAM
3 stage - 4 sec | Utilizes about 22Gb RAM and 10Gb VRAM
If you see that the final result is oversharped - just increase the value of blur_strength in the LayerFilter: Film V2 node (Stage 3)
V2:
No more errors with 42lux repository ("Model Sampling Flux Normalized" replaced with the stock node ModelSamplingFlux)
Added two additional postprocessing stages
Description
Initial version
FAQ
Comments (6)
Welcome to try Flux-Krea-nunchaku version,
20sec 1 photo , speed fast
https://civitai.com/models/1831687/flux1-krea-dev-nunchaku-my-20sec-workflow
Enjoying how simple the WF is so far. I'm just curious because I'm still quite ignorant. Do you have a moment to explain what the PAG is/doing. As well as the purpose of the Sampling Shift (max_shift/base_shift)? Thank you!
Hi! You're welcome! You can read about PAG here https://arxiv.org/abs/2403.17377 or here https://cvlab-kaist.github.io/Perturbed-Attention-Guidance/
And this is very good explanation about the "shifts" https://www.reddit.com/r/comfyui/comments/1eugsf7/comment/likjgk7/
Doesn't work. ModelSamplingFluxNormalized node is red. https://github.com/42lux/ComfyUI-42lux is unavailable. Error 404
🤔 Ok, thanks for tracking, I will fix it
Fixed with v2