[Feature Request] Implement MrFlow into image generation models in wan2GP

[omeganebula666-creator]

MrFlow can speed up most text2image models by upto 10 to 20 times.
https://github.com/Xingyu-Zheng/MrFlow

[Tophness]

Been testing an early draft of this. So far, not good.
The results are way off from the original, and kind of what I'd expect from generating at low res, upscaling then essentially inpainting the same image. I mean, generally if you generate a lower native res, it's not the same image at all anymore, even for the same seed.
It seems like I reproduced all the steps in the paper, but I'm probably missing something
here's the wgp.py and /models/_settings.json if anyone wants to test.

Probably more importantly, can we confirm this actually works in comfyui?
It looks impressive if true, but it could also be one of those papers that promises a lot and doesn't deliver.

[Basthet]

To make it short. Of course it works with ComfyUI as the method per se is nothing new. Just make a low res picture, scale it up and make a new sample path with low denoise + with as view steps as possible. What's new, that it was proven, to be indeed faster by achieving almost the same quality (but finding the correct, model specific settings - denoise and step value). The tricky part was to find the correct settings for employed models (That's all what the comfy-UI implementation does: offer the correct model specific presets).

For normal that method was used to improve details and fidelity. Create a picture - but at target resolution, noise injection, and a second sampling with low denoise.
Good old SDXL times... Some are using it till today. There's even a quite unfair discussion on the MrFlow github page about this: Xingyu-Zheng/MrFlow#3.

Btw: the author was open with that from the beginning.
If you ask me: The author is right. He's focusing on speed, not on the other details/fidelity aspect + he objectively shows what to expect. I resembled his approach with the old method and got about 30% speed improvement with Krea 2 + Flux Klein 9b in ComfyUI. Only 30% as I am only on 8GB VRAM.

What you have done wrong: sorry I'm not a programmer. Did you use low denoise on the second sampling? It should be small - I used 0.16. The higher it is, the more the final picture will change.

[Tophness]

To make it short. Of course it works with ComfyUI as the method per se is nothing new. Just make a low res picture, scale it up and make a new sample path with low denoise + with as view steps as possible. What's new, that it was proven, to be indeed faster by achieving almost the same quality (but finding the correct, model specific settings - denoise and step value). The tricky part was to find the correct settings for employed models (That's all what the comfy-UI implementation does: offer the correct model specific presets).

For normal that method was used to improve details and fidelity. Create a picture - but at target resolution, noise injection, and a second sampling with low denoise. Good old SDXL times... Some are using it till today. There's even a quite unfair discussion on the MrFlow github page about this: Xingyu-Zheng/MrFlow#3.

Btw: the author was open with that from the beginning. If you ask me: The author is right. He's focusing on speed, not on the other details/fidelity aspect + he objectively shows what to expect. I resembled his approach with the old method and got about 30% speed improvement with Krea 2 + Flux Klein 9b in ComfyUI. Only 30% as I am only on 8GB VRAM.

That explains a lot, thanks.
I was expecting too much from this due to the video I saw on it that was claiming it was not only better quality than alternatives like teacache but basically imperceptible from the original.

What you have done wrong: sorry I'm not a programmer. Did you use low denoise on the second sampling? It should be small - I used 0.16. The higher it is, the more the final picture will change.

I did include all the parameters as adjustable in the UI, but I just tested it on the defaults and didn't get good results.
In that case, I'll leave it up to anyone else who wants to try.
It would be trivial to use all the correct values for each model and include them as presets.