Add resolution selector

[rkfg]

One of the powerful features of Hunyuan Video is its ability to create nice videos at very low resolution. For example, Wan really struggles with it and mostly makes garbage at low res. So we should exploit it and allow user to choose a lower resolution to speed up inference tremendously.

For example, I went from 4.3s/it (640p) to 1.7s/it (320p) to 1.08 s/it (240p) on my 3090 Ti with TeaCache enabled. That's x2.5-x4 speed increase compared to the original 640p! I just added more buckets (320p and 240p) and a selector. The default is still 640p. Also the chosen bucket is printed to the console now.

[morganavr]

Why did not you add these "mirrored" resolutions for 240p bucket? You added mirrored res (where width and height are inversed) for 320p bucket.

Missing resolutions:

(336, 224),
(304, 256),
(272, 288),
(240, 320),
(208, 352),
(176, 384),

[rkfg]

They're not exactly mirrored. I used 32px steps (from the middle resolution, up and down) and these just didn't match. I tried some other sizes and there was an error like RuntimeError: The size of tensor a (2658) must match the size of tensor b (2628) at non-singleton dimension 1. I'm not sure what exactly causes it but most probably an invalid size.

[brandon929]

I added similar support on a fork I created to support MPS: https://github.com/brandon929/FramePack/blob/9a5f8291df4550831b4f0fa8c694c44f7252d1ec/diffusers_helper/bucket_tools.py#L19C1-L35C23.

My approach was to make the 640 bucket the source of truth for aspect-ratios, and to just scale things to the desired resolution. This prevents having to build tables for all the resolutions you wish to support.

[rkfg]

Yeah, this is more flexible. If you want to submit a PR with your solution I'll close mine.

[brandon929]

I merged my changes with yours, and added the following PR: #158