Add resolution slider (from 240 -> 720 in 16-step increments)

[brandon929]

This incorporates the changes from @rkfg in #69 and adds the ability to support "any" resolution instead of a few predefined buckets.

This uses the buckets defined in the original implementation for "640" resolution as the source-of-truth for calculating aspect ratios for other resolution buckets.

[yar85]

I found that the calculated resolution doesn't match the input resolution from UI, so I used a little hack:

if resolution != 640:
    scale = lambda x, factor: round(x * factor / 16 ) * 16
    # find input resolution which gives output resolution closest to desired
    is_portrait_orientation = w < h
    res = resolution - 16
    bucket_size = best_bucket[1] if is_portrait_orientation else best_bucket[0]
    scaled_size = 0
    while scaled_size < resolution:
        res += 16
        scale_factor = res / 640.0
        scaled_size = scale(bucket_size, scale_factor)
    # no need to recalc scale_factor with decrementing res
    if is_portrait_orientation:
        best_bucket = (scale(best_bucket[0], scale_factor), scaled_size)
    else:
        best_bucket = (scaled_size, scale(best_bucket[1], scale_factor))

This increments input value by 16 until the resulting resolution height matches the input. I'm weak at math (and bad at Python programming 😸), so I'd appreciate some advice on how to avoid this awful brute-force loop.

upd.: New hack code, added portrait orientation support.

[brandon929]

FramePack currently generates videos that, if you have a square image, is roughly 640x640. If you have a different aspect ratio, it alters the actual image resolution to approximate that aspect ratio without really changing the pixel count.

This PR maintains that same logic, but lets you adjust the pixel count with the resolution mapping to approximately what the X and Y dimension would be if you have a square image.

I think this is a pretty good user experience as you can drop in nearly any photograph and get consistent performance no matter the aspect ratio.

[yar85]

This PR maintains that same logic, but lets you adjust the pixel count with the resolution mapping to approximately what the X and Y dimension would be if you have a square image.

Oh, I see now. Thank you for the clarification!
Still, since my user experience is more focused on the outcome of the generation (output video) rather than the generation process itself, I’ll leave a comment with the code for those who share the same perspective.
It might be useful to have a dedicated UI setting for toggling the logic behind output dimensions - but of course, that idea beyond the scope of this demo repository and your PR, which preserves the original logic. And I'm definitely not pushing the idea, but it should be mentioned here as relevant, I guess :)