192 Frame Limit (Bug and Proposed Fix)

[pftq]

I'm coming from debugging the SkyReels code and it seems to point at a bug in the Hunyuan VAE that hard limits the video at 192 frames (even if there is good output going beyond that, it just gets thrown out).

In autoencoder_kl_hunyuan_video.py. The original code runs into an out-of-bounds at the 25th tile (193rd frame at 8 frames per tile), which then causes the video to loop from 193 onwards instead of using the correct frames.

Find:

for i in range(0, num_frames, tile_latent_stride_num_frames):
    tile = z[:, :, i : i + tile_latent_min_num_frames + 1, :, :]

Replace:

for i in range(0, num_frames, tile_latent_stride_num_frames):
    end_idx = min(i + tile_latent_min_num_frames + 1, num_frames)
    tile = z[:, :, i : end_idx, :, :]
    if i > 0:
        tile = torch.cat([z[:, :, i - 1 : i, :, :], tile], dim=2)   # 20250222 pftq: Add previous frame for continuity

Making this change should remove the hard limit but I'm not quite sure how to test and apply it.