Fix swapped height/width dimensions in I2VDenoiser

[Mr-Neutr0n]

Summary

• Fix dimension unpacking order in I2VDenoiser.denoise() where masked_ref.size() was incorrectly unpacked as (b, c, t, w, h) instead of (b, c, t, h, w)
• The standard PyTorch video tensor layout is (B, C, T, H, W), matching how masked_ref is constructed in prepare_inference_condition() (B, C, T, H, W = z.shape)
• The swapped variables cause the image_gs guidance scale tensor (built via .repeat(b, c, 1, h, w)) to have incorrect spatial dimensions when scale_temporal_osci is enabled, leading to wrong guidance scaling or shape errors for non-square video resolutions

Details

In opensora/utils/sampling.py, line 186:

# Before (incorrect):
b, c, t, w, h = masked_ref.size()

# After (correct):
b, c, t, h, w = masked_ref.size()

The h and w variables are used downstream in the temporal oscillation scaling branch:

image_gs = torch.linspace(1.0, step_upper_image_gs, t)[
    None, None, :, None, None
].repeat(b, c, 1, h, w)

With the old code, for a non-square video (e.g., 720x480 latent), h would hold the width value and w the height value, producing an image_gs tensor with swapped spatial dimensions that cannot correctly broadcast against cond, uncond, and uncond_2 in the guidance computation.

Test plan

• Verified that prepare_inference_condition() in inference.py creates tensors with (B, C, T, H, W) layout
• Confirmed the fix aligns the unpacking order with all downstream usage of h and w
• Generate non-square videos with scale_temporal_osci=True to verify correct guidance scaling