fix: add missing n_frames attribute in StandardDiffusionLoss

[Mr-Neutr0n]

Summary

• StandardDiffusionLoss._forward() references self.n_frames to determine the offset noise shape, but this attribute is never initialized in __init__(), causing an AttributeError when offset_noise_level > 0.0.
• Added n_frames as an optional parameter to __init__() (defaulting to None) and assigned it to self.n_frames, preserving the existing behavior where None selects the 2D (image) offset noise shape.

Details

In sgm/modules/diffusionmodules/loss.py, the _forward method contains:

if self.offset_noise_level > 0.0:
    offset_shape = (
        (input.shape[0], 1, input.shape[2])
        if self.n_frames is not None
        else (input.shape[0], input.shape[1])
    )

self.n_frames is checked but never defined anywhere in the class, so any use of offset_noise_level > 0.0 raises:

AttributeError: 'StandardDiffusionLoss' object has no attribute 'n_frames'

This fix adds n_frames: Optional[int] = None to the constructor and stores it as self.n_frames. The default of None means existing configs and image-based training are unaffected, while video-based training can pass n_frames to get the correct 3D offset noise shape.

Test plan

• Verified that existing configs (which do not pass n_frames) continue to work unchanged since the default is None
• Confirmed the offset noise code path no longer raises AttributeError