What causes "Unsupported input dimension: {x.dim()}" in hpcaitech/Open-Sora?

Ensure input is 4D (B,C,H,W) or 5D (B,C,T,H,W) with an explicit batch dimension

hpcaitech/Open-Sora · error · ValueError

Unsupported input dimension: {x.dim()}

Error message

Unsupported input dimension: {x.dim()}

What it means

PixelUnshuffleChannelAveragingDownSampleLayer.forward only handles 4D (B,C,H,W) and 5D (B,C,T,H,W) inputs; other rank tensors raise this ValueError. The branch chosen also depends on temporal_downsample, so an incorrectly shaped tensor (e.g. an unbroadcast feature map or a 3D tensor) fails this check.

Source

Thrown at opensora/models/dc_ae/models/nn/ops.py:227

            x = x.view(B, self.out_channels, group_size, H, W)
            x = x.mean(dim=2)
        elif x.dim() == 5:  # [B, C, T, H, W]
            _, _, T, _, _ = x.shape
            if self.temporal_downsample and T != 1:  # 3d pixel unshuffle
                x = pixel_unshuffle_3d(x, self.factor)
                assert self.in_channels * self.factor**3 % self.out_channels == 0
                group_size = self.in_channels * self.factor**3 // self.out_channels
            else:  # 2d pixel unshuffle
                x = x.permute(0, 2, 1, 3, 4)  # [B, T, C, H, W]
                x = F.pixel_unshuffle(x, self.factor)
                x = x.permute(0, 2, 1, 3, 4)  # [B, C, T, H, W]
                assert self.in_channels * self.factor**2 % self.out_channels == 0
                group_size = self.in_channels * self.factor**2 // self.out_channels
            B, C, T, H, W = x.shape
            x = x.view(B, self.out_channels, group_size, T, H, W)
            x = x.mean(dim=2)
        else:
            raise ValueError(f"Unsupported input dimension: {x.dim()}")
        return x

    def __repr__(self):
        return f"PixelUnshuffleChannelAveragingDownSampleLayer(in_channels={self.in_channels}, out_channels={self.out_channels}, factor={self.factor}), temporal_downsample={self.temporal_downsample}"


class ConvPixelShuffleUpSampleLayer(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: int,
        factor: int,
    ):
        super().__init__()
        self.factor = factor
        out_ratio = factor**2
        self.conv = ConvLayer(

View on GitHub (pinned to 7ad6a96a13)

Solutions

Ensure input is 4D (B,C,H,W) or 5D (B,C,T,H,W) with an explicit batch dimension
Check upstream transforms for squeeze/view calls that drop a dimension
For single samples pass x.unsqueeze(0)

Example fix

# before
x = frame_3d  # (C,H,W)
y = layer(x)
# after
x = frame_3d.unsqueeze(0)  # (1,C,H,W)
y = layer(x)

Defensive patterns

Strategy: type-guard

Validate before calling

if x.dim() not in (4, 5):
    raise ValueError(f'expected 4D/5D input, got {tuple(x.shape)}')

Type guard

def is_valid_offload_shape(x: torch.Tensor) -> bool:
    return x.dim() in (4, 5)

Prevention

Always carry an explicit batch dimension through the pipeline
Add shape assertions at stage boundaries
Log tensor shapes in debug builds to catch rank drift early

When it happens

Trigger: Feeding a tensor with dim() != 4 and != 5 into the averaging shortcut layer, e.g. a per-frame 3D tensor or a 6D tensor from prior reshaping; occurs during encoder forward/downsample shortcut computation.

Common situations: Custom preprocessing that squeezes the batch or time dimension; feeding single images without a batch dim; intermediate reshapes that change tensor rank before the autoencoder.

Related errors

AI-assisted analysis of hpcaitech/Open-Sora@7ad6a96a13 (2026-08-28). Data as JSON: /api/errors/0090541f1f5b03c3. Report an issue: GitHub.