Stability-AI/generative-models · error · NotImplementedError

rearranging not available for {len(in_shape)}-dimensional in

Error message

rearranging not available for {len(in_shape)}-dimensional input.

What it means

The quantizer's output was flattened to a token sequence and must be rearranged back to spatial form. Only 4D (b,h,w,c tokens) and 5D (b,t,h,w) inputs are supported; any other rank raises NotImplementedError.

Source

Thrown at sgm/modules/autoencoding/regularizers/quantize.py:483

        rearr = False
        in_shape = z.shape

        if z.ndim > 3:
            rearr = self.output_dim is not None
            z = rearrange(z, "b c ... -> b (...) c")
        z = self.proj_in(z)
        z_q, loss_dict = super().forward(z)

        z_q = self.proj_out(z_q)
        if rearr:
            if len(in_shape) == 4:
                z_q = rearrange(z_q, "b (h w) c -> b c h w ", w=in_shape[-1])
            elif len(in_shape) == 5:
                z_q = rearrange(
                    z_q, "b (t h w) c -> b c t h w ", w=in_shape[-1], h=in_shape[-2]
                )
            else:
                raise NotImplementedError(
                    f"rearranging not available for {len(in_shape)}-dimensional input."
                )

        return z_q, loss_dict

View on GitHub (pinned to e8cd657656)

Solutions

1. Reshape the latent so it has 4D (image) or 5D (video) form before calling forward
2. Check the model config: use a regularizer variant matching your data dimensionality
3. Extend the else-branch with a rearrange pattern for your specific rank

Example fix

// before: passing a 3D latent (b c l) into the video quantizer
z_q, loss = quantizer(z, image_only_indicator)
// after: unsqueeze to 5D video layout (b c t h w)
z = z.unsqueeze(2).unsqueeze(3)  # give l a t/h/w interpretation as appropriate

def validate_quantizer_input(z):
    if z.dim() not in (4, 5):
        raise ValueError(f"quantizer expects 4D or 5D latent, got {z.dim()}D")
validate_quantizer_input(z)

def is_spatial_latent(z) -> bool:
    return z.dim() in (4, 5)

try:
    z_q, loss = quantizer(z)
except NotImplementedError as e:
    raise RuntimeError(f"latent rank {z.dim()} unsupported: reshape to 4D/5D") from e

When it happens

Trigger: Calling the quantizer's forward with z whose shape has neither 4 nor 5 dimensions, e.g. a 2D (b,c) or 3D (b,c,l) latent tensor with an in_shape other than len 4 or 5.

Common situations: Feeding an audio/1D-signal autoencoder into a video/image VQ model, custom encoder producing extra/missing dims, or passing unbatched tensors.

Related errors

• NotImplementedError
• Unknown loss type {self.loss_type}
• provide num_res_blocks either as an int (globally constant)
• NotImplementedError
• unknown merge strategy {self.merge_strategy}

AI-assisted analysis of Stability-AI/generative-models@e8cd657656 (2026-08-29). Data as JSON: /api/errors/d3c690b955de8fda. Report an issue: GitHub.