Comfy-Org/ComfyUI · error · ValueError

SeedVR2 VAE cache input is too short for convolution: input_

Error message

SeedVR2 VAE cache input is too short for convolution: input_len={input_len}, pad_len={pad_len}.

What it means

get_cache_size computes how many input frames a cached (streaming) convolution must retain between slices. If the effective input length (chunk + cached frames + padding) is smaller than the dilated kernel size, output_len <= 0 — the conv would produce zero output frames — and the helper raises. This only happens with tiled/sliced VAE execution where the slice length is configured below the receptive field of the conv.

Source

Thrown at comfy/ldm/seedvr/vae.py:266

        model.padding = orig_padding

class MemoryState(Enum):
    DISABLED = 0
    INITIALIZING = 1
    ACTIVE = 2
    UNSET = 3

def get_cache_size(conv_module, input_len, pad_len, dim=0):
    dilated_kernel_size = conv_module.dilation[dim] * (conv_module.kernel_size[dim] - 1) + 1
    output_len = (input_len + pad_len - dilated_kernel_size) // conv_module.stride[dim] + 1
    remain_len = (
        input_len + pad_len - ((output_len - 1) * conv_module.stride[dim] + dilated_kernel_size)
    )
    overlap_len = dilated_kernel_size - conv_module.stride[dim]
    cache_len = overlap_len + remain_len

    if output_len <= 0:
        raise ValueError(
            f"SeedVR2 VAE cache input is too short for convolution: input_len={input_len}, pad_len={pad_len}."
        )
    return cache_len

class DiagonalGaussianDistribution(object):
    def __init__(self, parameters: torch.Tensor):
        self.parameters = parameters
        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
        self.logvar = torch.clamp(self.logvar, BYTEDANCE_LOGVAR_CLAMP_MIN, BYTEDANCE_LOGVAR_CLAMP_MAX)

    def mode(self):
        return self.mean

class SpatialNorm(nn.Module):
    def __init__(
        self,
        f_channels: int,
        zq_channels: int,

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Increase the tile/split size (or the VAE tile settings) so each slice is at least the dilated kernel length in the split dimension.
2. Raise or disable the memory limit for the VAE pass so slicing does not shrink below kernel size.
3. For very short videos, avoid temporal slicing in the temporal dimension altogether.
4. Check conv kernel/dilation in the message-adjacent config to compute the minimum valid slice length: dilation*(kernel_size-1)+1.

Example fix

# before
vae.decode(latent, tile=(1, 256, 256))  # temporal tile 1 < kernel 3
# after
vae.decode(latent, tile=(3, 256, 256))  # >= dilated kernel size

def min_slice_len(conv, dim=0):
    return conv.dilation[dim] * (conv.kernel_size[dim] - 1) + 1

def validate_tile(conv, tile_len, dim=0):
    need = min_slice_len(conv, dim)
    if tile_len < need:
        raise ValueError(f"tile length {tile_len} below minimum {need} for this conv")
    return tile_len

When it happens

Trigger: Running SeedVR2 VAE encode/decode with a tile/split size (or extremely short video) whose per-slice temporal length is smaller than dilation*(kernel-1)+1; aggressive memory-limit settings that force tiny temporal slices.

Common situations: Low-VRAM users setting very small tiling lengths; processing 1-frame videos with 3D convs whose temporal kernel is 3+; memory_limit values that shrink slices below kernel size.

Related errors

• SeedVR2 VAE cache size {next_cache_size} exceeds split size
• SeedVR2 VAE cache size {cache_size} exceeds input length {in
• SeedVR2 VAE convolution requires an explicit MemoryState.
• SeedVR2 upsample expected {self.channels} channels, got {hid
• SeedVR2 downsample expected {self.channels} channels, got {h

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/52343055cb29aea4. Report an issue: GitHub.