Comfy-Org/ComfyUI · error · ValueError

Unknown SeedVR2 VAE forward mode: {mode}

Error message

Unknown SeedVR2 VAE forward mode: {mode}

What it means

SeedVR2 VAE forward() accepts only the literal modes "encode", "decode", and "all" (encode then decode). Any other string falls through the if-chain and raises a ValueError. This is a narrow dispatcher, so an unrecognized mode is always a caller bug, never a runtime data condition.

Source

Thrown at comfy/ldm/seedvr/vae.py:1440

                    self._decode(z_slices[z_idx], memory_state=MemoryState.ACTIVE, memory_cache=memory_cache)
                )
            out = torch.cat(decoded_slices, dim=2)
            return out
        else:
            return self._decode(z)

    def forward(self, x: torch.FloatTensor, mode: Literal["encode", "decode", "all"] = "all"):
        def _unwrap(value):
            return value[0] if isinstance(value, tuple) else value

        if mode == "encode":
            return _unwrap(self.encode(x))
        if mode == "decode":
            return _unwrap(self.decode_(x))
        if mode == "all":
            latent = _unwrap(self.encode(x))
            return _unwrap(self.decode_(latent))
        raise ValueError(f"Unknown SeedVR2 VAE forward mode: {mode}")

class VideoAutoencoderKLWrapper(VideoAutoencoderKL):
    def __init__(
        self,
        spatial_downsample_factor = 8,
        temporal_downsample_factor = 4,
    ):
        self.spatial_downsample_factor = spatial_downsample_factor
        self.temporal_downsample_factor = temporal_downsample_factor
        super().__init__()
        self.set_memory_limit(BYTEDANCE_VAE_CONV_MEM_GIB, BYTEDANCE_VAE_NORM_MEM_GIB)

    def forward(self, x: torch.FloatTensor):
        z, p = self._encode_with_raw_latent(x)
        x = self.decode(z)
        return x, z, p

    def _encode_with_raw_latent(self, x):

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Pass one of the three supported literals: "encode", "decode", or "all".
2. If the mode comes from user input, validate it against {"encode", "decode", "all"} before calling forward.
3. Prefer calling encode()/decode_() directly instead of the mode dispatcher when only one operation is needed.

Example fix

# before
out = vae(x, mode="enc")
# after
out = vae(x, mode="encode")

VALID_MODES = {"encode", "decode", "all"}
if mode not in VALID_MODES:
    raise ValueError(f"mode must be one of {VALID_MODES}, got {mode!r}")
out = vae.forward(x, mode=mode)

def is_seedvr_forward_mode(mode) -> bool:
    return mode in ("encode", "decode", "all")

When it happens

Trigger: Calling vae.forward(x, mode="enc"), mode="auto", mode=None, or passing a mode variable that was never validated.

Common situations: Typos in workflow scripts; passing a mode read from a user config file or JSON without validation; refactoring code that used a different VAE class with more modes.

Related errors

• SeedVR2 decoder only supports UpDecoderBlock3D, got {up_bloc
• SeedVR2 VideoAutoencoderKLWrapper.decode: `seedvr2_tiling` m
• SeedVR2 VideoAutoencoderKLWrapper.decode: 5-D latent input m
• SeedVR2 VideoAutoencoderKLWrapper.decode: 4-D latent input m
• SeedVR2 VideoAutoencoderKLWrapper.decode: latent input must

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/b4db3cdb23bfb2a1. Report an issue: GitHub.