Comfy-Org/ComfyUI · critical · ValueError

Unknown patch method: {self.patch_method}

Error message

Unknown patch method: {self.patch_method}

What it means

Patcher (forward patchify) in the Cosmos tokenizer supports two methods: 'haar' (learned-free Haar wavelet via _haar) and 'rearrange' (space-to-depth via _arrange). The method is fixed from the patch_method config string at construction; any other value raises 'Unknown patch method' on the first forward call. The valid names are keys of the _WAVELETS table ( haar wavelet) plus the plain rearrange path.

Source

Thrown at comfy/ldm/cosmos/cosmos_tokenizer/patching.py:71

        self.register_buffer(
            "wavelets", _WAVELETS[patch_method], persistent=_PERSISTENT
        )
        self.range = range(int(torch.log2(torch.tensor(self.patch_size)).item()))
        self.register_buffer(
            "_arange",
            torch.arange(_WAVELETS[patch_method].shape[0]),
            persistent=_PERSISTENT,
        )
        for param in self.parameters():
            param.requires_grad = False

    def forward(self, x):
        if self.patch_method == "haar":
            return self._haar(x)
        elif self.patch_method == "rearrange":
            return self._arrange(x)
        else:
            raise ValueError("Unknown patch method: " + self.patch_method)

    def _dwt(self, x, mode="reflect", rescale=False):
        dtype = x.dtype
        h = self.wavelets.to(device=x.device)

        n = h.shape[0]
        g = x.shape[1]
        hl = h.flip(0).reshape(1, 1, -1).repeat(g, 1, 1)
        hh = (h * ((-1) ** self._arange.to(device=x.device))).reshape(1, 1, -1).repeat(g, 1, 1)
        hh = hh.to(dtype=dtype)
        hl = hl.to(dtype=dtype)

        x = F.pad(x, pad=(n - 2, n - 1, n - 2, n - 1), mode=mode).to(dtype)
        xl = F.conv2d(x, hl.unsqueeze(2), groups=g, stride=(1, 2))
        xh = F.conv2d(x, hh.unsqueeze(2), groups=g, stride=(1, 2))
        xll = F.conv2d(xl, hl.unsqueeze(3), groups=g, stride=(2, 1))
        xlh = F.conv2d(xl, hh.unsqueeze(3), groups=g, stride=(2, 1))
        xhl = F.conv2d(xh, hl.unsqueeze(3), groups=g, stride=(2, 1))

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Set patch_method to 'haar' or 'rearrange' exactly
2. Verify the tokenizer checkpoint config; if it names another wavelet, use a ComfyUI version that maps it
3. For custom wavelets, add the wavelet to _WAVELETS and a matching forward branch in a fork

Example fix

# before
Patcher(patch_method="db2", patch_size=patch_size)

# after
Patcher(patch_method="haar", patch_size=patch_size)

if patch_method not in {"haar", "rearrange"}:
    raise ValueError(f"patch_method {patch_method!r} unsupported; use 'haar' or 'rearrange'")

def is_supported_patch_method(m: str) -> bool:
    return m in {"haar", "rearrange"}

When it happens

Trigger: Constructing cosmos_tokenizer Patcher with patch_method like 'db2', 'haar-learned', or 'space2depth' - e.g. loading a Cosmos tokenizer checkpoint variant that stores a different wavelet name in its config.

Common situations: Loading Cosmos-Predict2 or custom VAE/tokenizer checkpoints with wavelet configs beyond haar; porting configs from NVIDIA's cosmos repo where more wavelets exist in _WAVELETS; typos in custom configs.

Related errors

• Normalization {name} not found
• Normalization mode {self.qkv_norm_mode} not found, only supp
• Unknown pos_emb_cls {self.pos_emb_cls}
• Unknown interpolation method {self.interpolation}
• Unknown pos_emb_cls {self.pos_emb_cls}

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/8dd4020540a1fee9. Report an issue: GitHub.