Comfy-Org/ComfyUI · error · ValueError

Unknown activation type: {activation_type}

Error message

Unknown activation type: {activation_type}

What it means

Factory guard in comfy/ldm/ace/vae/autoencoder_dc.py: get_activation() supports exactly relu, relu6, silu, and leaky_relu. Any other act_fn string from the model config raises ValueError at build time, before any weights load.

Source

Thrown at comfy/ldm/ace/vae/autoencoder_dc.py:49

    elif norm_type == "layer_norm":
        return ops.LayerNorm(num_features)
    elif norm_type == "rms_norm":
        return RMSNorm(num_features, eps=eps, elementwise_affine=True, bias=True)
    else:
        raise ValueError(f"Unknown normalization type: {norm_type}")


def get_activation(activation_type):
    if activation_type == "relu":
        return nn.ReLU()
    elif activation_type == "relu6":
        return nn.ReLU6()
    elif activation_type == "silu":
        return nn.SiLU()
    elif activation_type == "leaky_relu":
        return nn.LeakyReLU(0.2)
    else:
        raise ValueError(f"Unknown activation type: {activation_type}")


class ResBlock(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        norm_type: str = "batch_norm",
        act_fn: str = "relu6",
    ) -> None:
        super().__init__()

        self.norm_type = norm_type
        self.nonlinearity = get_activation(act_fn) if act_fn is not None else nn.Identity()
        self.conv1 = ops.Conv2d(in_channels, in_channels, 3, 1, 1)
        self.conv2 = ops.Conv2d(in_channels, out_channels, 3, 1, 1, bias=False)
        self.norm = get_normalization(norm_type, out_channels)

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Inspect the checkpoint config's act_fn values and correct them to one of relu|relu6|silu|leaky_relu
2. Check for casing/whitespace issues in the config string and normalize it
3. If the model truly requires another activation, extend get_activation in this fork rather than skipping the check

Example fix

# before (checkpoint config)
"act_fn": "swish"

# after
"act_fn": "silu"

SUPPORTED_ACTS = {"relu", "relu6", "silu", "leaky_relu"}
if cfg.get('act_fn') not in SUPPORTED_ACTS:
    raise SystemExit(f"unsupported act_fn {cfg.get('act_fn')}; supported: {SUPPORTED_ACTS}")

def is_supported_activation(name: str) -> bool:
    return name in {"relu", "relu6", "silu", "leaky_relu"}

When it happens

Trigger: Loading an AutoencoderDC whose config specifies an unsupported activation, e.g. 'gelu', 'Swish', 'silu ' (trailing space), or 'mish'. The act_fn is defined per-block in the checkpoint's architecture config (ResBlock default is relu6).

Common situations: Hand-edited or upstream-drifted ACE VAE configs; configs exported from a different codebase using different activation naming; typo in a custom fine-tune config.

Related errors

• activation incorrectly specified. check the config file and
• Unknown normalization type: {norm_type}
• Block with {block_type=} is not supported.
• Unknown activation function: {act_fn}
• Minimum cutoff must be larger than zero.

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/5df818aec007a5d4. Report an issue: GitHub.