lllyasviel/Fooocus · error · ValueError

Wrong activation value in EqualLinear: {activation}Supported

Error message

Wrong activation value in EqualLinear: {activation}Supported ones are: ['fused_lrelu', None].

What it means

Identical guard to the one in stylegan2_arch.py, but in the bilinear-backbone variant ldm_patched/pfn/architecture/face/stylegan2_bilinear_arch.py. This EqualLinear only supports activation=None or 'fused_lrelu' because the bilinear variant replaces fused up/down sampling and thus has no other fused activation paths. Any other activation string raises ValueError at construction.

Source

Thrown at ldm_patched/pfn/architecture/face/stylegan2_bilinear_arch.py:52

            Supported: 'fused_lrelu', None. Default: None.
    """

    def __init__(
        self,
        in_channels,
        out_channels,
        bias=True,
        bias_init_val=0,
        lr_mul=1,
        activation=None,
    ):
        super(EqualLinear, self).__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.lr_mul = lr_mul
        self.activation = activation
        if self.activation not in ["fused_lrelu", None]:
            raise ValueError(
                f"Wrong activation value in EqualLinear: {activation}"
                "Supported ones are: ['fused_lrelu', None]."
            )
        self.scale = (1 / math.sqrt(in_channels)) * lr_mul

        self.weight = nn.Parameter(torch.randn(out_channels, in_channels).div_(lr_mul))
        if bias:
            self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
        else:
            self.register_parameter("bias", None)

    def forward(self, x):
        if self.bias is None:
            bias = None
        else:
            bias = self.bias * self.lr_mul
        if self.activation == "fused_lrelu":
            out = F.linear(x, self.weight * self.scale)

View on GitHub (pinned to ae05379cc9)

Solutions

Pass activation=None or 'fused_lrelu'.
If you patched stylegan2_arch.py to support a new activation, apply the matching patch to stylegan2_bilinear_arch.py (and vice versa) so the two vendored copies stay consistent.
Apply custom activations outside EqualLinear in the caller's forward().

Example fix

// before
EqualLinear(num_style_feat, in_channels, activation='relu')
// after
EqualLinear(num_style_feat, in_channels, activation=None)

Defensive patterns

Strategy: validation

Validate before calling

valid = {None, 'fused_lrelu'}
assert activation in valid, f'activation must be one of {valid}, got {activation!r}'

Type guard

def is_equal_linear_activation(v) -> bool:
    return v is None or (isinstance(v, str) and v == 'fused_lrelu')

Prevention

Treat stylegan2_arch.py and stylegan2_bilinear_arch.py as a pair: any change to accepted values must go into both.
Diff the two vendored files before modifying either.

When it happens

Trigger: Building the bilinear StyleGAN2 face model with EqualLinear(..., activation=<anything except 'fused_lrelu' or None>), typically when copying layer code from the non-bilinear file that was modified to use another activation.

Common situations: Diverging the two stylegan2_arch / stylegan2_bilinear_arch copies and applying an activation change to only one; loading a face-model config tuned against a different implementation.

Related errors

AI-assisted analysis of lllyasviel/Fooocus@ae05379cc9 (2026-08-15). Data as JSON: /api/errors/15f4a9eac869e461. Report an issue: GitHub.