deepfakes/faceswap · error · FaceswapError

The TFLambdaOp '{name}' is not supported

Error message

The TFLambdaOp '{name}' is not supported

What it means

FaceswapError raised while converting a legacy Keras 2 model config to Keras 3: a TFLambdaOp layer is encountered whose operation (the last '.'-separated segment of its name) is not one of multiply, truediv, add, subtract. Those four are the only lambda ops the converter can map to Keras 3 ScalarOp layers. Anything else (e.g. custom tf functions) has no automatic equivalent.

Source

Thrown at plugins/train/model/_base/update.py:153

    def _convert_lambda_config(self, layer: dict[str, T.Any]):
        """Keras 2 TFLambdaOps are not compatible with Keras 3. Scalar operations can be
        relatively easily substituted with a :class:`~lib.model.layers.ScalarOp` layer

        Parameters
        ----------
        layer
            An existing Keras 2 TFLambdaOp layer

        Raises
        ------
        FaceswapError
            If the TFLambdaOp is not currently supported
        """
        name = layer["config"]["name"]
        operation = name.rsplit(".", maxsplit=1)[-1]
        if operation not in ("multiply", "truediv", "add", "subtract"):
            raise FaceswapError(f"The TFLambdaOp '{name}' is not supported")
        value = layer["inbound_nodes"][0][-1]["y"]

        if isinstance(layer["config"]["dtype"], str):
            dtype = layer["config"]["dtype"]
        else:
            dtype = layer["config"]["dtype"]["config"]["name"]
        new_layer = ScalarOp(operation, value, name=name, dtype=dtype)

        logger.debug("Converting legacy TFLambdaOp: %s", layer)

        layer["class_name"] = "ScalarOp"
        layer["config"] = new_layer.get_config()
        for n in layer["inbound_nodes"]:
            n[-1] = {}
        layer["inbound_nodes"] = [layer["inbound_nodes"]]
        logger.debug("Converted legacy TFLambdaOp to %s", layer)

    def _process_deprecations(self, layer: dict[str, T.Any]) -> None:  # noqa[C901]

View on GitHub (pinned to f530cb7508)

Solutions

1. Check the layer name reported in the error to identify the unsupported operation
2. Re-save/rebuild the model in Faceswap 2 using only the supported arithmetic lambda ops, then port
3. If the op is one you added yourself, implement an equivalent ScalarOp mapping upstream in update.py before porting

Example fix

# before: legacy config layer named
# 'model/tf.math.reduce_mean_3' -> raises (reduce_mean unsupported)

# after: rebuild in FS2 using supported op, e.g.
# 'model/tf.math.multiply_1' -> converts to ScalarOp

import json
cfg = json.loads(config_str)
for layer in cfg["config"]["layers"]:
    if layer["class_name"] == "TFLambdaOp":
        op = layer["config"]["name"].rsplit(".", 1)[-1]
        if op not in ("multiply", "truediv", "add", "subtract"):
            raise SystemExit(f"Unsupported lambda op {op!r}; rebuild model with supported ops")

from lib.exceptions import FaceswapError
try:
    updater.port_model()
except FaceswapError as err:
    if "TFLambdaOp" in str(err):
        log_unportable_model(old_file)  # record and skip, do not crash the batch
    else:
        raise

When it happens

Trigger: Porting a Faceswap 2 model that contains an exotic TFLambdaOp layer, e.g. from a fork or plugin that added custom lambda operations to the graph. operation = name.rsplit('.', 1)[-1] fails the whitelist check.

Common situations: Porting models produced by modified Faceswap forks, very old versions, or models whose config was hand-edited to insert custom lambdas.

Related errors

• '{self._old_model_file}' is not a valid Faceswap 2 model fil
• Unable to load the model from '{self.filename}'. This may be
• Error loading weights file {self._weights_file}.
• The state file '{state_file}' does not exist. This model can
• Clip network could not be found in '{state_file}'. Discovere

AI-assisted analysis of deepfakes/faceswap@f530cb7508 (2026-08-15). Data as JSON: /api/errors/f87fa197408dc849. Report an issue: GitHub.