invoke-ai/InvokeAI · error · ValueError

per_layer_weights must be comma-separated numbers: {e}

Error message

per_layer_weights must be comma-separated numbers: {e}

What it means

Krea2ConditioningRebalanceInvocation._parse_weights splits the per_layer_weights string on commas and converts each token with float(); if any token is not a valid decimal number (e.g. '1.2.3' or 'abc'), float() raises ValueError, which is re-raised with this message. The field is expected to be comma-separated numeric gains for the 12 tapped Krea2 text-encoder layers.

Source

Thrown at invokeai/app/invocations/krea2_conditioning_rebalance.py:53

    conditioning: Krea2ConditioningField = InputField(
        description=FieldDescriptions.cond, input=Input.Connection, title="Conditioning"
    )
    per_layer_weights: str = InputField(
        default="1.0,1.0,1.0,1.0,1.0,1.0,1.0,2.5,5.0,1.1,4.0,1.0",
        description=f"Comma-separated gains for the {_NUM_TEXT_LAYERS} tapped encoder layers (exactly "
        f"{_NUM_TEXT_LAYERS} values).",
    )
    multiplier: float = InputField(
        default=4.0,
        allow_inf_nan=False,
        description="Overall multiplier applied to the conditioning after per-layer weighting.",
    )

    def _parse_weights(self) -> list[float]:
        try:
            weights = [float(x.strip()) for x in self.per_layer_weights.split(",") if x.strip() != ""]
        except ValueError as e:
            raise ValueError(f"per_layer_weights must be comma-separated numbers: {e}") from e
        if len(weights) != _NUM_TEXT_LAYERS:
            raise ValueError(f"per_layer_weights must have exactly {_NUM_TEXT_LAYERS} values, got {len(weights)}.")
        if not all(math.isfinite(weight) for weight in weights):
            raise ValueError("per_layer_weights must contain only finite values.")
        return weights

    @torch.no_grad()
    def invoke(self, context: InvocationContext) -> Krea2ConditioningOutput:
        weights = self._parse_weights()

        cond_data = context.conditioning.load(self.conditioning.conditioning_name)
        assert len(cond_data.conditionings) == 1
        conditioning = cond_data.conditionings[0]
        assert isinstance(conditioning, Krea2ConditioningInfo)

        embeds = conditioning.prompt_embeds  # (B, seq, 12, hidden)
        gains = torch.tensor(weights, dtype=embeds.dtype, device=embeds.device).view(1, 1, _NUM_TEXT_LAYERS, 1)
        embeds = embeds * gains * self.multiplier

View on GitHub (pinned to 0b6a024f2f)

Solutions

1. Fix the string so every comma-separated token parses with Python float(), e.g. '1.0,0.8,...' (12 values).
2. Use dots, not commas, as decimal separators (the comma is the list separator).
3. Validate the string programmatically before building the graph: try [float(x) for x in s.split(',') if x.strip()].

Example fix

// before
per_layer_weights="1.0, 0,8, 1.2, ..."  // '0,8' -> float() ValueError
// after
per_layer_weights="1.0, 0.8, 1.2, ..."

def valid_weights(s: str, n: int = 12) -> bool:
    try:
        vals = [float(x) for x in s.split(",") if x.strip() != ""]
    except ValueError:
        return False
    return len(vals) == n

try:
    out = invoke(node)
except ValueError as e:
    if "comma-separated numbers" in str(e):
        node.per_layer_weights = ",".join(["1.0"] * 12)  # neutral default
        out = invoke(node)

When it happens

Trigger: Invoking a krea2_conditioning_rebalance node with per_layer_weights containing a token that float() cannot parse, such as '0.5,,x' with a non-numeric entry, or with a decimal comma ('0,5') instead of a dot.

Common situations: Locale confusion (comma as decimal separator) breaking the comma-separated list; typos when hand-typing 12 values; whitespace/odd characters pasted from a spreadsheet.

Related errors

• per_layer_weights must have exactly {_NUM_TEXT_LAYERS} value
• per_layer_weights must contain only finite values.
• cfg_scale values must be finite.
• shift must be finite.
• At least one Krea-2 conditioning is required.

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/6d7f858b6aba4123. Report an issue: GitHub.