invoke-ai/InvokeAI · error · ValueError

Krea-2 conditioning mask shape {tuple(mask.shape)} does not

Error message

Krea-2 conditioning mask shape {tuple(mask.shape)} does not match prompt embedding shape {tuple(embeds.shape[:2])}.

What it means

Each Krea-2 conditioning carries prompt embeddings and a boolean attention mask. The mask must have shape (batch, seq_len), i.e. exactly the first two dims of the prompt embedding tensor. This ValueError fires when the loaded mask shape is inconsistent with its embeddings, indicating corrupted or mismatched conditioning data.

Source

Thrown at invokeai/app/invocations/krea2_denoise.py:164

    ) -> Krea2RegionalPromptingExtension:
        conditioning_fields = (
            [conditioning_field] if isinstance(conditioning_field, Krea2ConditioningField) else conditioning_field
        )
        if not conditioning_fields:
            raise ValueError("At least one Krea-2 conditioning is required.")

        text_conditionings: list[Krea2TextConditioning] = []
        for field in conditioning_fields:
            cond_data = context.conditioning.load(field.conditioning_name)
            assert len(cond_data.conditionings) == 1
            conditioning = cond_data.conditionings[0]
            assert isinstance(conditioning, Krea2ConditioningInfo)
            conditioning = conditioning.to(dtype=dtype, device=device)
            embeds = conditioning.prompt_embeds
            if conditioning.prompt_embeds_mask is not None:
                mask = conditioning.prompt_embeds_mask.to(device=device, dtype=torch.bool)
                if mask.shape != embeds.shape[:2]:
                    raise ValueError(
                        f"Krea-2 conditioning mask shape {tuple(mask.shape)} does not match "
                        f"prompt embedding shape {tuple(embeds.shape[:2])}."
                    )
                valid_token_counts = mask.sum(dim=1)
                if not torch.equal(valid_token_counts, valid_token_counts[:1].expand_as(valid_token_counts)):
                    raise ValueError("All Krea-2 conditioning batch items must have the same valid token count.")
                embeds = torch.stack(
                    [batch_embeds[batch_mask] for batch_embeds, batch_mask in zip(embeds, mask, strict=True)]
                )
            regional_mask = None
            if field.mask is not None:
                mask = context.tensors.load(field.mask.tensor_name)
                regional_mask = Krea2RegionalPromptingExtension.preprocess_regional_prompt_mask(
                    mask=mask,
                    grid_height=grid_height,
                    grid_width=grid_width,
                    dtype=dtype,
                    device=device,

View on GitHub (pinned to 0b6a024f2f)

Solutions

1. Regenerate the conditioning with the current Krea-2 text encoder instead of reusing cached/saved conditioning data.
2. If constructing Krea2ConditioningInfo manually, slice or pad prompt_embeds_mask so its shape equals prompt_embeds.shape[:2].
3. Verify the text-encoder/model version that produced the conditioning matches the one used at denoise time.

Example fix

// before: mask with wrong seq_len
Krea2ConditioningInfo(prompt_embeds=embeds, prompt_embeds_mask=mask[:, :embeds.shape[1]-1])
// after: mask matches embeds' (batch, seq) dims
assert mask.shape == embeds.shape[:2]
Krea2ConditioningInfo(prompt_embeds=embeds, prompt_embeds_mask=mask)

if cond.prompt_embeds_mask is not None:
    assert cond.prompt_embeds_mask.shape == cond.prompt_embeds.shape[:2], \
        f"mask {tuple(cond.prompt_embeds_mask.shape)} != embeds {tuple(cond.prompt_embeds.shape[:2])}"

def has_consistent_mask(cond: Krea2ConditioningInfo) -> bool:
    if cond.prompt_embeds_mask is None:
        return True
    return tuple(cond.prompt_embeds_mask.shape) == tuple(cond.prompt_embeds.shape[:2])

try:
    out = invoke_krea2_denoise(...)
except ValueError as e:
    if "mask shape" in str(e) and "prompt embedding shape" in str(e):
        cond = regenerate_conditioning_with_current_text_encoder()
        out = invoke_krea2_denoise(...)
    else:
        raise

When it happens

Trigger: Loading a Krea2ConditioningInfo via context.conditioning.load where prompt_embeds_mask.shape != prompt_embeds.shape[:2] — e.g. a conditioning object saved with embeddings from one tokenizer config and a mask from another, or a manually constructed Krea2ConditioningInfo with mismatched tensors.

Common situations: Custom nodes constructing Krea2ConditioningInfo by hand with wrong mask dimensions; conditioning data persisted by an older InvokeAI version whose embedding width/token count changed after a model or library update; truncated/corrupted saved conditioning files.

Understand the failure class

Background: Tensor shape mismatch errors ("must have shape", "expected shape ... got ..."): when tensor dimensions disagree with what an op or layer was told to expect — this error's family across 6 libraries.

Related errors

• per_layer_weights must be comma-separated numbers: {e}
• 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.

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/943cce831da6d8ae. Report an issue: GitHub.