invoke-ai/InvokeAI · error · ValueError

Initial latents are required when a denoise mask is provided

Error message

Initial latents are required when a denoise mask is provided.

What it means

A denoise mask tells the pipeline which latent regions to preserve, which only makes sense when there are initial latents to mask. `_validate_inputs` raises this ValueError when denoise_mask is provided but the `latents` input is None.

Source

Thrown at invokeai/app/invocations/krea2_denoise.py:233

        raise ValueError(f"Invalid CFG scale type: {type(self.cfg_scale)}")

    @staticmethod
    def _should_apply_cfg_for_step(cfg_scale: float, *, has_negative_conditioning: bool) -> bool:
        return has_negative_conditioning and cfg_scale > 1.0

    @staticmethod
    def _validate_effective_schedule(*, start_idx: int, end_idx: int) -> None:
        if end_idx <= start_idx:
            raise ValueError(
                "The requested denoising range does not contain any effective denoising steps at the configured "
                "step count. Increase denoising_end, decrease denoising_start, or increase steps."
            )

    def _validate_inputs(self) -> None:
        if self.denoising_start >= self.denoising_end:
            raise ValueError("denoising_start must be less than denoising_end.")
        if self.denoise_mask is not None and self.latents is None:
            raise ValueError("Initial latents are required when a denoise mask is provided.")

    def _is_distilled(self, context: InvocationContext) -> bool:
        """Whether the transformer is the distilled Turbo checkpoint (fixed mu) vs. Raw (dynamic mu).

        Prefer the classified variant (works for diffusers, single-file and GGUF alike); fall back to
        the pipeline-level ``is_distilled`` flag in model_index.json, then default to distilled.

        A failed config lookup is a real error and is allowed to propagate — silently defaulting to the
        Turbo shift would apply the wrong sampling schedule to a Raw model.
        """
        from invokeai.backend.model_manager.taxonomy import Krea2VariantType

        config = context.models.get_config(self.transformer.transformer)
        variant = getattr(config, "variant", None)
        if variant is not None:
            return variant != Krea2VariantType.Base
        # No classified variant (unexpected for Krea-2) — fall back to the pipeline-level flag. Only a
        # missing/malformed model_index.json is tolerated here; it defaults to the distilled behavior.

View on GitHub (pinned to 0b6a024f2f)

Solutions

  1. Connect initial latents (from a VAE Encode or Resize Latents node output) to the invocation's latents input.
  2. If no initial image is intended, remove the denoise_mask connection.
  3. Check the workflow for a disabled or deleted node upstream of the latents input.

Example fix

// before: mask without latents
DenoiseInvocation(denoise_mask=mask, latents=None)
// after: provide initial latents
DenoiseInvocation(denoise_mask=mask, latents=vae_encode.latents)
Defensive patterns

Strategy: validation

Validate before calling

if denoise_mask is not None and latents is None:
    raise ValueError("Provide initial latents (e.g. VAE Encode output) when using a denoise mask.")

Type guard

def mask_has_latents(denoise_mask, latents) -> bool:
    return denoise_mask is None or latents is not None

Try / catch

try:
    out = invoke_krea2_denoise(denoise_mask=mask, latents=latents)
except ValueError as e:
    if "Initial latents are required" in str(e):
        latents = vae_encode(image).latents
        out = invoke_krea2_denoise(denoise_mask=mask, latents=latents)
    else:
        raise

Prevention

When it happens

Trigger: Connecting a DenoiseMaskField to the krea2_denoise node while leaving the latents input unconnected — e.g. running txt2img with a mask instead of img2img/inpaint.

Common situations: Inpainting graphs where the initial-image/VAE-encode branch was disconnected; users expecting mask-based txt2img; workflow templates missing the latents edge.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/d75f442bf30cde1a. Report an issue: GitHub.