invoke-ai/InvokeAI · error · ValueError

denoising_start should be 0 when initial latents are not pro

Error message

denoising_start should be 0 when initial latents are not provided.

What it means

Without initial latents, denoising starts from pure noise at sigma_0, which corresponds to denoising_start == 0. `_run_diffusion` raises this ValueError when no init latents are supplied but denoising_start is greater than a small epsilon (1e-5), because skipping the early steps requires an initial latent to start from.

Source

Thrown at invokeai/app/invocations/krea2_denoise.py:373

                device,
            )
            neg_prompt_embeds = neg_extension.regional_text_conditioning.prompt_embeds

        # Load initial latents (img2img).
        init_latents = context.tensors.load(self.latents.latents_name) if self.latents else None
        if init_latents is not None:
            init_latents = init_latents.to(device=device, dtype=inference_dtype)
            if init_latents.dim() == 5:
                init_latents = init_latents.squeeze(2)

        noise = self._get_noise(self.height, self.width, inference_dtype, device, self.seed)

        if init_latents is not None:
            s_0 = sigmas_sched[0].item()
            latents = s_0 * noise + (1.0 - s_0) * init_latents
        else:
            if self.denoising_start > 1e-5:
                raise ValueError("denoising_start should be 0 when initial latents are not provided.")
            latents = noise

        # Pack latents into 2x2 patches: (B, C, H, W) -> (B, grid_h*grid_w, C*4).
        latents = pack_latents(latents, 1, KREA2_LATENT_CHANNELS, latent_height, latent_width)

        # Position ids: text tokens at origin, image tokens carry their grid coords.
        text_seq_len = pos_prompt_embeds.shape[1]
        position_ids = prepare_position_ids(text_seq_len, grid_height, grid_width, device)
        # The negative prompt can tokenize to a different length than the positive prompt, so it needs its
        # own position ids. Reusing the positive ids would leave the rotary embedding (text + image tokens)
        # a different length than the uncond query sequence and crash in the transformer's apply_rotary_emb.
        neg_position_ids = (
            prepare_position_ids(neg_prompt_embeds.shape[1], grid_height, grid_width, device)
            if neg_prompt_embeds is not None
            else None
        )

        # Inpaint extension operates in 4D, so unpack/repack around each merge.

View on GitHub (pinned to 0b6a024f2f)

Solutions

1. Set denoising_start to 0 when no initial latents are provided.
2. Provide initial latents (VAE Encode of an image) if a nonzero denoising_start is intended (img2img mode).
3. Guard caller code: only apply the refiner start fraction when init latents exist.

Example fix

// before: txt2img with a refiner start
denoise = Krea2Denoise(latents=None, denoising_start=0.7, denoising_end=1.0)
// after
denoise = Krea2Denoise(latents=None, denoising_start=0.0, denoising_end=1.0)
# or wire latents=vae_encode.latents for img2img

if latents is None and denoising_start > 1e-5:
    raise ValueError("Set denoising_start=0 for txt2img, or supply initial latents for img2img.")

def start_matches_mode(latents, denoising_start: float) -> bool:
    return latents is not None or denoising_start <= 1e-5

try:
    out = invoke_krea2_denoise(latents=latents, denoising_start=start)
except ValueError as e:
    if "denoising_start should be 0" in str(e):
        out = invoke_krea2_denoise(latents=latents, denoising_start=0.0, denoising_end=end)
    else:
        raise

When it happens

Trigger: Running the krea2_denoise invocation in pure txt2img mode (latents=None) with denoising_start > 1e-5 — e.g. a refiner-style start fraction applied without an img2img latent input.

Common situations: Users copying refiner settings (denoising_start=0.7) onto a txt2img graph; template workflows that enable a start fraction but lack the image-to-latents branch; scripts setting start values unconditionally.

Related errors

• The requested denoising range does not contain any effective
• denoising_start must be less than denoising_end.
• 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.

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/b5b5c92619bd3300. Report an issue: GitHub.