invoke-ai/InvokeAI · warning · NotAMatchError

unable to determine model variant from state dict

Error message

unable to determine model variant from state dict

What it means

Variant detection reads the first conv layer 'model.diffusion_model.input_blocks.0.0.weight' and maps its in_channels to a variant (4=Normal, 5=Depth/SD2, 9=Inpaint). If that key is entirely absent from the state dict, the model cannot be classified as any variant and NotAMatchError is raised — the file likely is not a standard SD UNet checkpoint.

Source

Thrown at invokeai/backend/model_manager/configs/main.py:412

            if key_name in state_dict and state_dict[key_name].shape[-1] == 1024:
                if "global_step" in state_dict:
                    if state_dict["global_step"] == 220000:
                        return SchedulerPredictionType.Epsilon
                    elif state_dict["global_step"] == 110000:
                        return SchedulerPredictionType.VPrediction
            return SchedulerPredictionType.VPrediction
        else:
            return SchedulerPredictionType.Epsilon

    @classmethod
    def _get_variant_or_raise(cls, mod: ModelOnDisk) -> ModelVariantType:
        base = cls.model_fields["base"].default

        state_dict = mod.load_state_dict()
        key_name = "model.diffusion_model.input_blocks.0.0.weight"

        if key_name not in state_dict:
            raise NotAMatchError("unable to determine model variant from state dict")

        in_channels = state_dict["model.diffusion_model.input_blocks.0.0.weight"].shape[1]

        match in_channels:
            case 4:
                return ModelVariantType.Normal
            case 5:
                # Only SD2 has a depth variant
                assert base is BaseModelType.StableDiffusion2, f"unexpected unet in_channels 5 for base '{base}'"
                return ModelVariantType.Depth
            case 9:
                return ModelVariantType.Inpaint
            case _:
                raise NotAMatchError(f"unrecognized unet in_channels {in_channels} for base '{base}'")

    @classmethod
    def _validate_looks_like_main_model(cls, mod: ModelOnDisk) -> None:
        has_main_model_keys = _has_main_keys(mod.load_state_dict())

View on GitHub (pinned to 0b6a024f2f)

Solutions

1. Point the scan at the actual full-checkpoint file containing the UNet keys, not a VAE/text-encoder file.
2. If the model is diffusers-format (unet/ folders), ensure you are using the diffusers config class / directory scan, not the single-file checkpoint scan.
3. Verify the file downloaded completely and contains model.diffusion_model.* keys.
4. Re-export or convert the model to the legacy checkpoint layout if needed.

Example fix

// before
scan('model/vae/diffusion_pytorch_model.safetensors')  # no UNet keys
// after
scan('model/sd_xl_base_1.0.safetensors')               # full checkpoint

sd = load_file('model.safetensors')
if 'model.diffusion_model.input_blocks.0.0.weight' not in sd:
    print('Missing UNet input conv key — not a legacy main checkpoint')

def has_unet_input_conv(sd: dict) -> bool:
    return 'model.diffusion_model.input_blocks.0.0.weight' in sd

try:
    cfg = probe_model(path)
except NotAMatchError as e:
    if 'unable to determine model variant' in str(e):
        print('Scan the full checkpoint file, not a VAE/subcomponent')

When it happens

Trigger: from_model_on_disk → _get_variant_or_raise on a state dict missing 'model.diffusion_model.input_blocks.0.0.weight' — e.g. a VAE, a text-encoder-only dump, a diffusers-unet-style checkpoint without the legacy key prefix, or an empty/partial state dict.

Common situations: Pointing the scanner at the wrong file inside a checkpoint repo (e.g. the VAE or CLIP files); diffusers-format checkpoints probed by legacy checkpoint configs; truncated downloads.

Related errors

• Unexpected key: {k}
• missing keys after fp8 load: {missing[:10]}
• unable to determine base type from state dict
• unrecognized unet in_channels {in_channels} for base '{base}
• state dict does not look like a main model

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/935baf17566f30ed. Report an issue: GitHub.