invoke-ai/InvokeAI · error · RuntimeError

Provided model was not a diffusers model/pipeline, as expect

Error message

Provided model was not a diffusers model/pipeline, as expected.

What it means

apply_hidiffusion (hidiffusion.py:2063) requires the target to be a diffusers DiffusionPipeline or ModelMixin (checked via isinstance_str) because it relies on pipeline attributes like `.unet` and `name_or_path`. Passing anything else — a bare torch module that is not a diffusers ModelMixin, a dict, a wrapper, or None — raises this RuntimeError before any patching.

Source

Thrown at invokeai/backend/hidiffusion/hidiffusion.py:2063

    generator: torch.Generator | None = None,
    has_controlnet: bool = False,
    is_controlnet_text_to_image: bool = False,
):
    """
    model: diffusers model. We support SD 1.5, 2.1, XL, XL Turbo.

    apply_raunet: whether to apply RAU-Net

    apply_window_attn: whether to apply MSW-MSA.
    """

    # Make sure the module is not currently patched
    remove_hidiffusion(model)

    is_diffusers = isinstance_str(model, "DiffusionPipeline") or isinstance_str(model, "ModelMixin")

    if not is_diffusers:
        raise RuntimeError("Provided model was not a diffusers model/pipeline, as expected.")
    else:
        # Check if the pipeline is a ControlNet pipeline. InvokeAI's modular
        # denoise passes a bare UNet, so it reports ControlNet separately.
        has_controlnet = has_controlnet or hasattr(model, "controlnet")
        is_sdxl_controlnet = hasattr(model, "controlnet") and isinstance_str(
            model, "StableDiffusionXLControlNet", prefix=True
        )
        is_sd_controlnet = hasattr(model, "controlnet") and isinstance_str(
            model, "StableDiffusionControlNet", prefix=True
        )

        # Check for ControlNet Inpaint pipelines
        is_sdxl_controlnet_inpaint = is_sdxl_controlnet and isinstance_str(model, "Inpaint", contains=True)
        is_sd_controlnet_inpaint = is_sd_controlnet and isinstance_str(model, "Inpaint", contains=True)

        if is_sdxl_controlnet_inpaint or is_sd_controlnet_inpaint:
            # For ControlNet Inpaint pipelines, we don't patch the pipeline class
            # because they already have all the necessary inpainting logic

View on GitHub (pinned to 0b6a024f2f)

Solutions

Pass the diffusers pipeline object itself (e.g. StableDiffusionXLPipeline instance), not an inner component or wrapper.
Apply HiDiffusion before torch.compile/DDP/accelerate wrapping so isinstance checks still match.
Verify the argument is not None — check the from_pretrained call succeeded.
If you only have a bare UNet, wrap/load it via a diffusers pipeline, or use a diffusers ModelMixin subclass.

Example fix

// before
apply_hidiffusion(pipe.unet, apply_raunet=True)
// after
apply_hidiffusion(pipe, apply_raunet=True, apply_window_attn=True)

Defensive patterns

Strategy: type-guard

Validate before calling

def is_diffusers_model(model) -> bool:
    from invokeai.backend.util import isinstance_str
    return isinstance_str(model, "DiffusionPipeline") or isinstance_str(model, "ModelMixin")

if not is_diffusers_model(model):
    raise TypeError("Pass the diffusers pipeline, not a wrapped/inner module")
apply_hidiffusion(model, apply_raunet=True)

Type guard

def is_hidiffusion_target(model) -> bool:
    from invokeai.backend.util import isinstance_str
    return model is not None and (
        isinstance_str(model, "DiffusionPipeline") or isinstance_str(model, "ModelMixin")
    )

Try / catch

try:
    apply_hidiffusion(model, apply_raunet=True, apply_window_attn=True)
except RuntimeError as e:
    if "not a diffusers model" in str(e):
        logger.error("apply_hidiffusion requires the raw diffusers pipeline (pre-compile, pre-wrap)")
    else:
        raise

Prevention

Apply HiDiffusion before torch.compile / DDP / accelerate wrapping.
Pass the pipeline object itself, never pipe.unet or a scheduler.
Null-check the model after from_pretrained to avoid passing None.
Avoid custom pipeline wrappers that don't subclass DiffusionPipeline.

When it happens

Trigger: Calling hidiffusion_patch/apply_hidiffusion with a raw torch.nn.Module UNet that is not a diffusers ModelMixin subclass, a compiled (torch.compile) or DDP-wrapped model whose type no longer reports as DiffusionPipeline/ModelMixin, a None value from a failed pipeline load, or the InvokeAI modular denoise passing an unexpected object.

Common situations: Applying HiDiffusion inside custom inference code where the model was already unwrapped/compiled; wrapping the pipeline in a accelerator/distributed wrapper first; typos passing scheduler or text_encoder instead of the pipeline.

Understand the failure class

Background: Invalid argument type errors: "must be of type string", "expected X, got Y", and ERR_INVALID_ARG_TYPE explained — this error's family across 15 libraries.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/d46b4df1004b613d. Report an issue: GitHub.