invoke-ai/InvokeAI · error · TypeError

Expected AutoencoderKLWan or FluxAutoEncoder, got {type(vae)

Error message

Expected AutoencoderKLWan or FluxAutoEncoder, got {type(vae).__name__}.

What it means

A post-device-residency re-check in the decode path: after model_on_device yields the VAE, the code re-validates it is AutoencoderKLWan or FluxAutoEncoder and raises this TypeError otherwise. The initial invoke() check passed, so this usually indicates the on-device object changed between checks.

Source

Thrown at invokeai/app/invocations/anima_latents_to_image.py:144

            )
            use_tiling = self._use_tiled_decode(TorchDevice.choose_torch_device(), full_decode_working_memory)
            estimated_working_memory = estimate_vae_working_memory_anima(
                operation="decode",
                image_tensor=latents,
                vae=vae_info.model,
                tile_size=ANIMA_VAE_TILE_SIZE if use_tiling else None,
            )
        else:
            estimated_working_memory = estimate_vae_working_memory_flux(
                operation="decode",
                image_tensor=latents,
                vae=vae_info.model,
            )

        with vae_info.model_on_device(working_mem_bytes=estimated_working_memory) as (_, vae):
            context.util.signal_progress("Running Anima VAE decode")
            if not isinstance(vae, (AutoencoderKLWan, FluxAutoEncoder)):
                raise TypeError(f"Expected AutoencoderKLWan or FluxAutoEncoder, got {type(vae).__name__}.")

            vae_dtype = next(iter(vae.parameters())).dtype
            # Use the VAE's intended compute device (CUDA/MPS, or CPU if configured cpu_only). Do NOT infer it from
            # current param residency: partial loading may have temporarily offloaded all weights to RAM, which would
            # wrongly place the latents (and thus the whole decode) on the CPU (see #9373).
            latents = latents.to(device=vae_info.compute_device, dtype=vae_dtype)

            TorchDevice.empty_cache()

            with torch.inference_mode():
                if isinstance(vae, FluxAutoEncoder):
                    # FLUX VAE handles scaling internally, expects 4D [B, C, H, W]
                    img = vae.decode(latents)
                else:
                    # The cached VAE instance is shared across invocations, so always set
                    # the tiling state explicitly rather than leaving it as-is.
                    if use_tiling:
                        vae.enable_tiling(

View on GitHub (pinned to 0b6a024f2f)

Solutions

  1. Re-run the generation so the model loads freshly and consistently.
  2. Update InvokeAI; on-device yield behavior should match the validated type.
  3. Avoid concurrent edits to the model list while a workflow runs; restart the app to clear residency state.

Example fix

# before: VAE offloaded/swapped mid-run causing mismatched on-device object
# after: restart backend / pin the model in memory (disable eager offload) so model_on_device yields the validated VAE
Defensive patterns

Strategy: type-guard

Validate before calling

with vae_info.model_on_device() as (_, vae):
    if not is_anima_vae(vae):
        raise TypeError("on-device VAE type mismatch; reload before decode")

Type guard

def is_anima_vae(vae) -> bool:
    from diffusers import AutoencoderKLWan
    from invokeai.backend.flux.model import FluxAutoEncoder
    return isinstance(vae, (AutoencoderKLWan, FluxAutoEncoder))

Try / catch

try:
    latents2img.invoke(context)
except TypeError as e:
    if "Expected AutoencoderKLWan or FluxAutoEncoder" in str(e) and "got" in str(e):
        restart_or_reload_model(vae_key)  # clear residency cache and retry

Prevention

When it happens

Trigger: The object yielded by model_on_device during decode is not the expected VAE instance — e.g. model record swapped/offloaded concurrently or an incompatibility between the loaded wrapper and current InvokeAI internals.

Common situations: Long-running generation during which the model was unloaded/reloaded; InvokeAI version mismatch in model residency code; a corrupted on-device cache.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/414c6445fddf399a. Report an issue: GitHub.