invoke-ai/InvokeAI · error · TypeError

Expected AutoencoderKL or FluxAutoEncoder, got {type(vae).__

Error message

Expected AutoencoderKL or FluxAutoEncoder, got {type(vae).__name__}. VAE model type changed unexpectedly after loading.

What it means

Inside the device-context manager (model_on_device), the VAE object handed back is re-checked: it must still be AutoencoderKL or FluxAutoEncoder. This is a defensive invariant check - the type was already validated at load time, so failure here means the model wrapper returned a different object than expected (e.g. a swapped, re-loaded, or partially-converted model). The message says 'changed unexpectedly after loading'.

Source

Thrown at invokeai/app/invocations/z_image_latents_to_image.py:72

        is_flux_vae = isinstance(vae_info.model, FluxAutoEncoder)

        # Estimate working memory needed for VAE decode
        estimated_working_memory = estimate_vae_working_memory_flux(
            operation="decode",
            image_tensor=latents,
            vae=vae_info.model,
        )

        # FLUX VAE doesn't support seamless, so only apply for AutoencoderKL
        seamless_context = (
            nullcontext() if is_flux_vae else SeamlessExt.static_patch_model(vae_info.model, self.vae.seamless_axes)
        )

        with seamless_context, vae_info.model_on_device(working_mem_bytes=estimated_working_memory) as (_, vae):
            context.util.signal_progress("Running VAE")
            if not isinstance(vae, (AutoencoderKL, FluxAutoEncoder)):
                raise TypeError(
                    f"Expected AutoencoderKL or FluxAutoEncoder, got {type(vae).__name__}. "
                    "VAE model type changed unexpectedly after loading."
                )

            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)

            # Disable tiling for AutoencoderKL
            if isinstance(vae, AutoencoderKL):
                vae.disable_tiling()

            # Clear memory as VAE decode can request a lot
            TorchDevice.empty_cache()

            with torch.inference_mode():

View on GitHub (pinned to 0b6a024f2f)

Solutions

Re-run the invocation; transient model-cache swaps are the usual cause.
Verify the VAE field still points at a Z-Image/Flux compatible VAE and re-open the workflow.
Clear/restart the model RAM cache or restart InvokeAI to flush stale loaded models.
Check for multiple concurrent workflows sharing the VAE and serialize them; update InvokeAI if reproducible.

Defensive patterns

Strategy: try-catch

Type guard

def vae_stable_after_load(vae) -> bool:
    from diffusers import AutoencoderKL
    from invokeai.backend.flux.vae import FluxAutoEncoder
    return isinstance(vae, (AutoencoderKL, FluxAutoEncoder))

Try / catch

try:
    out = z_image_l2i.invoke(context)
except TypeError as e:
    if "changed unexpectedly" in str(e):
        context.logger.warning("Model cache swapped VAE mid-run; retrying once.")
        out = retry(z_image_l2i.invoke, context)
    else:
        raise

Prevention

Avoid running many concurrent workflows that compete for the same VAE in RAM cache.
Restart InvokeAI / clear model cache if this error recurs after model changes.
Re-run the invocation first - transient swaps often resolve on retry.

When it happens

Trigger: Within invoke() of ZImageLatentsToImage, entering `with ... vae_info.model_on_device(...) as (_, vae)` and the yielded object fails isinstance(vae, (AutoencoderKL, FluxAutoEncoder)).

Common situations: Concurrency/memory pressure causing the model to be unloaded and swapped mid-session; a patched loader returning a wrapper or converted dtype object; race with another invocation replacing the model; exotic SeamlessVae wrapper logic interacting with a non-standard VAE.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/54244273492530d9. Report an issue: GitHub.