FunUI VAE Loader:
FunUIVAELoader, displayed as “FunUI VAE Loader” in the FunUI/Loaders category, loads a VAE with an explicitly chosen computation precision. A VAE is the component that helps translate between ordinary media and the compressed representation, called a latent, used by many generative models. Choosing a compatible VAE is important because the generation model and the component that reconstructs your images, video, or audio must understand the same representation. This node prepares the VAE for other nodes to use; it does not generate, encode, or decode media by itself. During loading, it reads both the model weights and available file metadata, adjusts a specific naming convention used by some audio VAE files, and checks that ComfyUI recognizes the resulting VAE as valid. Its main benefit is giving you direct control over the VAE’s working precision while supporting loading through ComfyUI’s VAE implementation.
FunUI VAE Loader Input Parameters:
vae_name
Required COMBO parameter that selects the VAE file to load. The available choices are supplied dynamically by the node’s VAE file-listing helper, so they depend on the files available in your installation rather than a fixed list included in the source. The selected name is resolved through ComfyUI’s registered VAE folders, and the file’s weights and metadata are then loaded. Choose a VAE intended for your generation model and media type: a file that loads successfully is not necessarily compatible with every workflow. For example, an audio VAE should only be used where the workflow expects that kind of VAE. Options: the filenames returned by the VAE file-listing helper; the supplied context does not include that helper’s implementation or its exact filtering rules. Default: no explicit default is declared in the schema. Minimum and maximum: not applicable because this is a file-selection menu.
dtype
Required COMBO parameter that determines the numerical precision the VAE uses for computation. Precision affects how numerical values are represented during encoding and decoding, which can influence memory use, hardware compatibility, and numerical stability. The selected setting is mapped to a computation dtype before the VAE is constructed; this is not a setting for image resolution, compression quality, or output file format. Default: bfloat16. The node’s tooltip specifically states that MiniMax H3 needs bfloat16, making this setting particularly important for that workflow. Options: the keys of the shared DTYPES mapping; its complete definition is not provided, so the full menu cannot be reliably enumerated from this context. Minimum and maximum: not applicable because this is a selection menu. Keep the default unless your model documentation or hardware requirements call for another option available in your installed node.
FunUI VAE Loader Output Parameters:
vae
A VAE object with type VAE, ready to connect to compatible encoding, decoding, or other VAE-dependent nodes. This object contains the loaded autoencoder, not an image, video, audio clip, or latent result. The loader constructs it using the selected file’s weights, the chosen computation precision, and any metadata returned by the file loader. Before returning it, the node asks ComfyUI to check that the VAE is valid; loading stops with an error if that check fails. For certain audio VAE files, weight names beginning with audio_vae. are changed to begin with autoencoder. before construction, accommodating the naming convention identified in the source. Successful validation means ComfyUI accepted the VAE, but you still need to ensure it matches the model and downstream nodes in your workflow.
status
A STRING containing a brief confirmation of successful loading in the format filename loaded as dtype. For example, if the selected filename is example_vae.safetensors and the precision is bfloat16, the returned text is example_vae.safetensors loaded as bfloat16. This message helps you confirm which file and precision were used, especially when your workflow has several loaders. The same confirmation text is also logged. It is returned only after the VAE has been constructed and passed its validity check; it is not a progress report, an error output, or a guarantee that later encoding or decoding operations will succeed.
FunUI VAE Loader Usage Tips:
- Start with
bfloat16, the node’s declared default and the setting used by the provided FunUI system and LTX pipeline examples. For MiniMax H3, keepbfloat16as required by the node’s tooltip. - Choose the VAE recommended for your exact model family and task. Do not assume that every VAE listed in the menu can decode the latents produced by your selected diffusion model.
- Place your VAE file in a registered ComfyUI VAE directory, commonly
ComfyUI/models/vae. If a newly added file does not appear, refresh the model list if your installation supports it, or restart ComfyUI and reopen the selection menu. - Connect
vaeto a compatible VAE encode or decode node. Loading the VAE alone does not convert media; the downstream node performs that work. - In workflows with separate video and audio VAEs, use separate loader instances and select the correct file for each branch. The provided LTX pipeline includes distinct VAE and audio VAE loaders, both configured with
bfloat16. - Preserve metadata when obtaining or converting audio VAE files. The source notes that an audio VAE may store its model configuration in the file header, and this loader passes that metadata to ComfyUI when constructing the VAE.
- If you encounter memory problems during encoding or decoding, reduce the downstream workload or use compatible tiled processing where available. Changing the loader’s precision can affect memory use, but it should not replace checking the model’s precision requirements and your hardware’s capabilities.
FunUI VAE Loader Common Errors and Solutions:
Selected VAE file cannot be found
- Explanation: The loader resolves the selected filename through ComfyUI’s registered VAE folders and raises an error if the file cannot be located. This can happen after a file is moved, renamed, or deleted, or when a workflow references a file that is not installed on your computer. The exact error wording depends on ComfyUI.
- Solution: Confirm that the file exists in a registered VAE directory, refresh the available files or restart ComfyUI, and select the installed filename again.
VAE file cannot be read or deserialized
- Explanation: Errors during file loading can indicate an incomplete download, a damaged file, an unsupported file format, or a file that is not actually model weights. File-reader messages vary by format and ComfyUI version; this node does not define its own message for these failures.
- Solution: Download the file again from a trusted source, verify its size or checksum when available, and confirm that it is a VAE file supported by your ComfyUI installation. Renaming a file extension does not convert its contents.
VAE validation failed or no supported VAE was detected
- Explanation: The node explicitly checks whether the constructed VAE is valid. Validation can fail when the selected weights do not describe a supported VAE, when necessary information is missing, or when your installed ComfyUI version does not recognize the architecture. The supplied code does not show the exact exception message.
- Solution: Verify that you selected the intended VAE rather than a diffusion model, text encoder, or unrelated checkpoint. Follow the model’s installation instructions, preserve required metadata, and check whether the model requires a newer compatible ComfyUI version.
Selected dtype is not available
- Explanation: A saved or manually edited workflow can reference a precision value that is absent from the installed node’s
DTYPESmapping. Depending on where it is detected, this can appear as a selection-validation error or aKeyErrorfor the requested value. - Solution: Reselect a precision from the node’s current menu. Use
bfloat16as the documented default, particularly for MiniMax H3, and avoid manually entering precision names that the menu does not offer.
Computation dtype is not supported on the current device
- Explanation: Some hardware and processing operations do not support every numerical precision. A device or operation-specific error may occur while preparing the VAE or later when it is used, even if the file itself is valid.
- Solution: Check the model’s hardware and precision requirements. If the model permits it, select another supported precision from the actual menu; if it requires
bfloat16, use a compatible execution environment rather than assuming another precision is interchangeable.
CUDA out of memory
- Explanation: This is a possible PyTorch runtime error, not a message defined by the node. It means that available GPU memory is insufficient while loading or using the VAE. Encoding or decoding large images, long videos, or large batches can require substantially more memory than loading alone.
- Solution: Close other GPU-heavy applications and unload unnecessary models. If the failure happens downstream, reduce batch size, resolution, or frame count, or use compatible tiled encoding or decoding. Only change precision when the model and your hardware support the alternative.
Tensor size, shape, or channel mismatch during encoding or decoding
- Explanation: A VAE can pass its loading check but still be incompatible with the latents or media supplied by another node. Downstream errors mentioning mismatched dimensions, sizes, or channels may indicate that the VAE belongs to a different model family or media type.
- Solution: Use the VAE specified for your generation model, check that audio and visual branches are connected correctly, and verify that the downstream nodes support that VAE’s architecture. Successful loading alone does not establish workflow compatibility.
