FunUI Decode:
FunUIDecode turns a generated latent—the compressed, not-yet-viewable representation of a picture or video—into visual content you can preview, process, or save in your workflow. It also supports decoding sound for models that generate audio alongside video. Its main benefit is that it accommodates both ordinary latents and models that require a specialized decoding process: it first checks whether an installed FunUI module can handle the model’s particular format, and otherwise uses standard VAE decoding when appropriate. A VAE is the component that translates between compressed representations and visible pictures. The node also offers tile-based decoding to help reduce graphics-memory requirements. To protect the result, it stops with an explanatory error rather than attempting an inappropriate fallback when a specialized decoder fails or a multipart latent has no compatible decoder.
FunUI Decode Input Parameters:
samples
Required input of type LATENT. This is the generated latent you want to convert into viewable content, typically connected from a sampler such as FunUISampler. The node reads the latent data stored under the samples entry of this input and checks whether an installed decoder can handle it. Ordinary latents can proceed through standard VAE decoding, while specialized or multipart latents need a compatible decoding module. Connect the completed sampling result rather than an image or a model connection; decoding translates the supplied representation rather than running another round of generation. No default value, numeric range, or selectable options are defined.
vae
Required input of type VAE. This is the picture VAE, which translates the visual latent into images or video frames. For the standard decoding path, this component performs the actual image reconstruction; it is also passed to specialized decoders when they are checked. Use a VAE compatible with the model and latent format you generated, because an incompatible VAE can fail or produce incorrect-looking results. In the supplied SD1.5 workflow example, this connection comes from the same model loader used for generation. This input remains required even when you provide a model-specific decoder or an audio VAE. No default value, numeric range, or selectable options are defined.
model
Optional input of type MODEL, with a default of None when unconnected. Supply the model used to generate the latent when its format requires specialized decoding. The node passes this connection to the installed decoder modules so they can recognize and handle the appropriate model; ordinary single-latent decoding does not require it. Connecting it is particularly useful for specialized video or audiovisual workflows, although it does not replace the need to install the appropriate FunUI module. If a multipart latent cannot be decoded and this connection is missing, the node includes an additional message explaining that no model was wired for recognition. No numeric range or selectable options are defined.
audio_vae
Optional input of type VAE, with a default of None when unconnected. This connection provides the audio VAE for models that generate sound alongside video. It is passed to specialized decoding modules, allowing those modules to use it when their decoding process requires it. It does not add sound to an ordinary image latent, and the standard single-latent decoding path does not use it or produce audio. Connect the compatible audio VAE required by your particular audiovisual model rather than assuming the picture VAE can perform both jobs. No numeric range or selectable options are defined.
tile_size
Optional input of type INT that controls the requested decoding tile size in pixels. The default is 0, the minimum is 0, the maximum is 4096, and the interface step is 64. Tiled decoding processes the picture in smaller sections to help fit decoding into less graphics memory; this setting concerns decoding, not the intended output resolution. A value of 0 means use the model’s own decoding behavior, not necessarily disable tiling. The interface specifically notes that MiniMax H3 uses ComfyUI’s 256-pixel tiles at 0, matching the tile size its decoder was trained on, and that larger tiles show a grid. It also states that an X2 Detail VAE is always tiled, with a minimum tile size of 256. The value is passed to specialized decoders as well as the ordinary decoding helper, so its precise handling depends on the decoding path.
FunUI Decode Output Parameters:
images
Output of type IMAGE containing the decoded picture or sequence of video frames. Connect this to an image preview, further image-processing nodes, or a compatible saving node. For ordinary decoding, if the result has five dimensions representing a batch of clips, the node combines the clip and frame dimensions into a single frame batch so downstream nodes receive a sequence of images. This output is decoded visual data, not an encoded video file, and no frame-rate information is explicitly returned by this node. For a specialized decoding path, the image result is supplied by the module that handled decoding.
audio
Output of type AUDIO containing the audio result returned by a specialized decoder, when available. This lets you pass generated sound onward to compatible processing or saving nodes alongside the decoded frames. The output does not guarantee that every decoded result includes sound: the ordinary single-latent decoding path explicitly returns None, and specialized decoding only provides audio when supported by the model and its decoder. An absent audio result is therefore normal for image-only generation or workflows without generated sound.
status
Output of type STRING reporting which decoding route completed successfully. Standard decoding returns the exact text decoded as a single latent. When an installed module handles decoding, the text contains that module’s identifier followed by decoded this. This is useful for confirming whether the result came from ordinary VAE decoding or a model-specific implementation. It is a short execution report rather than an image-quality assessment, performance measurement, or saved-file location. If decoding raises an error, the node stops instead of returning a successful status.
FunUI Decode Usage Tips:
- Connect the sampler’s latent result to
samplesand a compatible picture VAE tovae. For a straightforward workflow, follow the supplied SD1.5 example by taking the model and VAE from the same loader used for generation. - Leave
tile_sizeat 0 initially. This preserves model-specific default behavior and avoids assuming that every model should use the same tile size. - If decoding runs out of graphics memory, try a smaller supported nonzero
tile_sizewhere the decoder permits it. This setting targets decoding memory; it does not reduce the memory already needed for sampling. - For MiniMax H3, prefer
tile_size0 so decoding uses the documented 256-pixel tiles. Avoid increasing the tile size to chase performance, because the interface explicitly warns that larger tiles show a grid. - When using an X2 Detail VAE, expect tiled decoding even at the default setting. Its documented minimum tile size is 256, so a smaller requested value should not be treated as a way to force smaller tiles or disable tiling.
- Connect
modelfor specialized video or audiovisual workflows, and ensure the corresponding FunUI module is installed. Providing the model helps compatible modules recognize the latent, but cannot supply a missing decoder. - Connect
audio_vaeonly when your model generates sound and its decoder requires that component. For silent or image-only workflows, an emptyaudiooutput is expected. - Use
statusto check which decoding path ran. When saving video, configure the frame rate in the downstream saving workflow; FunUIDecode does not expose a frame-rate control or return frame-rate metadata as a separate output.
FunUI Decode Common Errors and Solutions:
“MODULE_ID handles this model's decode and failed: EXCEPTION_TYPE: DETAILS. Refusing to decode it as a plain latent, which would not be the same picture.”
- Explanation: This is the node’s error template for a failure inside a specialized decoding module; the capitalized placeholders are replaced with the actual module identifier, exception type, and details. The node deliberately stops because treating that model’s latent as an ordinary latent could produce an incorrect result.
- Solution: Read the exception details and ComfyUI log to identify the underlying issue. Check that the latent, model, picture VAE, and any required audio VAE belong to a compatible workflow. Verify that the relevant FunUI module and its dependencies are installed correctly. If the underlying error reports insufficient graphics memory, try an appropriate smaller tile size where supported rather than attempting to bypass specialized decoding.
“This latent has more than one part and no installed module knows how to decode it. The model input is not wired, so no module could recognise it. The module for this model is missing.”
- Explanation: The latent contains multiple parts, no installed decoder accepted it, and
modelwas left unconnected. The node cannot safely interpret it as one ordinary visual latent. The message flags both the missing model connection and the lack of a recognized decoder. - Solution: Connect the model that generated the latent to
modeland try again. If decoding is still unrecognized, install or enable the appropriate model-support module and check that it matches your FunUI installation. Also connect the correct picture VAE and any audio VAE required by that model.
“This latent has more than one part and no installed module knows how to decode it. The module for this model is missing.”
- Explanation: This version occurs when
modelis connected but no installed decoder accepts the multipart latent. Supplying a model connection alone is not enough when compatible decoding support is missing or cannot recognize the supplied combination. - Solution: Confirm that the connected model actually produced the latent and that its FunUI support module is installed and enabled. Check the ComfyUI startup and execution logs for module-loading problems, and use mutually compatible versions of the model-support module and FunUI.
“KeyError: 'samples'”
- Explanation: This can occur because the node directly reads
samples["samples"]and the supplied object does not contain that expected entry. It indicates malformed latent data or an upstream connection producing an unexpected structure, rather than a tile-size problem. - Solution: Connect a valid LATENT output from your sampler or another compatible latent-producing node. Inspect any custom nodes that modify or construct the latent, and confirm that their output preserves the required
samplesentry.
“CUDA out of memory” or “torch.OutOfMemoryError”
- Explanation: These are possible underlying decoder errors, not fixed messages defined by FunUIDecode. They indicate that decoding needs more graphics memory than is available. When the failure happens inside a specialized module, this message may appear within the node’s longer module-failure error.
- Solution: Close other GPU-heavy applications and try a smaller supported decoding tile size. If necessary, regenerate at a lower resolution or with fewer frames or a smaller batch. Respect model-specific constraints: MiniMax H3 should retain its documented 256-pixel tile behavior, and X2 Detail uses tiles of at least 256.
Decoder shape, channel, or tensor-size mismatch errors
- Explanation: These are possible underlying errors whose exact text depends on the VAE or specialized decoder. They can indicate that the decoder received a latent format it does not support, such as a latent paired with an incompatible VAE or model.
- Solution: Reconnect the picture VAE and model intended for the latent’s model family, and avoid mixing latent outputs from unrelated workflows. For specialized formats, confirm that the appropriate decoder module is available and that
modelis connected. Use the detailed exception in the ComfyUI log to distinguish an incompatible component from another decoder issue.
