Open-H3 IC · Compose Target Latent:
OpenH3ICComposeLatent, displayed in ComfyUI as “Open-H3 IC · Compose Target Latent,” prepares the combined video-and-audio latent representation used for the Open-H3 Regenerate 2K IC workflow. A latent is a compressed representation that the model processes before it is decoded into visible frames or audio. This node combines an already upscaled video representation with the official silence representation, giving the subsequent sampling stage a prepared audiovisual starting point. It can also create a mask that allows the video to be refined while keeping the audio portion pinned to silence. This supports a workflow focused on regenerating the picture rather than generating a new soundtrack; in the supplied example workflow, the draft’s original audio is attached separately when the final video is assembled. The node prepares and validates this combined representation but does not itself upscale video, run sampling, or decode media.
Open-H3 IC · Compose Target Latent Input Parameters:
video_latent
Required input of type LATENT containing the video representation you want to use as the starting point for regeneration. In the supplied workflow, this comes from MiniMaxH3LatentUpscaleWithModel after the draft has been VAE-encoded and upscaled to the target resolution. If this input contains a combined representation, the node takes its first component as the video. It automatically converts the video’s numerical precision and processing device to match the reference video component, but it does not resize or reshape the video: its complete tensor shape must already match the reference, including all dimensions, not just width and height. This determines the actual video content placed into the composed latent. Additional metadata from this input is carried forward and takes precedence over matching reference metadata, except for fields subsequently set or removed by the node. No default value, numerical minimum, maximum, or selectable options are defined; you must connect a compatible LATENT.
reference_latent
Required input of type LATENT providing the joint audiovisual target representation from MiniMax H3 Reference to Video. It must contain exactly two components: a video component followed by an audio component. The node uses these components as templates for the required shapes, numerical precision, and processing devices, and uses the reference latent’s metadata as the starting point for the result. The reference video content is replaced by the supplied video representation, and the reference audio content is replaced by the official silence representation. The video shapes must match exactly. For audio, the first three dimensions of the prepared silence must match the reference audio; only its final dimension is adjusted automatically through trimming or zero-padding. No default value, numerical minimum, maximum, or selectable options are defined; you must connect the compatible joint audiovisual LATENT rather than a video-only latent.
pin_audio
Required BOOLEAN setting with a default of True and the options True or False; numerical minimum and maximum values do not apply. When True, the node creates a noise mask filled with ones for the video component and zeros for the silence component, marking video as available for sampling changes while protecting the audio portion in downstream sampling that honors this mask. When False, the node removes any existing noise mask from the composed result, so audio is no longer protected by that mask. Both settings still initialize the audio component with official silence: False does not restore the reference audio or copy the draft’s soundtrack. Leave this enabled for the supplied picture-regeneration workflow, which attaches the original soundtrack separately at the end.
Open-H3 IC · Compose Target Latent Output Parameters:
latent
Output of type LATENT containing the prepared joint video-and-audio representation for subsequent sampling. Its samples consist of two components: the supplied video, converted to match the reference video’s device and numerical precision, and the official silence audio, fitted to the reference audio’s final dimension. The result retains reference metadata and incorporates additional metadata from the video latent, with matching video metadata taking precedence before the node applies its own final settings. When pin_audio is True, it includes a two-component noise mask that permits video changes and protects audio; when False, no noise_mask field is retained. It also includes an open_h3_ic metadata entry recording audio as “official_silence” and pinned as the selected Boolean value. This output is a prepared model representation, not a finished video or audio file, and must pass through the appropriate sampling and decoding stages.
Open-H3 IC · Compose Target Latent Usage Tips:
- Connect the upscaled draft latent from MiniMaxH3LatentUpscaleWithModel to video_latent and the joint audiovisual latent from MiniMax H3 Reference to Video to reference_latent. A video-only latent cannot serve as the reference.
- Set the latent upscaler and reference-generation stage to matching target dimensions and compatible frame settings. The supplied workflow uses 2560×1440, but this node does not impose that resolution; it requires the complete video latent shapes to match.
- Keep pin_audio set to True when you want to regenerate the picture while protecting the silent audio component during sampling. Preserve the draft’s original soundtrack through a separate audio connection to your final video assembly stage, as shown in the supplied workflow.
- Set pin_audio to False only when you intentionally want to remove this node’s audio-protection mask. The audio still starts as official silence, and this setting alone does not guarantee useful generated audio.
- If VRAM is limited, the supplied workflow notes suggest lowering the target to 1920×1088 and using the pruned INT8 reference model. Update both reference preparation and latent upscaling consistently; changing only one side will cause a shape mismatch.
- Use the composed latent with the appropriate Open-H3 IC sampling setup. The supplied workflow pairs it with the up3 sigma schedule, using 0.834942, 0.696852, 0.468753, and 0; this node does not select or apply those values itself.
Open-H3 IC · Compose Target Latent Common Errors and Solutions:
Open-H3 IC · Compose Target Latent must run inside ComfyUI
- Explanation: The node could not import ComfyUI’s nested-tensor support or the safetensors loader. This can happen when running the code outside ComfyUI or when the active Python environment lacks a required dependency.
- Solution: Run the node through ComfyUI with the custom-node extension installed. If the error persists, check the underlying import error in the console, verify the extension’s requirements in ComfyUI’s Python environment, and restart ComfyUI.
reference_latent must be the joint AV latent from MiniMax H3 Reference to Video
- Explanation: The reference samples are not a combined audiovisual representation with exactly two components. A standard video-only latent does not meet this requirement.
- Solution: Connect the joint audiovisual LATENT from MiniMax H3 Reference to Video to reference_latent. Check that an intermediate node has not replaced it with a video-only representation.
Upscaled video latent shape … does not match reference AV target shape …
- Explanation: The complete shape of the supplied video differs from the reference video shape. Differences can involve frame-related dimensions, spatial dimensions, batch size, or the latent format. The actual error includes both shapes to help identify the mismatch.
- Solution: Compare the two reported shapes and align the draft preparation, reference settings, and latent upscaler. Ensure both paths use compatible models, frame settings, batch sizes, and target resolutions; this node does not resize incompatible video latents.
Silence latent … does not match target …
- Explanation: The first three dimensions of the loaded silence representation do not match the reference audio. The node can trim or zero-pad only the final dimension, so it cannot repair other shape differences.
- Solution: Use the official silence asset supplied for this extension and a compatible MiniMax H3 audiovisual reference. Check for incompatible batch settings or assets from another model version. Consult the reported shapes to identify which dimensions differ.
FileNotFoundError or OSError while loading the official silence asset
- Explanation: The node loads the silence asset from the location configured by the extension. If that file is missing, inaccessible, or unreadable, composition cannot continue. The exact error depends on the file-loading failure.
- Solution: Check the console for the affected path, complete the extension’s asset setup, and confirm that ComfyUI can read the file. Restore the official asset rather than substituting an unrelated audio file.
KeyError: 'audio_latents'
- Explanation: The silence file was opened, but it does not contain the audio_latents entry expected by the node. This usually indicates an incorrect or incompatible asset.
- Solution: Restore the correct official silence safetensors file expected by the extension. Renaming another file will not provide the required contents.
KeyError: 'samples'
- Explanation: One of the connected LATENT objects lacks the samples field that stores its actual latent data. Although the socket is labeled LATENT, the supplied object is incomplete or incompatible.
- Solution: Reconnect the output from the appropriate upstream latent-producing node. If you use intermediate custom nodes, check whether one removes or replaces the samples field.
AttributeError involving 'is_nested' or 'tensors'
- Explanation: The node received samples in an unexpected data structure, or the installed ComfyUI version does not provide the expected nested audiovisual representation.
- Solution: Verify that both connections come from the intended H3 workflow nodes. If the wiring is correct, check compatibility between your ComfyUI version and the extension, update as appropriate, and restart ComfyUI.
Safetensors loading or deserialization error
- Explanation: The silence asset cannot be read as a valid safetensors file. It may be damaged, incomplete, or a different file type saved with a safetensors extension.
- Solution: Download or restore the official silence asset again using the extension’s setup instructions. Confirm that the download completed and that the file is not an HTML error page or another unrelated file.
Out of memory or CUDA out of memory
- Explanation: Preparing the result may require device transfers and, when audio pinning is enabled, full-size mask tensors. A high-resolution workflow can run out of memory at this stage even if earlier steps completed.
- Solution: Free unused GPU memory and reduce the target resolution or batch size while keeping both video paths consistent. Consider the lower-resolution and pruned-model configuration described in the supplied workflow. Do not disable pin_audio solely to save memory unless you also accept removing audio protection during sampling.
