ComfyUI 0.44 adds controls for video export and model processing ComfyUI 0.44 adds controls for video export and model processing

Export settings can fall back

The video nodes separate the image representation from the file container. Create Video offers full-resolution color or subsampling that reduces color detail horizontally, or along both axes. Save Video then chooses the codec and container. Its automatic setting selects MOV for ProRes, WebM for AV1 and MP4 otherwise. ProRes is not offered inside MP4. The underlying encoder checks whether a requested color sampling mode is supported. If it is not, it chooses a supported alternative and logs a warning. Transparency can likewise be discarded, with a warning, when the selected output cannot preserve it. A saved file therefore does not by itself establish that every requested property survived. The implementation selects a ProRes HQ profile for 4:2:2 output and a 4444 profile otherwise.

Windows follow the model’s frame structure

The MiniMax H3 node applies context windows during sampling. Its interface expresses length and overlap in real frames, with both following a seventeen-times-an-integer-plus-five pattern. Window starts align with the model’s seventeen-frame chunks to keep audio and video synchronized. The default window is 124 frames, with 39 frames of overlap. The node uses a static schedule and defaults to pyramid fusion, with FreeNoise noise shuffling enabled. Splitting multiple conditionings across windows is disabled by default. Its implementation converts the frame counts into latent units and caps overlap below the window length. The node inherits an experimental designation. These controls describe how a workflow divides and combines processing; they do not establish a measured memory saving or guarantee that joins will be invisible.

Compute precision stays on its branch

The VAE Compute Dtype node controls the numerical type used for encoding and decoding. It offers default, fp32, fp16 and bf16 choices. Default preserves the incoming VAE setting. For an explicit selection, the node clones the VAE and updates both its compute type and the associated model patcher. Other branches retain their own settings. That separation lets a workflow compare precision choices without changing every consumer of the original VAE. It is different from choosing the bit depth of an encoded video file: one concerns model computation, the other the saved representation. The node lives in the advanced debugging category, making it an explicit intervention rather than an automatic quality upgrade. These implementation changes landed before the versioned release; the announcement does not make each underlying model newly available.