MiniMaxH3ContextWindows - ComfyUI Built-in Node Documentation

Splits sampling into overlapping context windows so a MiniMax H3 video can be longer than the model’s single-pass window. The windows are placed on the model’s 17-frame chunks, which keeps the generated video and audio aligned across window boundaries. Connect the returned model to the sampler in place of the original model.

Inputs

Parameter	Description	Data Type	Required	Range
`model`	The model to apply context windows to during sampling.	MODEL	Yes	-
`context_length`	The length of the context window in real frames. Must be 17*n + 5 (default: 124).	INT	Yes	22 to 16384 (MAX_RESOLUTION), step 17
`context_overlap`	The overlap of the context window in real frames. Must be 17*n + 5 (default: 39).	INT	Yes	5 or higher, step 17
`fuse_method`	The method used to fuse the context windows (default: `"pyramid"`).	COMBO	Yes	`"pyramid"` `"relative"` `"flat"` `"overlap-linear"`
`freenoise`	Whether to apply FreeNoise noise shuffling, which improves window blending (default: True, advanced parameter).	BOOLEAN	Yes	true false
`split_conds_to_windows`	Whether to split multiple conditionings (created by ConditionCombine) to each window based on region index (default: False, advanced parameter).	BOOLEAN	Yes	true false

Both lengths are given in real frames and converted to latents before sampling: context_length becomes 2 latents at 5 frames or fewer, otherwise ((frames - 5) / 17) * 5 + 2. context_overlap is converted the same way and then capped at context_length minus 5, so the overlap always leaves a full chunk to advance. The node always uses the static schedule with a stride of 1, applies the windows on the latent time dimension, and does not add the causal fix frame, because every window starts on a chunk boundary. split_conds_to_windows expects multiple conditionings created by the ConditionCombine node.

Outputs

Output Name	Description	Data Type
`model`	The model with the applied context window configuration.	MODEL

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): e9e019d0bc870abfe226ef1aafb2607f55c0cc1b297884db35fe98c465cc0d91

​Inputs

​Outputs

Inputs

Outputs