A repeating H3 clip needs its final movement and sound to lead naturally into its beginning. The community ComfyUI-MiniMaxH3-Loop pack addresses that join during sampling and decoding. It is useful for animated backgrounds, gentle character motion, and recurring environmental scenes, rather than simply fading an ordinary video's last frame over its first.
This guide reviews the author's repository, prompting guide, and supplied workflow descriptions on October 23, 2026. A Comfy Cloud demonstration on October 22 provided the discovery signal. We have not run the graphs; performance and quality observations below are the author's, and H3Vid does not execute this patch.
Install the loop components together
Clone pxl-pshr/ComfyUI-MiniMaxH3-Loop into ComfyUI/custom_nodes and restart. The current README requires ComfyUI v0.42.0 or newer. Import a supplied template rather than adding only the model patch to an unrelated graph: the paired decode is part of the documented result.
The main path is model patch, guider and scheduler, sampler, Loop Decode, optional Loop Interpolate, then video creation. Loop Interpolate needs the separate ComfyUI-Frame-Interpolation pack. The Control workflow also uses VideoHelperSuite to load a driving clip. Confirm dependencies before investigating a missing-node error as a model problem.
Choose the template around the actual checkpoint:
| Template | Generation path | Starting steps |
|---|---|---|
MiniMax-H3-Loop.json | Ref2VA without acceleration LoRA | 25 |
MiniMax-H3-Loop-Turbo.json | Ref2VA plus its Turbo LoRA | 8 |
MiniMax-H3-Loop-FastH3.json | FastH3 OmniRef | 8 |
MiniMax-H3-Loop-Control.json | Ref2VA, Turbo, Fun ControlNet | 8 |
The FastH3 model card and H3 base files supply different diffusion weights. Keep the encoder and both VAEs selected as the template expects. A shared eight-step label does not make every acceleration checkpoint interchangeable.
Describe a continuing state
Write a scene that can already be happening at the start and can continue after the end. A lantern gently rocking in a steady breeze is easier to repeat than a person entering, picking it up, and leaving. Avoid a timeline of one-way events when the desired result is a cycle.
An original, untested brief could describe a locked camera, a suspended paper lantern slowly swaying above a still wooden table, fixed warm lighting, and continuous quiet room tone. Keep the environment stable. If water accumulates or footprints appear, their disappearance at the join can reveal the reset even when the motion itself is smooth.
Use reference conditioning to guide appearance. The author recommends a reference picture over first/last-frame anchors for image-based loops: a pinned frame can be interpreted as a shot start, while the ending motion may not approach it naturally. Anchors are supported in the documented graph, but support is different from a guarantee of an invisible join.
Calculate the delivered cycle
The graph's input length follows H3's 17k+5 frame pattern. The delivered loop contains 17k frames. That difference matters when fitting a background to music or choosing interpolation settings.
| Input length | Delivered frames | Runtime at 24 fps |
|---|---|---|
| 107 | 102 | 4.25 seconds |
| 158 | 153 | 6.375 seconds |
| 209 | 204 | 8.5 seconds |
| 260 | 255 | 10.625 seconds |
H3 generates at 24 fps. Increasing the playback fps alone shortens the clip and speeds its motion. Loop Interpolate instead creates intermediate frames across the entire cycle, including the final-to-first transition, and supplies the corresponding fps for video creation.
Wrapped context also increases the sequence seen by the model. With the default padding, an input length of 362 expands beyond the documented 362-frame model limit. Start with the supplied shorter examples instead of assuming the largest ordinary H3 length remains suitable after adding context.
Keep sound and ControlNet settings compatible
The default two-view option evaluates a second, half-rotated view to improve audio near the wrap point. The README reports approximately 1.8 times the sampling cost. It combines audio predictions according to their distance from each view's seam; it does not average two independent visual performances.
The Fun ControlNet route uses one view and requires a specific order: Loop first, then Apply MiniMax H3 Fun ControlNet. A pre-existing repeating depth or geometry clip can anchor motion while H3 changes the appearance. Feed the control preparation node the same length and padding values as the sampling path.
For an exact control fit, the source cycle should divide the delivered loop frames. For example, a 34-frame source cycle fits the 102-frame loop three times. That is a timing calculation, not evidence that every source clip will preserve geometry perfectly after generation.
Inspect several complete repetitions
Watch the exported file repeat several times at its intended speed and listen around the join. Check exposure, small details, camera drift, and a sudden audio burst. Try another seed before assuming every isolated visual pop comes from the wrapping implementation.
The author favors the unaccelerated 25-step path for faster movement and Turbo or FastH3 for drafts and slower subjects. Their timing examples use an RTX 4090 and specified dimensions; they are not a speed promise for other GPUs. For general local preparation, use the H3 ComfyUI guide. For multi-scene continuation rather than repeating one cycle, see Continuum.