H3 Project Suite manages a long MiniMax H3 video as a chain of short clips, with reviewable takes and saved context between them. Its distinguishing feature is a latent handoff: the next clip receives model-space video and sound context instead of only the previous clip's last decoded picture. The project manager decides which take becomes the foundation for the next generation.
This article explains the maintainer's documentation, reviewed on October 13, 2026. We have not installed the pack or generated a chain. It is a local ComfyUI extension, not a new MiniMax base model or an H3Vid online preset. The registry dates version 1.5.1 to September 26; renewed discussion is not a new release date.
What latent chaining changes
A final-frame handoff preserves a picture, but loses the sequence that led into it. A moving camera or spoken phrase has more context than one still can express. The suite carries a short overlap from the approved clip into the next generation, including audio, so the model has a sequence to continue.
The default video-source setting uses saved latents. That avoids decoding and then re-encoding the context at each handoff. It does not eliminate generative drift, guarantee a perfect join, or make a chain lossless. The model can still change exposure, texture, motion or voice. Review those changes in the actual export, especially after several extensions.
A chain is also not one unlimited-duration inference request. Each continuation consumes its own sampling time, and each saved take consumes disk space. The practical limit depends on your machine, project size and willingness to review accumulating changes.
The nodes have separate jobs
| Node | Role in the documented graph |
|---|---|
| H3 Project Hub | Selects a project and opens its review, take and export controls |
| H3 Context | Supplies continuation conditioning and the sampler's starting latent |
| H3 Context Trim | Removes the overlap from the newly generated clip |
| H3 Project Save | Stores the clip, context and production record |
The Hub sends its context to H3 Context, its chain-active state to the context logic, and its project identity to Save. Context conditioning goes to the guider. The sampled result is decoded, trimmed and saved.
One connection deserves special attention: the sampler's latent_image should come from H3 Context's latent output when using seed head. If it still comes directly from the ordinary MiniMax node, the documentation says seed head has no effect. A visible toggle does not prove the intended starting latent reaches the sampler.
A documentation-based starting workflow
Begin with a working local MiniMax H3 graph. The ComfyUI overview covers the base environment; this pack does not supply the model weights.
- Use the maintainer's example workflow and installation instructions. Restart ComfyUI and refresh its frontend after adding the pack.
- Set a project name and choose the resolution before the first clip. The chain keeps that size; a different aspect ratio belongs in another project.
- Check the Context-to-sampler latent connection and the Trim-to-Save path. Keep the default latent video source and matching 24 fps while establishing a baseline.
- Queue the first clip, open the project panel and review it. Approve the take you want to continue.
- Change the prompt for the next action and queue again. Inspect the transition as well as the new clip's ending.
- Export a short chain before extending it further. Save the graph and settings with the project.
Approval is an execution gate. Queueing before approving generates another take of the same clip; it does not advance the chain. Auto-approve removes that review gate for the project, so decide deliberately whether an unreviewed result should become the next clip's context.
The README's installation section warns about the original Motion Context pack, while its newer compatibility notes discuss coexistence with related descendants. Check your exact installed pack and the startup diagnostics. Do not wire two context managers into one chain or assume every similarly named fork is compatible.
Overlap trades new footage for continuity
The documented context-length choices are 1, 5, 22, 39 and 56 frames, with 22 as the default. A larger overlap gives the next clip more history but reduces newly contributed footage. The overlap must be trimmed once; retaining it in both clips creates repeated action or sound at the join.
Seed head places carried content into the starting latent. The head-hold control adjusts how firmly that content is retained. The experimental hold-framing option attempts to continue the shot after its overlap ends, but the maintainer notes possible ghosting. Try a prompt that clearly continues the same shot before treating a stronger hold as an automatic fix.
For a useful first scene, choose one continuous action with a fixed environment. An intentional cut, sudden lighting change or new speaker makes it harder to tell a legitimate creative change from unintended drift.
Audio continuation and audio identity are different
Continuing the previous tail can preserve the flow of a soundtrack or conversation. It can also compound changes if each new clip imitates the last clip's already altered voice.
Version 1.5.0 adds an experimental anchor-latent connection from the Hub. It uses the first clip's saved sound as a reference for later clips, with the first ten seconds used. It does not require an external voice recording, and the prompt cannot refer to that anchor by a tag. Connect it only as the documented reference path, while preserving the normal tail continuation.
Measure drift reports picture changes and audio measures including loudness, spectral brightness, high-frequency energy and quiet-region noise. These are diagnostics, not a voice-identity score. A deliberately louder scene changes the numbers for a valid reason. Compare a continuous scene and listen to the join before deciding the anchor helped.
Imported footage is conformed to the project
The Hub's video and audio VAE connections let the importer identify encoders. The project resolution supplies the target size. Imported footage is converted to 24 fps by selecting frames and trimmed to a valid H3 window; the filmstrip shows the retained range.
Inspect that range rather than assuming the entire source survives. When dimensions differ, choose a crop or a fit with bars. Audio accompanies a source that has it. The import enters as a pending first clip and needs review before it becomes continuation context.
Preview and final export have different costs
The fast export path uses saved video clips. The from-latents option rebuilds the master from saved latent files and encodes the delivered frames together. The maintainer describes it as slower but potentially cleaner at joins. Per-join level matching affects the exported master, while leaving the saved clips unchanged.
Projects retain their history, clips, latents and per-take records under ComfyUI's output/h3_projects directory. Keep those files when moving a project; the exported MP4 alone is not the resumable chain. Alternate takes can be moved to project trash, while purging removes them permanently.
For a hosted source-video continuation, see MiniMax H3 Extender. It addresses extending a playable video through an online task. It does not expose this local pack's saved latent projects, take branching or audio-anchor controls. Choose the destination according to whether you need an online extension or a resumable local chain.
The cover is an AI-generated concept illustration of connected clips and sound. It is not a screenshot, a tested workflow, or a MiniMax H3 output.