This addon provides a standalone H3 Motion Context implementation plus automatic audio chunking, latent-based clip continuation, per-clip prompts, final-frame references, automatic prompt requeueing, and MP4 video stitching.

It is designed for creators who want to generate a long MiniMax H3 video as a sequence of connected clips while preserving motion, audio, characters, and scene continuity.

Features

Automatically split long audio into sequential MiniMax H3 video clips.
Calculate clip starts and audio cuts on a user-selected frame timeline (24 fps by default).
Preserve the original full audio timeline for gap-free final stitching.
Automatically queue the next clip without manual clip-number changes.
Carry video and audio continuity through H3 latent context.
Save and load retry-safe numbered H3 latent slots.
Use a shared style prompt plus a different prompt for every clip.
Reuse the previous clip's final frame as the next image reference.
Select original-image, last-frame, or disabled reference behavior.
Trim generated clips to their intended duration.
Save individual MP4 clips and stitch them into one final MP4.
Does not require or register the original H3 Motion Context package.

Requirements

This addon requires:

ComfyUI.
A working MiniMax H3 video workflow.
The Python packages used by the H3 workflow, including safetensors, numpy, Pillow, and imageio-ffmpeg.

Installation

Install this folder inside ComfyUI/custom_nodes/:

ComfyUI/custom_nodes/ComfyUI-H3-Motion-Context-Auto-Chain-addon/

Restart ComfyUI after installing the addon.

The original H3 Motion Context package is optional. If it is installed for other workflows, this addon does not register its node IDs; both packages can coexist. Do not install multiple different H3 patch packs that modify the same ComfyUI H3 internals.

Nodes included

H3 Auto Chain Motion Context

Use this node in automatic-chain workflows. It uses the addon’s private Motion Context implementation whenever previous context is available. On clip 1, when no previous latent or context frames exist, it passes the conditioning through and reports trim_frames as 0.

This first-clip pass-through is required because an automatic chain has no previous clip to pin yet.

H3 Auto Chain Audio

Splits the complete audio input into sequential clip-sized chunks and outputs the current clip's audio, prompt, clip number, and internal chain configuration.

Important inputs:

chain_id: unique name for the project.
chunk_seconds: normal model duration for each generated clip; the default is 20 seconds. A shorter final source segment is padded with silence for model sampling rather than shortening the model request.
fps: generated video frame rate; default 24. Clip boundaries and audio sample cuts are calculated from this frame rate.
trim_frames: leading frames removed by H3 Motion Context Trim from continuation clips. Match the Motion Context context length; the default 22 is about one second at 24 FPS. The addon requests this extra audio span so the trimmed clip still occupies its complete timeline slot.
final_tail_mode: choose exact audio duration or audio plus tail for the final partial clip.
final_tail_frames: silent frames retained after the source audio when audio plus tail is selected. The model input is padded with digital silence; no source audio is repeated.
reset: starts a new chain.
style_prompt: text shared by every clip.
clip_prompts: optional numbered prompts such as [1], [2], and [3].
start_clip: clip number to begin with.
end_clip: final clip number; 0 continues until the audio ends.

Connect its chain_config output to the other addon nodes.

The source_audio output is the complete original audio. Connect it to the optional audio input of H3 Auto Chain + Stitch for frame-aligned final audio instead of concatenating separately encoded clip audio.

H3 Auto Chain Load Latent

Loads the previous H3 video/audio latent for the original H3 Motion Context node.

Connect its LATENT output only to:

H3 Auto Chain Motion Context → context_latent

When chain_config is connected, the correct previous clip slot is selected automatically. The first clip correctly starts without a previous latent.

H3 Auto Chain Save Latent

Saves the current H3 sampler output so the next clip can continue from it.

Connect the H3 sampler's output latent to this node. Connect chain_config from H3 Auto Chain Audio to automatically select the correct numbered save slot.

The saved latent is for H3 Motion Context only. Do not connect it to a VAE decode node.

H3 Auto Chain Frame Reference

Provides an image reference for the current clip.

Connect:

chain_config from H3 Auto Chain Audio.
initial_image from the first image used by the workflow.
reference_image to the H3 reference-image input.

Reference modes:

original: use the initial image for every clip.
last_frame: use the initial image for clip 1 and the previous clip's final frame afterward.
off: use the image only for clip 1.

last_frame is the recommended mode for character and scene continuity.

H3 Auto Chain + Stitch

Saves the current generated video, extracts its final frame, queues the next clip, and stitches all completed clips into one MP4 when the chain finishes.

Connect the final generated VIDEO and the chain_config output from H3 Auto Chain Audio. Choose audio_source:

original audio input: use the complete original audio. Connect source_audio from H3 Auto Chain Audio to the stitch node's optional audio input. This is the recommended mode when the final video must match the source timeline.
generated video audio: use the audio carried by each generated video clip. No original audio connection is required.

Use delete_completed_latents to choose whether the chain's saved latent files are removed after a successful final stitch. It is disabled by default, so the files remain available for retry or resume.

Recommended workflow

Complete AUDIO
    ↓
H3 Auto Chain Audio
    ├── audio chunk → H3 video workflow
    ├── prompt → text-conditioning workflow
    └── chain_config → addon chain nodes

H3 Auto Chain Load Latent
    ↓
H3 Motion Context → context_latent

Initial IMAGE
    ↓
H3 Auto Chain Frame Reference
    ↓
H3 reference-image input

Sampler output LATENT
    ↓
H3 Auto Chain Save Latent

Decoded and trimmed VIDEO
    ↓
H3 Auto Chain + Stitch

Use H3 Auto Chain Motion Context for automatic-chain workflows. It uses the addon’s standalone Motion Context implementation for continuation clips and bypasses context pinning only for clip 1. Connect its trim_frames output to H3 Auto Chain Motion Context Trim before passing the final video to the stitch node.

Starting a new video chain

Use these settings for a new project:

chain_id:      a unique project name
chunk_seconds: 20
fps:           24
trim_frames:  22
reset:         True
start_clip:    1
end_clip:      0
reference_mode: last_frame

Queue the workflow once. The addon generates clip 1, saves its latent and final frame, automatically queues clip 2, and continues until the audio ends.

Per-clip prompting

Use style_prompt for text that should remain consistent:

cinematic lighting, realistic character, detailed face, consistent wardrobe

Use clip_prompts for changing actions or locations:

[1] The character walks through a rainy city street.
[2] The character enters a warm cafe.
[3] The character looks out of the cafe window at the night traffic.

The addon combines the shared style prompt with the prompt for the current clip.

Resume a chain

To continue after clips 1 and 2:

Keep the same chain_id.
Set start_clip to 3.
Set reset to True to initialize the resumed chain state.
Confirm that the previous latent files exist when Motion Context continuation is needed.
Queue the workflow.

Missing previous videos no longer prevent a partial render. The requested clip is generated and, when stitching finishes, only the existing clip MP4 files are included in timeline order. This supports rendering clip 3 alone, or rendering clips 2-4 while preserving already-rendered clips 1, 5, and 6 when they exist. Missing previous latents still cause Motion Context to be bypassed for that clip; the addon never reuses a latent from another chain.

Retry a clip safely

The addon stores numbered latent slots, for example:

h3_context/<chain>_clip_00001.safetensors
h3_context/<chain>_clip_00002.safetensors
h3_context/<chain>_clip_00003.safetensors

To retry a clip, use the same chain_id, set start_clip to the clip number, set reset to True, and queue again. The clip's own slot is overwritten, so a rejected render is not accidentally used as the previous context.

Output files

Generated files are stored under the ComfyUI output directory, typically:

output/video/<chain>_clip_001.mp4
output/video/<chain>_clip_001_last.png
output/video/<chain>_clip_002.mp4
output/video/<chain>_clip_002_last.png
output/video/<chain>.mp4

The final file without _clip_### is the stitched MP4.

When delete_completed_latents is enabled, a successful automatic-chain stitch removes that chain's temporary .safetensors files from output/h3_context. They remain available when the option is disabled or the chain fails before final stitching. During a chain, clip N loads clip N-1 and saves clip N; clip 1 starts without a previous latent.

Important limitations

Keep the same resolution

When using context_latent, every clip in a chain must use the same resolution. Start a new chain with a new chain_id if the resolution changes.

Use a unique chain ID

The chain ID controls the internal state, latent filenames, video filenames, and final-frame references. Do not reuse it for unrelated projects.

Do not decode the saved context latent

The saved latent contains H3's paired video/audio representation. It is intended only for this addon’s Auto Chain Motion Context context_latent input.

Avoid competing H3 patch packs

The original H3 Motion Context package is optional and can coexist with this addon. Do not run multiple different H3 patch packs that modify the same ComfyUI H3 internals; the first compatible patch remains active and an incompatible patch is rejected clearly.

Troubleshooting

The addon cannot find a previous latent

Check that:

The chain_id is unchanged.
start_clip is correct.
The previous clip was saved successfully.
The addon Load Latent node, not the stock ComfyUI Load Latent node, is connected.

The chain starts from clip 1 again

Check that H3 Auto Chain Audio is the only chain-audio node in the graph and that the same chain_id is used throughout the workflow.

The final video contains only some clips

This is expected when the chain is sparse. H3 Auto Chain + Stitch includes every matching clip MP4 that exists and skips missing clip numbers. To create a complete timeline, render the missing clips with the same chain_id and stitch again.

The character changes between clips

Try:

Setting reference_mode to last_frame.
Using a stronger shared style prompt.
Keeping the same resolution and model settings.
Connecting the previous latent through H3 Motion Context.

Summary

The H3 Motion Context Auto-Chain addon turns a standard MiniMax H3 ComfyUI workflow into an automatic long-video generation workflow:

H3 Auto Chain Audio splits the audio and manages clip state.
H3 Auto Chain Load Latent loads the previous clip context.
The original H3 Motion Context node preserves motion and sound.
H3 Auto Chain Frame Reference preserves image and character identity.
H3 Auto Chain Save Latent stores the current clip for continuation.
H3 Auto Chain + Stitch queues the next clip and creates the final MP4.

Recommended starting settings are 20-second chunks, 22 context frames, 24 audio context frames, and last_frame reference mode.