MiniMax H3 Chain Policy

Chain Policy is the part of the Contex Loop pack that keeps you from drowning in per-scene continuity widgets. In 0.4 of this pack you had to reason about continuation_mode, context_length, audio_mode, and audio-context lengths all at once, per scene. 0.5 collapsed that into one small node with five controls, and it's the first thing you wire when you start a production graph.

The pitch: you set how a scene enters (its boundary with the predecessor) and what the final soundtrack is in one place, and the Plan stores the resolved behavior rather than the graph topology. That last part matters more than it sounds - it's why your saved checkpoints stay compatible with a rebuilt or re-laid-out workflow. The policy is a h3_chain_policy_v1 record; Plan expands it into canonical audio and transition records before hashing, so compatibility follows what you chose, not how you wired it.

What the four transition choices mean

incoming_transition is the default boundary for every scene (a scene can still override it in Plan Studio):

• cut - no carried picture. A hard edit.
• guide - 22-frame RGB continuation of the previous scene's motion. The tested everyday default.
• hard_av - a protected 39-frame picture and audio prefix. The previous scene's tail is carried into the next one and shielded from denoising.
• soft_av - the same hard picture prefix, but with a short feathered audio release so the sound hands off less abruptly.

The 22- and 39-frame numbers aren't arbitrary: they line up with H3's own frame counts so the AV clocks meet exactly. That's the kind of detail the old pack made you know by heart, and this node just hides it.

Audio, on three independent axes

• final_audio - what gets muxed into the final MP4: H3's generated sound, the exact source track, or nothing. This doesn't touch generation.
• source_reference - use the matching Source Timeline audio as guidance while H3 generates this scene's sound (influence, not a copy).
• generated_continuity - carry the previous scene's sampled audio latent across the boundary, so sound doesn't restart from nothing each scene. Guide gets 22 frames of it automatically; AV follows the video boundary.

And there's lock_source_audio, the hammer: it encodes each scene's exact Source Timeline window directly into the target audio latent and protects the whole audio mask from denoising. If your dialogue must be word-for-word the source, this is the switch. It automatically overrides source_reference and generated_continuity, because they'd compete with an exact target clock. Pair it with final_audio: source and the final MP4 keeps the original waveform too.

Outputs: chain_policy (wire it to the Plan's chain_policy input) and a status string that tells you the resolved transition, soundtrack, reference, and lock summary.

Install

Part of ComfyUI-MiniMaxH3-Contex-Loop - install the pack once and you get all its nodes. ComfyUI Manager (search "ComfyUI-MiniMaxH3-Contex-Loop") or:

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Contex-Loop.git

then restart. No pip dependencies; you do need a current ComfyUI with the native Add Guide for MiniMax H3 support (ComfyUI PR #15439), and note the models aren't bundled - the MiniMax H3 weights themselves are license-locked out of the US, EU, UK and South Korea.

Gotchas

The easiest mistake is treating lock_source_audio as "make it sound like the source" in a soft way. It's not - it's exact, and it switches the other two audio axes off. If you just want a vibe, use source_reference instead. Also remember the transition describes how a scene consumes its predecessor; editing it after the predecessor is already rendered doesn't change the finished predecessor, and changing resolved behavior of a saved scene is exactly the kind of edit that invalidates a resume. Keep it simple for a first run: guide + generated + continuity on.