MiniMax H3 Chain Recovered AV

Recovered AV is the node that finishes the "de-rope" pipeline, and it's the most specialized thing in this pack's upscale half. De-rope is the term for correcting the temporal smearing that certain learned video upscalers introduce - the bundled chain-aware de-rope variant combines LBH 3D with MAINodes in a single second pass, and the result comes off the upscaler on a different timeline than your source scene. This node returns that result to the source scene's world clock.

To get why that matters, picture the data flow: an upscaled scene is dilated on its own time axis during the de-rope pass. Before it can be saved as a scene, checkpointed, or handed to a following Drift-Control scene, it has to be back on the original RAW clock - the exact frame timing the rest of the chain uses. Feed this node a still-dilated latent and it rejects it rather than producing a scene that silently disagrees with everything around it. That's the discipline: a wrong-timing save is worse than a failed one, because the failure is visible at the node while the wrong timing would poison the chain invisibly.

What it needs

• state - the current Upscale state whose RAW clock the recovered latent must match.
• video_latent - the H3 video VAE encode of the frames after H3 Exact Recover, back on the original RAW clock. (So you've already run the recovery on the picture side before this node sees it.)

Optional audio_latent - the H3 audio VAE encode of H3 Audio Recover. Connect it when you want save_latent to retain a complete recovered AV checkpoint rather than a video-only one. In the de-rope bundle, the recovered audio is optionally re-joined here and returned to the source clock before the checkpoint write.

What it gives back

• latent - the recovered single-video or joint AV latent, suitable for Upscale Segment Save and Loop End. This is the cleaned, re-clocked result that becomes the next scene's HQ context.
• status - a verified description of the recovered RAW clock and packed stream layout, so you can confirm the join before it hits disk.

The design intent, per the pack docs, is to "restore the prior HQ Drift-Control tail" and "return recovered video/audio to the source clock before checkpointing" - which is precisely what lets a following Drift-Control scene or an interrupted child run resume from a checkpoint whose frame contract matches the rest of the production.

Install

From ComfyUI-MiniMaxH3-Contex-Loop. ComfyUI Manager → search the pack, or:

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

Restart. No pip deps - but note the de-rope recipe this node completes expects the external ComfyUI-MAINodes pack (for LBH 3D) installed separately, per the bundled "Deferred Upscale + De-Rope" workflow. Current ComfyUI with native Add Guide for MiniMax H3 expected; H3 weights not bundled (geofenced out of the US/EU/UK/South Korea).

Gotchas

This node is emphatically not a general-purpose upscale node - it verifies clock and stream layout and rejects what it can't trust. Feed it a latent still on the dilated timeline and it refuses; that's a feature. And the audio re-join is opt-in: if you skip audio_latent, the recovered checkpoint is video-only, which is fine for a video-only recipe and wrong for one where the final assembly expects joint AV. For most people this node only ever appears inside the maintained de-rope example workflow - don't go hunting for it in a plain upscale build.