H3 Continuation · Prepare Accepted Contexts (T8 EXP)

Continuing a clip is the hard version of video generation. You have an accepted 5-second segment and you want the next one to carry on from it - same character, same camera, same room tone - instead of starting a fresh world that happens to cut well.

This pack's continuation family is built around an explicit chain of accepted parents. You don't loop automatically and you don't let a node decide what counts as good; you point the graph at a parent segment you already accepted, and the chain re-verifies that parent's actual bytes before using it. H3 Continuation · Prepare Accepted Contexts is the second step in that chain, and it's deliberately the boring one.

What it does

It takes accepted_source (the typed handoff from the accepted-parent selection node, which is where you supply the chain id, the parent candidate/revision and the previous job SHA) and video_vae, and it decodes the last 39 RGB frames of the accepted MP4, resizes them the existing way, and encodes them with your current video VAE. Out come contexts, a T8_CONTINUATION_STAGE_CONTEXTS token, and report_json.

That's the whole job - and the node description is careful to list what it does not touch: the HIGH completed AV and the shared audio pass through unchanged. There's no HIGH prompt, no model, no sampling, no accept, no timeline update. Neutral preparation means if this node is misbehaving you can see it immediately, instead of debugging a failure that could be anywhere in a two-phase graph.

Two things worth knowing about the frames:

• 39 frames is capacity, not a setting. How much context actually gets consumed is chosen upstream on the accepted-parent node (5, 22 or 39). The pack's docs are explicit that saving/decoding a 39-frame tail does not mean you must select 39 - the smaller options exist because context costs memory and can pull the new segment's motion around.
• The LOW phase gets motion guides from these frames; the HIGH phase does not. In this contract, only HIGH locks the completed video prefix. That asymmetry is the point: LOW is cheap and exploratory, HIGH inherits the continuity contract.

Where it sits in the chain

accepted parent (Source) → this node → Geometry + Stage Plan → ONE Phase Conditions
                                        (LOW and HIGH, separately) → LOW stage
                                        → learned 3D lift → HIGH handoff → HIGH stage
                                        → Deliver Completed Window

Note the 8-step native schedule split as LOW4/HIGH4, preserving the 22/39-frame context clock. This is not FastH3 V2, not HyperFlow and not VDN - the pack says so in as many words, and mixing those recipes will get you a mismatch rather than a mixture. Both phases must keep matching source audio, which is why the accepted-parent selection carries audio policy with it.

Install

Manager → search MiniMax H3 Audio T8, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Fully quit and restart ComfyUI, then refresh the browser tab. A recent ComfyUI with native H3 support is required, plus the H3 main model, Qwen text encoder and both VAEs (models/diffusion_models, models/text_encoders, models/vae). No extra pip packages: the pack keeps requirements.txt empty so installing it can't replace ComfyUI's Torch/CUDA. Reference graphs live in examples/workflows/45-progressive-continuation-split - 80 of them, covering 22/39-frame context, plain/EAV/Relay/both effect configurations, and full/cold/load-completed variants.

Where it goes wrong

Placeholder values in the example graphs. The shipped parent IDs, paths and SHAs are placeholders so you can read the wiring; they cannot be queued as-is. Fill in your real accepted parent first.

A parent that changed. If the accepted project, revision, trim or file bytes moved since you built the graph, the source identity is invalid and you should re-select explicitly. That's the guard doing its job - it stops you silently continuing from a clip that isn't the one you reviewed.

Expecting LOW to hold the look. LOW motion guides give you movement continuity; they aren't a pixel lock on identity. Style and identity continuity is a human-judgement problem the pack keeps flagging rather than promising to solve.

Duplicate pack installs. Missing nodes or mismatched sockets after an update usually mean an old copy of the pack is shadowing the real one in custom_nodes. Rename it with a .disabled suffix (a leading underscore doesn't disable a pack), restart, and check python_module in /object_info.