H3 Continuation · Select Accepted Parent (T8 EXP)

Long-video generation has one failure mode that wastes more GPU hours than anything else: you continue from a clip that is not the clip you think it is. You re-rendered segment 3, swapped the parent file, moved the project, or edited the first half - and now the new segment keys off a stale parent, so the seam is garbage and the character has quietly changed. Every "why does my 20-second video look like two videos" post traces back to this.

This node exists to make that impossible. It is the trust anchor at the head of the continuation chain.

What it does

You give it an identity - a chain name, a segment index, a parent candidate id, a revision - and it goes and finds that exact parent in the output chain. Then it does the unglamorous part: it checks the revision, checks the candidate record, hashes the actual MP4 and context bytes on disk, and compares them against the previous job's SHA you provided. If any of that disagrees, it errors instead of guessing.

What it does not do is equally important. No VAE, no sampling, no chain creation, no auto-accept, no composition. The author's description says it outright: it validates and hands you a typed source object.

The inputs that matter

Three of these decide whether anything downstream runs at all:

• chain_id - the namespace that ties this segment to its parent, default h3_progressive_long_video. Change models, LoRAs, source audio, prompts or duration and the pack's docs say the same thing repeatedly: start a new chain id. Dragging an old cache into a new experiment is how you get results you cannot reproduce.
• segment_index and parent_candidate_id / parent_revision - which parent in that chain. Revision matters: if you saved a second take of the parent, revision 1 is not it.
• previous_job_sha256 - authentication of the previous job. Blank works, but filling it in is what actually catches a swapped parent file.

The rest is geometry: context_frames (5, 22 or 39 - 22 is the usual pick), and the canvas fields width, height plus low_width, low_height for the reduced pass. 896×448 with a 448×224 low canvas is the reference 0.4MP setup. Stored capacity for 39 frames of tail context does not mean you have to use 39.

What comes out

accepted_source feeds "Prepare Accepted Contexts" alongside a video VAE, which decodes that parent's tail frames into the motion guides this segment continues from. The width, height, low_width, low_height integers are pass-throughs so you can wire them into downstream nodes without duplicating the numbers by hand. report_json is the receipt: saved project revision, original file SHA, adopted range, actual clocks.

The honest caveat, in the author's own words

previous_job_sha256 authenticates the selected previous job only. It is not proof that the stage models and prompts you edited this morning are the same job that produced that parent. The guard proves lineage, not identity of today's recipe. Read it that way and this node is exactly what you want; read it as a quality guarantee and you will be confused when a parent that passes validation still continues badly under a different LoRA.

Installing it

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

Quit ComfyUI completely and restart. Manager users can search "MiniMax H3 Audio T8". The pack's requirements.txt installs nothing on purpose - no pip step, no chance of this pack replacing ComfyUI's Torch/CUDA stack, which is more than most video node packs can say. You need a recent ComfyUI Core with native H3 support; on an older Core these nodes will not register properly.

When it throws

Most failures here are the node doing its job. A ValueError about revision, candidate or the previous job SHA means the parent moved, was re-rendered, or your sha belongs to a different take - reselect, or start a fresh chain_id if you have genuinely changed the recipe. Missing MP4 or context bytes means the output root moved or was cleaned; the node reads real files, it does not keep a private copy.

One confusion worth heading off: this is the native accepted-parent route. Continuing from an externally produced video is a different entry point in this pack, and it deliberately refuses to fake a parent latent or seed. Don't try to route an outside clip through here to dodge the ancestor check - the pack documents that as unsupported, and it will fail rather than approximate.