H3 HyperFlow P7 · Select Existing Accepted Parent (T8 EXP)

Long-form video with a dialogue-capable model is the thing everyone wants and almost nobody has a clean story for. The P7 route in this pack is that story for H3, and it has a rule baked into its bones that most long-video pipelines only pretend to follow: the next segment starts from the take you accepted, not from the take that was rendered.

This node is the "start from it" half. It reads an existing accepted segment out of a chain, verifies it - revision, candidate ID, the actual delivered MP4, LOW and HIGH contexts, completed audio, and the previous execution contract - and hands you a typed accepted_parent plus the geometry that segment ran at.

What P7 is, briefly

A total 8-second, 192-frame deliverable split into two segments: segment 0 renders 124 frames, segment 1 renders its own 124 frames but only delivers the 68 new ones, for a seam around 5.17 seconds. Each segment runs the HyperFlow fresh recipe - LOW absolute 0:4, learned 3D latent upscale, HIGH absolute 4:8 with fresh noise - and segment 1 builds its LOW video conditions from the last 39 decoded frames of segment 0's accepted MP4, re-encoded through the existing resize and VAE path. HIGH keeps the original high-resolution context, and the audio carries over from the completed HIGH rather than being regenerated.

That last-39-frames detail is why "accepted" has to mean something. If the pipeline chains off the last rendered take while you were looking at a different candidate, motion continuity breaks in a way no amount of colour matching fixes.

Inputs

You supply the chain identity and the parent's authentication: chain_id (default h3_hyperflow_long_video_exp), segment_index (default 1), parent_candidate_id, parent_revision (default 1), previous_job_sha256. All of those except the chain ID come from the accept node's outputs, and the docs are explicit that the previous job SHA authenticates that parent - it does not say anything about today's edited models or prompts.

Then the geometry: context_frames (5, 22 or 39; default 22), width and height (defaults 896x448), and low_width/low_height (defaults 448x224). Keep these consistent with the chain you're continuing - they're what the segment was built at, and they go back out through the width/height/low_width/low_height INT outputs so you can wire them into downstream nodes instead of retyping numbers and hoping.

Out: accepted_parent, report_json, and those four geometry integers.

What it deliberately doesn't do

No sampling. No chain creation. No acceptance - that's the other node's job, and it's a separate explicit act. And no implicit reuse of the old P7 stage cache: reading a parent is not the same as inheriting its cached tensors, and the pack keeps those separate on purpose.

Start a brand-new chain and this node has nothing to read; a new chain begins from the P7 initial-segment node on an empty chain, and only then do accepted parents exist. Use a unique chain ID for a new film - reusing an old chain name is how you get a resume against last week's settings.

Install

Manager → 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

Full restart. This route wants the H3 base model, Qwen3-VL encoder, both VAEs, the original HyperFlow adapter in models/hyperflow/loras/, and the learned 3D latent upscaler in models/latent_upscale_models/. Recent core required; no pip installs from the pack. On Windows, the docs report the two-loader HyperFlow setup needing an isolated ComfyUI started with --disable-comfy-compiler to dodge a native compile crash - if you see a memory-compile failure rather than a Python traceback, try that.

Where people get burned

Copying the wrong IDs. parent_candidate_id and parent_revision and previous_job_sha256 all matter, and the pack treats a mismatch as a refusal. Don't carry them across chains.

Assuming acceptance is implied. It isn't. A candidate that was rendered and previewed but never explicitly accepted is not a valid parent, and the node will say so.

And the scope nobody should gloss over: the qualified run reported here is one T2VA/native 8-second two-segment clip at 896x448 with no active content LoRA. That's evidence the chain works, not evidence about your multi-LoRA, multi-reference, 0.6MP film. Note also that this pack's own docs say the whole-film, listen-to-it-humanly review is still outstanding - which is a genuinely refreshing thing to read in a model-card ecosystem, and a fair expectation to carry into your own testing.