H3 Video Outpaint · 读取已准备缓存 (EXP)

Every other loader in this outpaint family is about the tail of a run - a candidate, a confirmation, a finished selection. MiniMaxH3VideoOutpaintLoadPreparedT8 is about the head. It reconstructs the prepared handle from what's already on disk, so you can rebuild a graph around a run that was encoded hours or days ago without paying for the source, audio and prompt encode again.

Two inputs: plan and run_name. One is the geometry, one is the address. Two outputs: prepared and reload_report.

If you're looking at a save workflow, a re-compose, a re-delivery with different output settings, or just a graph you closed last night, this node is the first one on the canvas. It replaces Prepare, Load Prepared is what you put where Prepare was.

What it does and doesn't do

The description is specific, and the specificity is the useful part: it reads the completed source-video, audio and prompt caches using the same source file, plan parameters and run_name. It does not load CLIP or the VAE, and it does not re-encode the prompt. Missing or corrupted caches raise an error rather than being rebuilt.

That last line is a design choice worth respecting. A loader that silently re-encodes when the cache is missing would produce a working graph with subtly different conditioning, and you'd never know which of your runs used the rebuilt version. Failing hard keeps the run's identity honest - and the identity is the whole point, because the pack binds caches to the source hash, the model, ComfyUI core, KJNodes and the sampling implementation.

The practical upside is speed and VRAM: no CLIP or VAE load, no text encoding, no video decode at this stage. On a recovery graph where you're only re-running the tail, that's the difference between minutes and a re-prepare of a long clip.

Which run name to type

The one it was prepared under. outpaint_01 for the ordinary route, outpaint_guided_01 for the regional route - those are the defaults the shipped workflows use, and the run name is not a free-form label. It's a directory key, combined with the first part of the plan hash to locate output/T8_H3_Outpaint_Cache/<run_name>-<hash16>/. Names are limited to 1–64 characters of letters, digits, underscores and hyphens.

The trap: you must feed it the same plan that produced the cache. Not an equivalent one - a margin two pixels off is a different plan hash and therefore a different directory. If the reload errors, compare the plan_report you're generating now against the report from the original run before you assume anything broke.

Where it fits

Typical recovery graph: source video → Plan (same settings) → Load Prepared → Sample/Continue → Compose. Note that you still need a real Plan node upstream, because Load Prepared takes the plan handle as input rather than a cached copy of it. The plan node is cheap and CPU-only, so that's not a hardship - but it does mean the geometry has to still exist somewhere in your graph or in a saved workflow.

For the regional route, a reloaded prepared carries the regional conditioning provider, so it must go to the Regional Model node, not the plain sampler - the two expect different conditioning, and that mismatch fails rather than degrading quietly.

Install

Manager: search MiniMax H3 Audio T8 and restart ComfyUI fully. Manual:

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

The pack's requirements.txt is intentionally empty - no pip installs, so your ComfyUI Torch/CUDA stack can't be replaced by a node install. This node is a disk read plus validation, so it needs neither the H3 base model nor the text encoder. What it does need is the cache directory in place under your ComfyUI output folder and ffmpeg on PATH if the graph downstream goes on to sample or deliver.

The usual failure

A stale run name, or a plan that drifted. Both produce the same symptom - a missing-identity error that reads scarier than it is. Check the run name matches the original, check the source file is the same one (same hash, not just the same name), and check you haven't updated ComfyUI core, KJNodes or the pack in between. That last one is a legitimate invalidation rather than a bug: the caches verify the execution identity precisely because updating a component mid-run can change results, and the pack's advice is to keep the old cache and old results rather than trying to edit the record to get past the check.