H3 SPEED · Load Exact Previous Stage (T8 EXP)

Here's the scenario this node exists for. You ran stage 0 and stage 1 of a three-stage SPEED clip. Stage 2's prompt was wrong. You do not want to sit through stages 0 and 1 again. MiniMaxH3SPEEDStageLoadEXPT8 starts from a saved stage and rebuilds everything needed to continue, without re-running the previous model, noise generation, conditioning encode or sampler.

It runs nothing. That's the point.

Inputs

• speed_plan - the H3_T8_SPEED_PLAN from the Advanced SPEED Plan node. The loader needs the plan because what comes out is only meaningful relative to the schedule the stages were built from.
• seed (default 2608184001) and shift_audio (default 3.0) - the two things that have to match the run that produced the saved stage for the next stage's spec to be reconstructed honestly.
• next_stage_index (default 1, 1–99) - which stage you're about to run, indexed from that saved point.
• artifact_path - the relative manifest path returned by MiniMaxH3SPEEDStageSaveEXPT8. Use the real value; the example graphs ship with empty placeholders and will not run.
• artifact_sha256 - the digest that came back from the same save. It's mandatory, and it's not decoration: the load compares it, plus the stage's own identity, before handing anything downstream.

Outputs are completed_stage (a T8_SPEED_STAGE_RESULT - feed it to MiniMaxH3SPEEDDCTTransitionEXPT8), previous_spec (a T8_SPEED_STAGE_SPEC - feed it to the next MiniMaxH3SPEEDStageSetupEXPT8 so it can reuse the earlier stage's T2VA text instead of re-encoding from scratch), and report_json.

The cache story, because it will bite you

ComfyUI's default behaviour is aggressive output caching: if nothing upstream changed, a node doesn't re-execute. For a disk-backed loader that's exactly wrong - you could swap the file on disk and ComfyUI would happily hand you the old latent. The pack handles this the way ComfyUI nodes are supposed to: the node implements an input fingerprint that returns NaN when the artefact is missing, corrupt or locked, and NaN != NaN, so the node is always considered dirty and re-reads. That's the same always-rerun idiom covered in the node-plumbing notes, and it's why you'll see this node actually hitting disk on every queue rather than skipping.

The flip side is the semantic you have to internalise: this freezes a past stage, it does not validate your current settings against it. Edit the LOW stage's LoRA, prompt or noise and then load an old stage artefact, and you have a mismatch that the report will flag rather than silently reconcile. The author's line is "not an automatic cache hit against edited past-stage settings," and that's a promise, not a limitation.

Install

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

Full restart of ComfyUI, then refresh the frontend. Manager search: MiniMax H3 Audio T8. There's no dependency install step - the pack's requirements file is empty on purpose so it can never replace ComfyUI's Torch/CUDA build. You do need a recent ComfyUI core (native H3 support) or everything in the pack shows red.

Gotchas

Artefacts live under the ComfyUI output directory in the pack's own SPEED store, and every save creates a new immutable file rather than overwriting - so your output/ folder will accumulate stage snapshots. Prune them knowingly, and note that deleting a file the Cold_* example graphs reference turns them into an error, which is the correct outcome.

Wrong SHA, wrong stage, a path that escapes the store, or a file still being written all fail loudly. If you're seeing a resume that "works" but produces drift, you're almost certainly continuing from a stage that predates a change you made to the earlier stage - rerun that stage rather than fighting the loader. And as always with H3 on a 16 GB card: one job at a time, smaller canvas before you blame the pack.