H3 HyperFlow P7 · Load Exact LOW for HIGH Only (T8 EXP)

This is the node that makes HIGH-only iteration actually cheap. You ran the full LOW 0:4 pass, saved it, then decided the HIGH prompt or the HIGH LoRA needed another go. Load Exact LOW for HIGH Only reads the frozen LOW artifact back, revalidates it against the current chain, and hands you the typed LOW result - without loading a LOW model, drawing LOW noise, or sampling LOW at all.

It's a revalidation, not a cache

The distinction the pack keeps hammering is worth internalising, because it changes how you use this node. ComfyUI's cache asks "did anything upstream change?" A frozen artifact load asks "do these bytes hash to what I recorded, and does the current chain still agree with the provenance they claim?" Those are different questions, and the second one has a sharp consequence: loading an old LOW does not mean today's LOW settings match it. If you edited the LOW phase's prompt, its model, its effects or its noise, the frozen file doesn't know and doesn't care. It's not going to tell you that today's graph and yesterday's artifact have diverged - the pack documents that as a deliberate limit, not a bug.

What it will do is recheck the accepted parent and the freshly prepared LOW phase. So it catches the case that actually matters: the parent segment changed, or the phase was rebuilt for a different chain, or the geometry moved. That's a fail-closed check, and it's why the node takes a live phase input at all.

Inputs

low_phase - the prepared LOW phase from One Phase Native Conditions with phase set to low. This is the live half of the revalidation, and it's a required input: there's no "load the file and skip the checks" mode. It also means the load node sits downstream of a phase that has to be built correctly first, so a resume graph is a real graph with a real LOW phase in it, not a shortcut.

artifact_path and artifact_sha256 - both required, both from Save Exact LOW 0:4. Wire them from the save node's sockets in the save-and-resume workflow, or paste them; the example graphs ship placeholders with a zeroed hash that fail until you fill them in with real ones.

Outputs

low_result for the learned lift, low_x0 for the LATENT, report_json for the receipt. Downstream, Original Learned 3D Lift can't tell this apart from a freshly sampled LOW - that's the intent.

The workflow it enables

[once]  LOW 0:4 Setup → Stage Sampler → Bind Completed LOW Result → Save Exact LOW 0:4
[again] Prepare LOW phase → Load Exact LOW for HIGH Only → Learned 3D Lift
        → HIGH Audio Reconcile → HIGH 4:8 Setup → Stage Sampler → Bind/Save HIGH

Two things to keep in mind while iterating. The HIGH side still needs its own real conditions - you're skipping LOW, not skipping the segment contract - and if HIGH itself used Prompt Relay, rebuilding the source can still pull a model in, so "load-only" isn't automatically a zero-model graph.

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

Restart ComfyUI fully and refresh. requirements.txt installs nothing by design. This node itself needs no checkpoint loaded - the rest of a P7 graph needs the H3 base, Qwen encoder, both VAEs, the HyperFlow adapter in models/hyperflow/loras/ and the learned 3D upscaler in models/latent_upscale_models.

Failure modes worth knowing

Wrong SHA, moved file, truncated artifact, path outside the chain directory, or a type mismatch between the stored stage and what you're asking for: all hard errors, no silent regeneration. That's the desired behaviour, but it means the error you get is often "this doesn't match" rather than "you picked the wrong file", so check the stage type first. LOW artifacts are stored as hyperflow_low_partial4 in the P7 store under output/MiniMaxH3/modular_p7_stages/<chain_id>/segment_00000/; a HIGH artifact from the same directory is a different type and won't load here.

Last one, and it catches people in the middle of a long session: the chain lease. Saves and loads take an OS lock on the chain directory, so running two graphs against the same chain id at once gives you a busy error rather than corrupted files. Queue one at a time.