H3 HyperFlow · Load Exact Frozen Fresh Stage (T8 EXP)

A two-stage H3 run is a real chunk of time. If you crash in the HIGH half, or you just want to iterate on the second half's prompt without paying for the first half again, you don't want to re-run the LOW stage to get back to where you were.

This node loads a previously saved stage artifact - path and exact SHA-256, both mandatory - and gives you back the latents and typed context you need to continue. No sampling, no model loading, no GPU work at all. That last part is the selling point: reading a frozen stage shouldn't pull a 30GB checkpoint into VRAM just so you can decode it.

What you get, and what you do with it

Outputs are output, denoised_output, stage_context, stage_result and report_json. Which ones you use depends on what you froze.

A frozen LOW full8 stage: take output, run it through the learned 3D upscaler and the external reconcile, build HIGH conditions at the upscaler's real output size, then execute an independent HIGH stage with its own fresh noise. You keep the full 12-evaluation recipe minus the first eight evaluations.

A frozen LOW partial4 stage: the clean picture is denoised_output, not output - same socket rule as everywhere else in this pack - and it's what feeds the upscaler.

A frozen HIGH stage: you can decode it directly. No sampler, no encoder, no upscale. This is the case people underuse, because a finished HIGH is a finished clip; the pack's docs note a completed HIGH can be decoded with no sampling at all, which is a cheap way to re-export at a different container or bitrate without running the model.

expected_stage is the widget that keeps you honest. It defaults to hyperflow_low_full8 and you set it to match the artifact you're actually loading. Mismatch produces an error, not a guess.

It is an explicit freeze, not a cache

Worth repeating because the whole pack hammers it: loading stage 5 does not mean ComfyUI decided stage 5 was already done. You pointed at a file. The artifact records what was executed, and the node will not claim today's edited LOW prompt, LOW LoRA or LOW settings would reproduce it. If you want to change LOW content, run LOW again and save a new artifact. The docs phrase it as: explicit freeze of the selected first-pass result, not automatic cache lookup.

Practically, that also means placeholders. The shipped resume examples use a zeroed 64-bit SHA you must replace with your real save-node output, and running them as-is will fail - by design, because the alternative is silently sampling something you didn't intend.

Storage lives under output/MiniMaxH3/stage_artifacts, one directory per save, safetensors only, no pickle, with an OS lock and a manifest committed last. Never overwrite: a new save is a new directory.

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 needs the H3 base model, the Qwen3-VL text encoder, video and audio VAEs, the original HyperFlow adapter in models/hyperflow/loras/, and the learned latent upscaler in models/latent_upscale_models/. Newer ComfyUI core required. The pack installs no pip packages by design, so it can't break your Torch/CUDA stack on update.

Where people get burned

The path. The pack is strict about it: missing files, wrong SHA, out-of-range paths, partial writes, type mismatches, corrupt tensors and lock contention all fail loudly, and the node does not go looking for a plausible substitute. That's a feature when you're resuming a client job and a nuisance when you just mistyped a directory.

Also, don't expect it to sidestep HIGH-stage reality. Resuming a frozen LOW still means the HIGH stage runs from scratch - model, content LoRAs, conditioning, fresh noise, effects, all independent and all yours to configure. You saved the first half, not the hard part.

And the standing caveat: the pack's qualified evidence here is tiny-model CPU runs, cross-process resume checks, and modest GPU probes. The load path is solid engineering; the picture you get at full resolution with your own LoRA stack is still up to you to review.