H3 HyperFlow · Fresh Route Loader / Original Endpoint (T8 EXP)

HyperFlow is the pack's two-time injection route for MiniMax H3 - a distilled adapter that makes an 8-step AV generation actually work at 8 steps, by patching both the backbone time tensors and a separate endpoint-time projection. It is emphatically not a normal acceleration LoRA, and this loader is where that distinction stops being theory.

Feed it a full native H3 base and a HyperFlow file, get a MODEL you can schedule independently. What it doesn't do is treat the file like any other LoRA, and that's the entire value.

Why a dedicated loader

Two reasons, both practical.

First, image: a generic LoRA loader will happily load a HyperFlow safetensors and patching it through the normal path gives you a model that runs at the right step count and generates plausible mush, because the two-time structure is lost. The pack's Director compiler will even reject a two-time HyperFlow artifact that's been dropped into a content-LoRA slot - including a copy under a different filename - and this loader rejects a directly preceding generic LoRA whose metadata declares hyperflow=true. There's no silent fallback to a single-time approximation; if the file isn't the real thing, or the base is pruned, or the endpoint branch is missing, you get an error.

Second, siblings. The fresh route in this pack runs two stages that each want the HyperFlow owner attached to a different MODEL - LOW and HIGH, possibly with different content LoRAs on each. The old behaviour of "unload the shared model, remove the LoRA, patch the next one" doesn't work when both branches live in the same graph. This loader instead recovers the original base time tensors read-only from ComfyUI's own weight backups, so it can build its endpoint snapshot even while a sibling model is resident. No shared-model unload, no LoRA removal, and the legacy loader keeps its old semantics for old graphs.

Inputs and outputs

model - a full native H3 base, produced by whatever compatibility loader chain you use, including any content LoRAs you want on that branch. Point it at a pruned base and you'll be told no.

hyperflow_file - a combo listing the HyperFlow adapters the node can see. The pack registers its own hyperflow model folder, which resolves to ComfyUI/models/hyperflow/loras, so that's where minimax_h3_hyperflow_8step_v1.0.safetensors goes. The original non-pruned file, from the mirror the docs name, is what's qualified - 2.79GB, 632 tensors, rank/alpha 256, gate 0.25, video/audio shifts 12/3.

Out: model and report_json. Read the report once. It's the thing that tells you whether your content LoRA patched the base time embedder, in which case the loader will note that the original HyperFlow endpoint branch is unchanged and that the stack isn't quality-qualified. That's a warning, not a wall - the author preserves unknown stacks rather than banning them, and just declines to certify them.

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 afterwards. There is no automatic download for HyperFlow weights - no pruned fallback, no substitution - so plan on fetching the adapter yourself and dropping it in models/hyperflow/loras/. You'll also need the H3 base, Qwen3-VL encoder, video/audio VAEs, and for the fresh routes the learned 3D upscaler in models/latent_upscale_models/. The pack's requirements.txt installs nothing on purpose.

Where people get burned

Running this as if it were a generic LoRA node is mistake number one, and it usually shows up later as "HyperFlow looks blurry" rather than as an error.

Mistake two is mixing routes. This loader is for the fresh route - independently scheduled stages where HIGH re-noises and re-runs absolute 4:8 after the learned upscale. It is not the continuous route (raw x_sigma handoff, no new noise) and not P7. The adapters and formulas overlap; the schedules don't.

And the caveat the docs keep repeating, which is worth taking at face value: this is an experiment. The qualified evidence is tiny-model CPU parity and a handful of GPU probe clips at modest resolutions. Nobody has human-reviewed a full film with it, and the author says outright not to sell it as a speed or quality win over your existing 4+4. Treat it as a tool for when you specifically want 8-step H3 output quality you can schedule yourself.