MiniMax H3 HyperFlow Loader (Two-Time EXP/T8)

HyperFlow is a two-time adapter for MiniMax H3: the whole point is that a step happens at two different noise times rather than one, which is how you collapse a full trajectory into 8 evaluations. MiniMaxH3HyperFlowLoaderT8Advanced is the node that installs it - and the first thing to understand is that this is not a LoRA loader wearing a costume.

It loads the original 632-tensor HyperFlow artifact, maps every LoRA target, and installs the adapter's separate endpoint-time branch. As the author puts it in the node description: it does not substitute a generic acceleration LoRA, and it does not silently degrade to a single time.

That last clause is the design principle for this entire node family.

Why the strictness matters

Most speed LoRAs are one thing: weight deltas you scale and forget. HyperFlow is two things - 210 backbone/base-time patches plus two independent endpoint-time projections, 52 QKV fusions included, rank/alpha 256, gate 0.25, and separate video and audio shifts (12 and 3) over nine raw sigma points. If the endpoint branch is missing, you haven't got a slightly worse HyperFlow; you've got a different sampler that will produce correctly-shaped garbage.

So the loader validates. It requires a full, non-pruned MiniMax H3 base - 50 DiT blocks plus 2 refiner blocks - and rejects a pruned model instead of falling back. It also rejects a directly preceding generic LoRA whose attached metadata declares hyperflow=true, which is the pack catching the specific mistake of loading the two-time artifact through an ordinary loader.

Using it

Two inputs: model and hyperflow_file. The combo is populated from models/hyperflow/loras - if the list is empty you get a pointed FileNotFoundError telling you to put the original safetensors there. That's a separate discipline from content LoRAs, which stay on their own loaders; the docs repeat the rule that you must never load the HyperFlow original through a generic LoRA loader.

Outputs are model and report_json. The patched MODEL goes onward to a Plan node - the plan is bound to that exact model, so you can't move a plan between two differently-patched models and hope it works.

Where you drop it in the graph is a fresh H3 AV graph: CLIP, VAEs, your native H3 conditioning, and the HyperFlow Plan → Sampler trio feeding a normal SamplerCustomAdvanced. The old H3 sampler and the legacy 4+4 route are untouched by any of this.

Install

ComfyUI 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

Fully restart ComfyUI and refresh the page. The pack's requirements.txt installs nothing extra by design; all the heavy lifting is torch work that ComfyUI already provides. Beyond the usual H3 stack you need the HyperFlow safetensors (~2.8 GB) in models/hyperflow/loras, and the H3 base unpruned in models/diffusion_models.

Licensing, briefly, because it's not the usual story here: the HyperFlow code is Apache-2.0, but the adapter weights report the MiniMax H3 Community License, and that license's grant is void in the US, EU, UK and South Korea - outputs included. The conversion and the node change nothing about that boundary. Check where you are before you build a pipeline on it.

Where it bites

Pruned checkpoint? Stop here. The loader's rejection is a feature. Swap in the full base - the pair of refiner blocks is not optional padding.

Runtime crashes on a two-owner graph. The pack's own notes record a Windows crash inside the native graph compiler when two separately-patched loader models are used in one graph, and that the identical split succeeded on a Core started with --disable-comfy-compiler. Worth knowing before you file it as a node bug - the split routes are where that combination lives.

Patches stack in order. If a content LoRA patches the base time embedder, the loader will tell you the original HyperFlow endpoint branch is unchanged - and that the resulting stack hasn't been quality-qualified. Read the report JSON rather than assuming your stack did what you intended.