MiniMax H3 FastH3 4-Step Setup (Advanced EXP/T8)

FastH3 is the NVIDIA FastVideo project's shot at making MiniMax H3 generate at speed: a DMD2-distilled recipe that collapses the usual many-step denoise into a 4-step joint audio-video sampling contract. Distillation quality claims are a minefield in this space - every Turbo/Lightning-style model loses something, and the KB's distillation doc is full of "4x faster, but..." stories. So it's refreshing that this node doesn't try to sell you on quality at all. It configures the published contract and lets you get on with testing it.

The one rule to obey before anything else: apply the matching FastH3 LoRA to your model first. This node expects a model that already has the distilled LoRA in place - it sets up the sampler, sigmas and sampling contract, it doesn't add the LoRA for you. The description is explicit: "Apply the matching FastH3 LoRA first."

What it actually emits

Three things you'll wire into your sampling path, plus a report:

• model - the model passed through (with the 4-step contract's patch applied if needed).
• sampler - the sampler configured for the contract.
• sigmas - the 4-step schedule for the joint-AV run.
• report_json - tells you what got configured, and whether any attention backend actually engaged.

Two inputs decide the flavor of the run:

• task_family - defaults to t2va_only, which is what the published FastH3 4-step contract covers. There are three more options (t2va_fl2va, t2va_fl2va_legacy_untrained_exp, ref2va_untrained_exp), and the ones with untrained_exp or exp in the name are exactly what they look like: experimental, not validated, don't trust them for a deliverable.
• attention_profile - dense_comfyui is the default and is always usable. external_vsa_if_available is the sparse-attention route (a learned 90%-sparse tile-64 kernel), and here's the catch worth reading twice: the node will only report it as active when a real external FastVideo runtime and kernel are actually installed. If they're not there, it falls back to dense and says so in the report. That's the pack refusing to fake a speed win you didn't earn.

Who should reach for it

People who are already comfortable running H3 and want to benchmark the 4-step route against their 8-step native baseline on the same content. This is a performance experiment, not an onboarding tool - it lives in the 10-speed workflow folder and the "Advanced EXP" suffix is doing real work. Start from the pack's example workflow, keep the same seed and prompt across your A/B, and read report_json before you believe the speed number.

Also keep the pack's standing rule in mind: don't stack multiple sampler-owning or attention-patching nodes on the same model. If you chain this node behind an Enhance-A-Video or Prompt Relay compositor that also owns attention, you're in unsupported territory and the report can't save you.

Install

Standard for this pack:

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

or ComfyUI Manager → "MiniMax H3 Audio T8" → restart. Update ComfyUI itself first (it needs comfy_api.latest and native H3 support), and note the requirements.txt is intentionally empty - torch, torchaudio, numpy, Pillow come from ComfyUI. The FastH3 LoRA itself goes in ComfyUI/models/loras, and no requirements file in this repo will install it for you.