MiniMax H3 SPEED Plan / 渐进分辨率计划 (Advanced)

The big idea behind SPEED is that you don't have to pay for full-resolution sampling the whole way down. Generate at a small canvas first, then progressively sample the rest of the denoising at bigger and bigger resolutions, carrying the structure along. Done right it looks near-identical to straight full-res sampling while skipping a big chunk of the compute. MiniMaxH3SPEEDPlanT8Advanced is the node that turns that idea into an actual, validated schedule for H3 - a H3_T8_SPEED_PLAN that the SPEED sampler will execute stage by stage.

What it builds

Give it a final width/height (defaults 1056×608, snapped to the 32-pixel grid H3 demands), the total steps, and a scales string - strictly increasing spatial scales ending in 1.0, like the default 0.5,1.0 (so: first pass at half resolution, finish at full). It then computes the per-stage schedule using the official SPEED equations - the kappa rescale and the sigma-alignment relation - plus your shift_video (12 is the H3 stock value). The plan also pins the frequency-domain transform (dct) that the sampler will use to expand the video between stages.

The honest part is the transition knobs:

• transition_mode - manual_sigmas (default) or delta_optimal. Manual is the safe default, and the description says why: "until an H3 dataset spectrum is validated." SPEED's automated transition depends on a spectral model of the target model's latents, and nobody has a vetted one for H3 yet. So you start by typing a manual_transition_sigmas value (one descending H3 video sigma per transition, default 0.85).
• profile_policy - require_validated_profile vs allow_research_profile. Default refuses to pretend a research probe is the real thing.
• fallback_policy - error (default) or full_resolution_passthrough. If the plan can't be built, do you want a loud failure or a silent "just render at full res"? Default says fail loudly.
• delta (0.01) - the delta-optimal math's tuning parameter; leave it alone until you know what you're doing.

Outputs: speed_plan (into the sampler) and report_json (the per-stage schedule to sanity-check). NFE is preserved exactly across stages - the same number of total sampling steps as a straight run, just distributed across resolutions.

The honest caveats

This is a research-grade node in an experimental pack, and the author's own docs hammer two points: the automated delta-optimal path needs a validated H3 dataset spectrum (which is what the Spectrum Harvester → Accumulate → Finalize chain exists to build), and the audio side of SPEED is an H3-specific extension that remains experimental until real GPU listening tests pass. Manual sigmas + stock 20-step T2VA is the tested territory; multimodal and Turbo8 modes are research only.

Installing and using it

Standard pack install: ComfyUI Manager → "MiniMax H3 Audio T8", or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into custom_nodes, restart. No pip deps; you supply the H3 weights. Wire Plan → Source → Sampler, keep transition_mode on manual, and accept that your first SPEED run is an experiment - check the report_json before trusting the output.