MiniMax H3 SPEED Whole-Chain Sampler / 整链采样 (Advanced)

This is the engine room of the SPEED experiment. Where a normal H3 sampler runs the whole denoising at one resolution, MiniMaxH3SPEEDSamplerT8Advanced runs native H3 Euler sampling over the spatial stages your Plan defined - rebuilding the AV layout and conditioning at every canvas, expanding the video between stages in the frequency domain (official DCT expansion), and carrying the audio along on the shared flow. Give it model, speed_plan, and speed_source, and it returns the finished av_latent plus mux_audio, ready for an H3 decode.

What it does, stage by stage

For each resolution in the plan it re-encodes the references, rebuilds the AV latent layout and conditioning, runs its slice of the Euler steps, and DCT-expands the video up to the next canvas. The dct_chunk_size input (64) is the batching knob for that expansion - the number of BCT spatial slices transformed per chunk; if you're VRAM-fragile, lowering it trades a bit of speed for smaller transient peaks.

The audio path is the genuinely H3-specific bit. Audio latents have no spatial dimension, so they don't get DCT-expanded - instead they're "reindexed on the shared public H3 flow," meaning the pack keeps the audio noise stream aligned as the video canvas changes. The author's honesty matters here: that audio extension is experimental until real GPU listening tests pass, so don't assume SPEED's audio quality is a solved problem yet.

The execution_scope you must pick

This combo is the guardrail, and the tooltip is a contract:

• strict_t2va_stock20 (default) - media-free native-audio T2VA, exactly 20 steps, shifts 12/3, unpatched stock H3 model. This is the only scope the author treats as tested territory.
• multimodal_research_exp - unlocks stage-rebuilt I2VA / FL2VA / L2VA / Ref2VA / Hybrid mechanics. Research only.
• turbo8_t2va_research_exp - media-free T2VA at exactly 8 steps with a weight-patched (Turbo) model. Note the tooltip's confession: the node "cannot prove a LoRA file's identity from patch tensors," so it can't verify your Turbo LoRA is the right one.

Outputs worth knowing: av_latent (feed the pack's H3 AV decode), mux_audio (the audio to actually save - the README's repeated warning applies here: don't grab audio from a generic decode when the sampler hands you muxed audio), conditioned_prompt, media_map_json, and report_json.

Installing and the honest bottom line

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; H3 weights on you. This is the most "build it, measure it, be surprised" node in the SPEED set - the README's blind-review tooling and quality-pair scripts exist because the author expects you to A/B against straight sampling rather than trust SPEED blindly. Start with strict_t2va_stock20, keep the manual-sigma plan, and compare a SPEED render against a normal one before you adopt it for anything real.