H3 SPEED · Prepare ONE Stage (T8 EXP)

If you were going to pick one node from the SPEED split family to actually understand, pick this one. MiniMaxH3SPEEDStageSetupEXPT8 rebuilds exactly one planned stage - the AV canvas, the conditioning, and the native-flow Euler sigma schedule for that stage's slice of the resolution ramp - and hands you the pieces to sample it yourself. It runs no sampling and has no hidden fallback.

The reason it shouldn't be confused with the old whole-chain SPEED sampler: this is deliberately one stage, callable twice in a graph with different models, LoRAs, prompts and noise. Both stages can see each other only through the typed spec you pass between them.

Inputs

• model - this stage's diffusion model. Independent branches are the point; you can swap a lighter or differently-LoRA'd model in for a later stage.
• speed_plan - H3_T8_SPEED_PLAN from the Advanced SPEED Plan node. It carries the stage list, per-stage targets and the transition schedule.
• speed_source - H3_T8_SPEED_SOURCE from the Advanced SPEED Stage Source node (the per-stage conditioning source).
• stage_index (default 0, 0–99) - which planned stage to prepare.
• shift_audio (default 3.0) - the audio clock shift for this stage.
• seed (default 2608184001) - the execution seed for this stage.
• execution_scope - strict_t2va_stock20 (default), multimodal_research_exp, or turbo8_t2va_research_exp. This is the honesty dial: it declares which execution contract you're claiming. Keep the strict default unless your plan really is one of the research modes.
• reuse_t2va_text (default true) - when you feed previous_spec in, this reuses the earlier stage's encoded T2VA text rather than rebuilding it. Set it false if you edited a stage's text, otherwise you'll be sampling against the old encoding.
• previous_spec (optional) - the T8_SPEED_STAGE_SPEC from the previous stage's setup (or from MiniMaxH3SPEEDStageLoadEXPT8 when you're cold-resuming).

Outputs

Eleven, and they split into "sample with these" and "keep these for the record": model, positive, av_latent, sampler, sigmas, stage_spec, mux_audio, conditioned_prompt (the resolved text - read it when a stage ignores you), media_map_json, report_json, and stage_context. The first five go into MiniMaxH3SPEEDStageSampleEXPT8; stage_spec goes into the sampler and into MiniMaxH3SPEEDDCTTransitionEXPT8 when you prepare the next stage; stage_context is what the EAV and Relay nodes validate against; mux_audio is the audio side of the joint target.

The install, once

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

Full exit and restart of ComfyUI, then refresh the browser. Manager search term: MiniMax H3 Audio T8 - and if the registry copy is behind the GitHub release, install from GitHub, because the pack publishes to both independently. There is no dependency install; the pack's requirements.txt is intentionally empty so it can never replace ComfyUI's Torch/CUDA build. You do need a recent ComfyUI core with native MiniMax H3 support, or you'll be looking at a canvas full of red.

Models: H3 diffusion model in models/diffusion_models, Qwen3-VL text encoder in models/text_encoders, video and audio VAEs in models/vae.

Gotchas

An empty previous_spec on stage 1+ is allowed but means no text reuse - you'll re-encode, which costs time and can shift the conditioning. Stage indices are 0-based and must line up with the plan's stage count; an out-of-range index fails rather than clamping.

The important expectation-setting: the pack's own tests found the fixed SPEED plan slower and less preferred than a plain full-resolution baseline on their benchmark, and the example docs say not to treat the shipped values as speed or quality recommendations. Use this for stage-level inspection and research, not as your default sampler. If you're on 16 GB, also keep one H3 job at a time - H3 canvas and frame count are the two things that actually decide whether you finish.