MiniMax H3 Dual-Clock AYS Schedule Contract (Advanced EXP)
An AYS schedule that knows it isn't an image-model AYS schedule
- model
- av_latent
- model
- sampler
- sigmas
- report_json
If you came from SDXL, "AYS" means Align Your Steps: a hand-tuned set of noise schedules that let you drop from the usual 20-30 steps to around 8-10 without the quality cliff, because the schedule spends its steps where they count instead of spreading them evenly. The temptation, when a new video model arrives, is to grab an AYS array that worked for SDXL and jam it in. This node is the pack's very deliberate answer to that temptation: no.
The full name is a contract, not marketing. This is the Dual-Clock AYS Schedule Contract, and its own description states the rule: it builds the native H3 schedule or imports H3-calibrated base-flow knots, then maps those same knots onto separate video and audio clocks. And critically - "It does not claim that image-model AYS schedules are valid for MiniMax H3." The pack's README is even more direct: there is no directly-applicable official H3 optimal schedule, so the default stays native flow. What this node gives you is a way to experiment with schedules on a model whose video and audio run on different shift values, without fooling yourself that you've found the official table.
Why two clocks?
This is the H3-specific part and it's the reason the node exists at all. H3 denoises video and audio jointly, but they want different sigma-shift settings - the pack's standard is video shift 12, audio shift 3. A single image-style schedule can't represent both. This node builds one base-flow knot set and then projects it onto the video clock and the audio clock separately, so both streams stay consistent with each other while getting their own pacing.
Inputs that matter:
steps(default 8) - the count, if you're using the built-in native schedule.shift_video/shift_audio(defaults 12 / 3) - the two clock shifts.schedule_profile-native_flow_baseline(the honest default) ormanual_h3_calibrated, which unlocksmanual_base_sigmas: asteps+1comma/space-separated list of strictly descending sigmas from 1.0 to 0.0. The default string[1.0, 0.875, 0.75, ... 0.0]is just a placeholder example, and the tooltip is careful to say a label is "not validation evidence."schedule_label- provenance for the report. Nothing more.
Outputs are the patched model, a sampler (default dual_clock_euler), the sigmas, and report_json. Wire the sampler and sigmas into your sampling node and keep the model.
The honest use case
This is a tool for people doing systematic schedule exploration on H3, not a speed hack to copy-paste from the SDXL world. Start with native_flow_baseline, get a reference result, then try a manually-calibrated knot set and A/B it. The report_json will tell you exactly which base knots you mapped and how they split across the two clocks, so your A/Bs are reproducible. Keep the README's reminder in your head: there's no official H3 AYS table, so any manual set you import is your own experiment, not a discovered optimum. And don't stack it with other sampler-owning nodes at once - that rule applies to everything in this pack.
Install
Same as the rest of the 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 before you blame the node for anything - an old ComfyUI breaks the whole pack at import with missing comfy_api.latest or comfy.ldm.minimax. No extra downloads for this one.
Inputs (9)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| av_latent | LATENT | — | |
| steps | INT | 81–1000 | — |
| shift_video | FLOAT | 12.000.01–100 | — |
| shift_audio | FLOAT | 3.000.01–100 | — |
| schedule_profile | COMBO | native_flow_baseline | 2 options: native_flow_baseline, manual_h3_calibrated |
| manual_base_sigmas | STRING | [1.0, 0.875, 0.75, 0.625, 0.5, 0.375, 0.25, 0.125, 0.0] | Used only by manual_h3_calibrated. Supply steps+1 strictly descending base-flow sigmas from 1.0 to 0.0. |
| schedule_label | STRING | unvalidated H3 calibration | Human-readable provenance; a label is not validation evidence. |
| sampler_name | COMBO | dual_clock_euler | 45 options: dual_clock_euler, euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, +39 |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |
| sampler | SAMPLER | — |
| sigmas | SIGMAS | — |
| report_json | STRING | — |