MiniMax H3 R40 · Isolated Stage-2 Sampling
Iterate shots without re-running everything
- model
- cine_linx
- shot_lab
- model
- noise
- sampler
- sigmas
- report
By the R40 generation of IAMCCS's MiniMax H3 pipeline, the architecture got more opinionated: stage-2 refinement became something you tune separately, driven by a Shot Lab plan instead of the main settings. This node is the stage-2 sampler for that world - the "isolated" in its display name means its seed, sigmas and acceleration are fully controlled by a separate shot_lab object so you can iterate on refine passes shot-by-shot without re-running or re-encoding stage 1.
The R40 Shot Lab concept (the IAMCCS_H3_R40_SHOT_LAB input) is essentially a per-shot control deck: it carries stage2_seed, stage2_seed_stride, a sigma preset, sparse-attention settings, and a stage-2 LoRA. Think of it as the difference between tweaking a global setting that affects every shot vs. dialing in "this specific take" the way a director would.
How it works
Same shape as the R38 refine sampler - builds noise, sampler, sigmas and a patched model for a custom sampler - but with R40's tuning knobs layered on:
- Seed = shot_lab's
stage2_seed+ segment_index ×stage2_seed_stride. Independent of stage 1, so refining shot 4 doesn't disturb shots 1–3. - LoRA stack: turbo LoRA, then secondary LoRA, then the shot_lab's own stage-2 LoRA. Three application layers, all from the plan.
- Sigma shift via
MiniMaxH3SigmaShift(video shift 12 / audio 3 from sampling config). - Sparse attention (
H3SparseAttention) if the shot lab asks: a video-budget fraction (default 0.30) with denser early/late steps. This is the "how much attention can we skip" lever that buys speed on long clips. - Sigma preset:
seedhunter_3(default),seedhunter_4,seedhunter_5- hand-tuned sigma schedules that end at 0.0. These replace the "just trust the scheduler" approach with explicit step boundaries tuned for the refine regime. - Default sampler
er_sde.
The _apply_sparse path depends on an H3-Optimizations node being registered; if you request sparse and it isn't available, it raises telling you exactly that.
Inputs
model,cine_linx,segment_index- as with R38.shot_lab- theIAMCCS_H3_R40_SHOT_LABobject from the Shot Lab control node upstream. This is what makes it "isolated".
Outputs
model, noise, sampler, sigmas, report - feed the first four to SamplerCustomAdvanced.
Install
Pack install (git clone https://github.com/IAMCCS/IAMCCS-nodes.git), MiniMax H3 support, and - only if you use sparse mode - the external H3-Optimizations node. The seedhunter sigma presets are built in; nothing to download for those.
Gotchas
- This node needs the R40 Shot Lab input. There's no fallback if
shot_labis missing - the node exists specifically because the author wanted stage-2 sampling decoupled from the main plan, so don't wire it like an R38 node. - Sparse attention budget is a budget (0.30 = aim to skip ~70% of video attention). Crank it too aggressive and refinement quality dips; the denser-early/late-steps default is there to protect the parts of the schedule that matter.
- sigma preset names come from IAMCCS's own seedhunter tuning runs. If you override the preset to
manual, you're on the list of hard-coded schedules - there isn't a free-form custom sigma editor on this node.
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| cine_linx | IAMCCS_SUPERNODE_LINX | — | |
| segment_index | INT | — | |
| shot_lab | IAMCCS_H3_R40_SHOT_LAB | — |
Outputs (5)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |
| noise | NOISE | — |
| sampler | SAMPLER | — |
| sigmas | SIGMAS | — |
| report | STRING | — |