Skip to content

v0.1.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 03 Sep 23:00
v0.1.0
482f70d

Initial public release of MiniMax H3 Flow-Aligned Regenerate.

Highlights

  • H3-native low-resolution trajectory capture and time-aligned high-resolution guidance.
  • Progressive Target Input handoff for Continuum: early H3 sampling can run on a private lower-resolution grid before switching to the final grid without a separate learned-refine replay.
  • Integrated Continuum refine-state guidance for the existing MiniMax H3 Latent Upscaler + Refine path.
  • Experimental resolution-aware refine sigmas, temporal correspondence, acceleration, downsample consistency, reference-budget diagnostics, and attention diagnostics.
  • Runtime metrics distinguish logical sampler calls, actual H3 NFEs, Spectrum forecasts, handoff probes, guidance events, and resolution-map events.
  • CI covers Python 3.10-3.13, source-contract checks, Ruff, formatting, 133 tests, compilation, package build, and isolated wheel validation.

Tested operating point

The strongest matched difficult-motion result tested during development used:

  • 14 SA-Solver-PECE outer steps
  • 736x736 private source -> 896x896 target
  • source_mode=scale, source_scale=0.83
  • fixed handoff 0.35 -> actual ~0.358 / index 9
  • direction guidance 0.25
  • acceleration / temporal / consistency 0
  • 54 logical calls / 36 actual H3 NFEs / 18 Spectrum forecasts across two chunks
  • effectively 9 low-grid + 5 high-grid outer steps per chunk

D12 was judged better than D10, and D14 slightly better again. These are tested quality/speed tradeoffs, not universal optima.

Decoded-media validation

Completed smoke coverage includes:

  • integrated two-pass C7=7+7, C6=7+6, C5=7+5, plus exploratory C4=7+4;
  • Progressive Target Input at D10/D12/D14;
  • fixed and auto-computed handoff selection;
  • corrected HiFlow-style acceleration;
  • conservative temporal correspondence;
  • downsample consistency;
  • matched E0/E1 resolution-aware learned-refine sigma mapping with the real learned refiner enabled in both arms.

Acceleration, downsample consistency, auto handoff, D14 temporal guidance, and resolution-aware refine sigmas were structurally valid but did not show a clear decoded-media advantage in the matched difficult-motion tests, so they remain non-promoted experimental controls.

Earlier temporal runs decoded with the accidental TensorRT w4a16_awq VAE are excluded from temporal-quality attribution.

Scope

This project is a training-free research implementation informed by public work. It does not reproduce MiniMax's closed H3-Regenerate-2K model or unreleased sparse-attention topology. Broad cross-prompt quality claims are intentionally withheld; the repository documents the larger optional benchmark matrix for publication-grade evaluation.

Assets 2
Loading