Skip to content

v0.0.5

Choose a tag to compare

View all tags
@lmangani lmangani released this 26 Sep 10:09
· 6 commits to main since this release
v0.0.5
4f961e6

Highlights

  • Faster Ref2VA on Metal (M3 Ultra measurements): Video VAE encoder, Qwen long-prompt attention, and DiT SDPA — large jobs much quicker (e.g. ref-video encode ~10→3 min; long Qwen encoder ~195→38 s)
  • Upstream h3.c Metal correctness: GQA F32 scale + barrier, 16-byte threadgroup align, finite VAE RGB clamp, warm VAE cache across ref edits (h3_cache_invalidate_inputs), Ultra long-SDPA query split
  • VAE tile seams fixed: decoder locked to official 256px tiles (Metal 320 path quilted a 16px grid)
  • Silent-video ingest + PyAV audio adapt for H3; Lanczos post-hoc “upscale” retired
  • Generation timings: per-phase wall clock in clip metadata (timings) and server log; phase marked complete when h3 advances (fixes undercounted denoise steps)

Install

Download H3-WS-macOS.dmg from this release after CI attaches it (~6–10 min on macos-26).

Apple Silicon only. Weights stay on your machine (models/MiniMax-H3); the DMG does not include checkpoints.

Assets 4
Loading