github mflux-community/mflux v.0.19.2
Release 0.19.2

5 hours ago

0.19.2

Fixed

  • Replaying a run with --config-from-metadata works again for runs made without --base-model; sidecars written by earlier versions replay too. (#696)
  • mflux-generate-flux2 and mflux-generate-z-image pick the family sibling from a custom checkpoint's name again (mlx-community/flux2-klein-9b-8bit runs as klein-9b without --base-model); a name that matches nothing keeps the command's default model. (#697)
  • The --steps default for a third-party or local checkpoint now follows the model it resolves to (via --base-model or the alias in its name) instead of always being 25. (#699)
  • mflux-generate-z-image now honors --base-model for custom checkpoints; family inference reads only the checkpoint's own name, so a parent directory can no longer reshape it; the flux2 guidance error names --base-model as the fix. (#702)

Internal

  • CHANGELOG.md is retired; release notes for each version live in GitHub Releases. (#700)

Changed

  • ci: keep dynamic values out of shell source in release and ci-extract (#679) [needs edit: no release-note block]
  • fix: give the Flux generation variants save_model, and make dev-redux saveable end-to-end (#680) [needs edit: no release-note block]
  • just install ensure current directory is a git repo (#683) [needs edit: no release-note block]
  • Release notes are now harvested from each PR's release-note block into a draft GitHub Release that a maintainer reviews before publishing. CHANGELOG.md is no longer edited by hand. (#685)
  • fix: flux2 --base-model arg handling (#688) [needs edit: no release-note block]
  • fix: SeedVR2 upscaling crashes on mlx 0.32.2 (#690) [needs edit: no release-note block]
  • CUDA ONLY - bump mlx to 0.32.2 (#692) [needs edit: no release-note block]
  • Shell completion now suggests --base-model values (the canonical model names). (#706)
  • The metadata sidecar now records the checkpoint path a run used, and --config-from-metadata replays from it instead of silently substituting the built-in model's weights. (#708)
  • Training: new opt-in "gradient_checkpointing": true config flag for FLUX.2, Z-Image and Ernie LoRA training, and the backward pass is now evaluated last-block-first whenever checkpointing is active (Krea 2 included), so the memory saving actually materializes. FLUX.2 klein-base-4B edit training fits on 32–36 GB Macs with it. (#711)
  • Update README.md (#720)

Don't miss a new mflux release

NewReleases is sending notifications on new releases.