Skip to content

FullSpectrum Studio v0.4.9 Official Release

Choose a tag to compare

@github-actions github-actions released this 30 May 22:59
· 10 commits to main since this release

FullSpectrum Studio v0.4.9 Official Release

This release is the first official package after the v0.4.8 community
reliability preview. It keeps the same honest scope: FullSpectrum reduces and
remaps existing Bambu painted .3mf states. It does not repaint, smooth or
clean up a badly painted model.

What changed

  • Smart quality is now the recommended default. It tests practical, balanced,
    detail and high-detail plans, scores the final validated palette, and keeps
    the best tradeoff for the painted usage in that model.
  • Physical anchor selection now favors filament choices that produce better
    final Bambu-reconstructed mixed swatches, not just closer direct colors.
  • macOS and Windows both expose a catalog-region selector for Bambu planning
    colors. The chosen region is written into warnings, JSON output and the
    shareable report.
  • Catalog regions are planning metadata only. FullSpectrum does not query live
    Bambu store stock, so check availability before buying filament.
  • The v0.4.8 fixes remain in place: real error reports, local debug logs,
    cancellable conversion, progress heartbeat text, readable dark-mode controls
    and safer large-model preview handling.

Compatibility

  • Primary input remains a painted Bambu .3mf.
  • Textured OBJ/GLB import remains experimental and size/format constrained.
  • Very complex or noisy painted files can still be slow or visually imperfect,
    because the output can only be as clean as the original paint-state data.
  • Every generated .3mf is reopened and validated before the app offers it to
    Bambu Studio or OrcaSlicer.

Validation used for this release

  • python3 -m unittest discover -s tests -v
  • swift build
  • ./script/build_and_run.sh build
  • Local conversion checks for catalog-region output, smart Bambu Core planning
    and CMYKW quality-100 planning on the public bald-eagle test model used
    during the v0.4.8 reliability pass.