| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2026-10-18 | 1.1 kB | |
| v4.10.1 - KohakuFA speedup, MOSS Image flow schedule shift fix source code.tar.gz | 2026-10-18 | 9.5 MB | |
| v4.10.1 - KohakuFA speedup, MOSS Image flow schedule shift fix source code.zip | 2026-10-18 | 12.0 MB | |
| Totals: 3 Items | 21.4 MB | 13 | |
Features
- MOSS Image can now train using edit / reference datasets like Krea2 or Flux can
- MOSS Image flow schedule shift now allows better training at 1024px or higher
Bugfixes
- MOSS Image batched training no longer loops over each sequence, instead running a single forward pass through the transformer
- KohakuFA CuTe attention now correctly reuses kernels across different text input (caption) lengths
What's Changed
- Add MOSS-Image edit and reference LoRA training by @bghira in https://github.com/bghira/SimpleTuner/pull/3277
- Correct MOSS-Image noise sampling and flow timestep grids by @bghira in https://github.com/bghira/SimpleTuner/pull/3278
- Batch MOSS-Image text and reference attention by @bghira in https://github.com/bghira/SimpleTuner/pull/3279
- Reuse Kohaku attention kernels across sequence lengths by @bghira in https://github.com/bghira/SimpleTuner/pull/3280
- merge by @bghira in https://github.com/bghira/SimpleTuner/pull/3281
Full Changelog: https://github.com/bghira/SimpleTuner/compare/v4.10.0...v4.10.1