Repository navigation
v4.10.1 - KohakuFA speedup, MOSS Image flow schedule shift fix
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 #3277
- Correct MOSS-Image noise sampling and flow timestep grids by @bghira in #3278
- Batch MOSS-Image text and reference attention by @bghira in #3279
- Reuse Kohaku attention kernels across sequence lengths by @bghira in #3280
- merge by @bghira in #3281
Full Changelog: v4.10.0...v4.10.1