Download Latest Version
v4.10.1 - KohakuFA speedup, MOSS Image flow schedule shift fix source code.zip (12.0 MB)
Add to Preferred Sources
| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2026-10-17 | 3.0 kB | |
| v4.10.0 - high-quality MOSS-Image _ Flux 3 Klein training source code.tar.gz | 2026-10-17 | 9.4 MB | |
| v4.10.0 - high-quality MOSS-Image _ Flux 3 Klein training source code.zip | 2026-10-17 | 11.9 MB | |
| Totals: 3 Items | 21.4 MB | 0 | |
Features
- MOSS-Image with VRAM-optimised configuration examples
- Flux 3 Klein Image with a distillation training adapter (assistant LoRA)
- KohakuFA attention is now set by default when
attention-cuteextra is installed
Bugfixes
- MultiGPU (FSDP2 and DDP) training correctness fixes
- WebUI metrics display and collection fixes
- VAE caching should be a little faster now with large datasets
What's Changed
- Fix Qwen-Image 2.1 FSDP metadata after model selection by @hjinnkim in https://github.com/bghira/SimpleTuner/pull/3258
- Fix rank-dependent VAE bucket geometry by @hjinnkim in https://github.com/bghira/SimpleTuner/pull/3254
- Fix redundant image resizing during VAE caching by @hjinnkim in https://github.com/bghira/SimpleTuner/pull/3255
- Document the LyCORIS 4 LoKR kernel-size workaround by @bghira in https://github.com/bghira/SimpleTuner/pull/3259
- Fix training reports with large metric and timestep histories by @bghira in https://github.com/bghira/SimpleTuner/pull/3260
- Keep Qwen Image VAE attention independent of denoiser kernel selection by @bghira in https://github.com/bghira/SimpleTuner/pull/3263
- Normalize text-file captions before training cache lookups by @bghira in https://github.com/bghira/SimpleTuner/pull/3262
- Preserve sparse evaluation metrics when downsampling training records by @bghira in https://github.com/bghira/SimpleTuner/pull/3261
- Add ComfyUI export for fused Krea2 full-matrix LoKR and DoRA by @bghira in https://github.com/bghira/SimpleTuner/pull/3264
- Support value gradient clipping under FSDP2 by @hjinnkim in https://github.com/bghira/SimpleTuner/pull/3265
- Fix LoRA and LyCORIS training under DDP and FSDP2 by @hjinnkim in https://github.com/bghira/SimpleTuner/pull/3266
- Fix LR scheduler position being overwritten on resume by @hjinnkim in https://github.com/bghira/SimpleTuner/pull/3267
- Require CaptionFlow 0.5.7 and PyTorch 2.13 by @bghira in https://github.com/bghira/SimpleTuner/pull/3269
- Keep container device-list paths out of CUDA_VISIBLE_DEVICES by @bghira in https://github.com/bghira/SimpleTuner/pull/3270
- Make automatic Kohaku attention the training default by @bghira in https://github.com/bghira/SimpleTuner/pull/3256
- Add FLUX.3 Image Klein training and image editing support by @bghira in https://github.com/bghira/SimpleTuner/pull/3271
- Preserve model-card base identity for local checkpoints by @bghira in https://github.com/bghira/SimpleTuner/pull/3272
- Fix Flux.3 Image Klein segmented checkpointing and add GPU measurements by @bghira in https://github.com/bghira/SimpleTuner/pull/3274
- Use the released FLUX.3 Klein training assistant by default by @bghira in https://github.com/bghira/SimpleTuner/pull/3275
- Add MOSS-Image Base and SFT training support by @bghira in https://github.com/bghira/SimpleTuner/pull/3273
- merge by @bghira in https://github.com/bghira/SimpleTuner/pull/3276
Full Changelog: https://github.com/bghira/SimpleTuner/compare/v4.9.3...v4.10.0
