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 | 2026-10-01 | 3.3 kB | |
| v4.9.3 - high quality qwen image 2.1 training source code.tar.gz | 2026-10-01 | 9.2 MB | |
| v4.9.3 - high quality qwen image 2.1 training source code.zip | 2026-10-01 | 11.6 MB | |
| Totals: 3 Items | 20.8 MB | 0 | |
Features
- Qwen Image 2.1 with high quality training assistant LoRA & quality-tuned config examples
- See this experiment log: https://huggingface.co/SimpleTuner/Qwen-Image-2.1-LoRA-experiments
- Caption-driven online assistant LoRA training method
- Shard-first aspect bucketing for webshart datasets
- Native JSON captions for webshart datasets
What's Changed
- Fix stochastic optimizer update scaling by @bghira in https://github.com/bghira/SimpleTuner/pull/3225
- Fix resolution labels for saved validation images by @bghira in https://github.com/bghira/SimpleTuner/pull/3226
- Use cached metadata-only reads for Webshart JSON captions by @bghira in https://github.com/bghira/SimpleTuner/pull/3227
- webshart native JSON caption handling by @bghira in https://github.com/bghira/SimpleTuner/pull/3224
- Add Qwen Image 2.1 support and measured VRAM presets by @bghira in https://github.com/bghira/SimpleTuner/pull/3228
- Preserve AnyFlow checkpoint RNG state and scope compilation variants by @bghira in https://github.com/bghira/SimpleTuner/pull/3229
- Add caption-driven assistant LoRA training and image generation by @bghira in https://github.com/bghira/SimpleTuner/pull/3230
- Cache Webshart VAE latents in shard-first aspect groups by @bghira in https://github.com/bghira/SimpleTuner/pull/3233
- Keep assistant frozen after bare-base regularisation targets by @bghira in https://github.com/bghira/SimpleTuner/pull/3234
- Use Ollin texture-fixed VAE for Qwen Image 2.1 by @bghira in https://github.com/bghira/SimpleTuner/pull/3235
- Fix Qwen Image 2.1 RGBA input to auxiliary feature encoders by @bghira in https://github.com/bghira/SimpleTuner/pull/3236
- Default Qwen Image 2.1 LoRA training to assistant v2 by @bghira in https://github.com/bghira/SimpleTuner/pull/3237
- Fix DINO patch features and Qwen 2.1 automatic flow shift by @bghira in https://github.com/bghira/SimpleTuner/pull/3238
- Preserve Qwen assistant opt-outs and update default coverage by @bghira in https://github.com/bghira/SimpleTuner/pull/3239
- Update Qwen Image 2.1 examples to assisted regularised REPA training by @bghira in https://github.com/bghira/SimpleTuner/pull/3240
- Fix per-rank RNG and sampler checkpoint saving under DDP by @bghira in https://github.com/bghira/SimpleTuner/pull/3241
- assistant v2 docs for qwen image 2.1 by @bghira in https://github.com/bghira/SimpleTuner/pull/3242
- perf(data): avoid repeated full-bucket scans in the sampler by @hjinnkim in https://github.com/bghira/SimpleTuner/pull/3243
- docs: add a dedicated Qwen Image 2.1 quickstart by @bghira in https://github.com/bghira/SimpleTuner/pull/3245
- merge by @bghira in https://github.com/bghira/SimpleTuner/pull/3246
- Strip filename-strategy captions to match text embed precache by @bls9707-claude in https://github.com/bghira/SimpleTuner/pull/3244
- merge by @bghira in https://github.com/bghira/SimpleTuner/pull/3247
- Use assistant v3 by default for Qwen Image 2.1 by @bghira in https://github.com/bghira/SimpleTuner/pull/3248
- merge by @bghira in https://github.com/bghira/SimpleTuner/pull/3249
New Contributors
- @bls9707-claude made their first contribution in https://github.com/bghira/SimpleTuner/pull/3244
Full Changelog: https://github.com/bghira/SimpleTuner/compare/v4.9.2...v4.9.3
