v4.9.3 - high quality qwen image 2.1 training

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 #3225
• Fix resolution labels for saved validation images by @bghira in #3226
• Use cached metadata-only reads for Webshart JSON captions by @bghira in #3227
• webshart native JSON caption handling by @bghira in #3224
• Add Qwen Image 2.1 support and measured VRAM presets by @bghira in #3228
• Preserve AnyFlow checkpoint RNG state and scope compilation variants by @bghira in #3229
• Add caption-driven assistant LoRA training and image generation by @bghira in #3230
• Cache Webshart VAE latents in shard-first aspect groups by @bghira in #3233
• Keep assistant frozen after bare-base regularisation targets by @bghira in #3234
• Use Ollin texture-fixed VAE for Qwen Image 2.1 by @bghira in #3235
• Fix Qwen Image 2.1 RGBA input to auxiliary feature encoders by @bghira in #3236
• Default Qwen Image 2.1 LoRA training to assistant v2 by @bghira in #3237
• Fix DINO patch features and Qwen 2.1 automatic flow shift by @bghira in #3238
• Preserve Qwen assistant opt-outs and update default coverage by @bghira in #3239
• Update Qwen Image 2.1 examples to assisted regularised REPA training by @bghira in #3240
• Fix per-rank RNG and sampler checkpoint saving under DDP by @bghira in #3241
• assistant v2 docs for qwen image 2.1 by @bghira in #3242
• perf(data): avoid repeated full-bucket scans in the sampler by @hjinnkim in #3243
• docs: add a dedicated Qwen Image 2.1 quickstart by @bghira in #3245
• merge by @bghira in #3246
• Strip filename-strategy captions to match text embed precache by @bls9707-claude in #3244
• merge by @bghira in #3247
• Use assistant v3 by default for Qwen Image 2.1 by @bghira in #3248
• merge by @bghira in #3249

New Contributors

• @bls9707-claude made their first contribution in #3244

Full Changelog: v4.9.2...v4.9.3