This document summarizes the key highlights for each released version. Detailed change descriptions live in the per-version files under this folder.
- 3.3.0 — DreamOmni2 trainer support with cumulative offset positioning and VLM prompt optimization.
- 3.2.1 — Bug fixes: scheduler state isolation and directory cleanup improvements.
- 3.2.0 — Multi-logger support (TensorBoard/wandb/SwanLab) with unified LoggerManager interface.
- 3.1.0 — Validation sampling during training with TensorBoard visualization.
- 3.0.2 — Fixed FSDP LoRA checkpoint saving issues and performance benchmarking.
- 3.0.1 — Bug fixes, test infrastructure improvements, and dependency updates.
- 3.0.0 — Multi-resolution mixed training pipeline, dynamic resolution candidate selection.
- 2.4.1 — Documentation polish and dataset guide updates.
- 2.4.0 — Dynamic shape support for Qwen-Image-Edit-Plus.
- 2.3.0 — Full Qwen-Image-Edit-Plus architecture integration.
- 2.2.0 — CSV metadata dataset support.
- 2.1.0 — FSDP training path and distributed optimizations.
- 2.0.0 — Major refactor with revised trainer stack.
- 1.6.0 — Training workflow improvements and LoRA upload automation.
- Older versions are tracked in the
legacysubfolder when applicable.
- Track pending features and bugfixes that target the next release.
- Update this section when pull requests land without an immediate release tag.