Repository navigation
Changes since v0.1.28:
- fix(lora): make checkpoint resume work end-to-end
- Reject unsupported image and video media roles before a task is queued.
- feat(lora): pick the best checkpoint from a same-condition candidate grid
- fix(lora): train Z-Image and Qwen faces so identity sticks (#76)
- Remove per-image LoRA captions and VLM auto-caption (#81)
- fix(lora): train every image with one caption
- fix(lora): keep face-dataset skip list and identity gates
- fix(lora): harden face-lock training contracts and wizard gates
- feat(frontend): share one model select for badge and description tip (#79)
- Add LTX 2.5 MSR Ref2VA end-to-end with registry-driven ref2va capability.
- Connect multi-reference image edits and upgrade Qwen Image Edit to 2511 (#75)
- fix(lora): train Z-Image and Qwen faces at 1024 with sigmoid noise
- Add optional audio denoise and route H3 Veda SDPA through the shared helper.
- fix(ace_step): VAE tiling, turbo shift, Haar DCW, and cover conditioning parity (#74)
- feat(ace_step): DiT cross-attention lyric timeline + gallery synced scrolling lyrics (#73)
- fix(ace_step): upstream parity fixes for timbre encoder, ResidualFSQ, APG guidance (#72)