Omnidreams edit-SFT style skins: offline pair generation, VLM filter, mid-stream LoRA trainer - #458

Open

wenqingw-nv wants to merge 8 commits into

NVIDIA:mainfrom

wenqingw-nv:omnidreams-style-skin-pr

Commits on Aug 13, 2026

Add edit-SFT style-skin pipeline: JoyAI pair generation, VLM filter, LoRA trainer

wenqingw-nv
and
claude
committed
d151d57
Fix pre-swap text cache built at LoRA scale 1

wenqingw-nv
and
claude
committed
0ce92eb

Commits on Aug 14, 2026

Add style-maintenance episodes to keep long windows stable

wenqingw-nv
and
claude
committed
d1b97d9

Commits on Aug 18, 2026

Fix pre-swap cache LoRA-scale leak; add v4 episodes and teacher-regen v5 recipe

wenqingw-nv
and
claude
committed
bbe4d12

Commits on Aug 22, 2026

Add pre-merged text-edit LoRA deploy hook with release_targets seam

wenqingw-nv
and
claude
committed
1713595
Add CUDA-graph-safe fused drift corrector with per-state dispatch

wenqingw-nv
and
claude
committed
4cd5e0c
Fix fp32 base-weight aliasing in TextEditLoRA's pre-merge

wenqingw-nv
and
claude
committed
630b66b
Add style-drift corrector data-gen, trainer, and gate calibration

wenqingw-nv
and
claude
committed
6221f0c