Open-sourced our LoRA training pipeline - LoRA Gym. Wan 2.1/2.2 support, 16 templates across Modal and RunPod, handles the Wan 2.2 dual-expert setup out of the box. Also releasing our captioning toolkit for the first time - per-LoRA-type strategies for characters, styles, motion, objects.
Still early, recommended settings coming next week.
github.com/alvdansen/lora-gym

.jpeg)