Post
1748
Just open-sourced LoRA Gym with Timothy - production-ready training pipeline for character, motion, aesthetic, and style LoRAs on Wan 2.1/2.2, built on musubi-tuner.
16 training templates across Modal (serverless) and RunPod (bare metal) covering T2V, I2V, Lightning-merged, and vanilla variants.
Our current experimentation focus is Wan 2.2, which is why we built on musubi-tuner (kohya-ss). Wan 2.2's DiT uses a Mixture-of-Experts architecture with two separate experts gated by a hard timestep switch - you're training two LoRAs per concept, one for high-noise (composition/motion) and one for low-noise (texture/identity), and loading both at inference. Musubi handles this dual-expert training natively, and our templates build on top of it to manage the correct timestep boundaries, precision settings, and flow shift values so you don't have to debug those yourself. We've also documented bug fixes for undocumented issues in musubi-tuner and validated hyperparameter defaults derived from cross-referencing multiple practitioners' results rather than untested community defaults.
Also releasing our auto-captioning toolkit for the first time. Per-LoRA-type captioning strategies for characters, styles, motion, and objects. Gemini (free) or Replicate backends.
Current hyperparameters reflect consolidated community findings. We've started our own refinement and plan to release specific recommendations and methodology as soon as next week.
Repo: github.com/alvdansen/lora-gym
- 2 replies
