Video LoRA

Where training actually lives

The core Wan2.2 repo is inference-first. LoRA training is documented in the community stack the README itself lists: DiffSynth-Studio (LoRA + full finetune + FP8 + offload), Musubi Tuner, and ComfyUI-oriented trainers that export WanVideoWrapper-compatible adapters.

MoE pairing

• Train or download two adapters aligned to the high-noise and low-noise experts.
• Apply each only on the SNR region that expert owns, or quality collapses at the switch.
• Rank and α conventions follow the trainer (DiffSynth vs Musubi) — do not mix loaders blindly.

Animate exception

Wan-Animate ships a relighting LoRA for replacement mode (--use_relighting_lora). The README explicitly recommends against using LoRA models trained on Wan2.2 T2V/I2V, because those deltas were not trained with the animation heads.

FAQ

Is there an official Alibaba LoRA trainer in Wan2.2?

Not in the main generate.py path. Use DiffSynth-Studio or another listed community trainer.

FP8 + LoRA?

Yes in DiffSynth. Merge or apply LoRA in higher precision when possible, then quantize.