LoGo checkpoints

The LoRA adapter each model was post-trained to with the LoGo reward.

file	model	base weights
logo-lyra2-lora.pt	Lyra-2 (use DMD scheduler)	nvidia/Lyra-2.0
logo-lingbot-world-v2-lora.pt	LingBot-World 2.0	robbyant/lingbot-world-v2-14b-causal-fast
logo-uniworld-view-lora.safetensors	UniWorld-View	Drexubery/UniView + Wan-AI/Wan2.1-VACE-14B-diffusers + CausVid LoRA

Usage, with the code at https://github.com/ziqi-ma/logo.

# Lyra-2
torchrun --standalone --nproc_per_node=8 -m lyra_2._src.rl.inference.evaluate \
  --adapter logo-lyra2-lora.pt --scenes-root <scene dir> --scenes "<ids>"

# LingBot-World 2.0
python -m wan.rl.inference.gen_scenes --input_base <scene root> --videos_base <out> \
  --lora_path logo-lingbot-world-v2-lora.pt

# UniWorld-View
python -m rl.inference.eval_gen --ws_prefix <out> \
  --rl-ckpt logo-uniworld-view-lora.safetensors --pose-scale 1.5

Evaluation on TrajectoryBench, including the per-group generation settings, is described in eval/README.md in that repository.