š How it works:
ā¢ Embed camera trajectories into diffusion noise, zero extra modules
ā¢ 3D rewards from Depth Anything 3 + Qwen3-VL as geometry critics
ā¢ Periodic decoupled training: buildings stay rigid, flags still wave šļøš©
ā¢ 3K text prompts only, no video data
Weijie Wang
