GeoART extends the Predict-and-Perturb self-refinement framework to camera-conditioned video diffusion.
Geometry-aware test-time scaling improves camera control in video diffusion without retraining
GeoART scores multiple denoising candidates with a confidence-gated geometry prior and selects the next refinement state probabilistically, improving camera alignment and multi-view consistency on two backbones.
Chinese Tech
Liudi Yang · Ziyuan Liu
University of Freiburg · Huawei Heisenberg Research Center (Munich)
Research Digest··2 min read
The authors present GeoART, a training-free self-refinement method for camera-conditioned video diffusion.
§