Guided Trajectory Optimization with Sparse Scaling for Test-Time Diffusion
The episode discusses a paper titled "Guided Trajectory Optimization with Sparse Scaling for Test-Time Diffusion," which introduces RTS, a Reward-guided Trajectory Scaling method. The hosts explain how RTS uses reward guidance and PCA-based dimensionality reduction to efficiently find optimal noise trajectories during diffusion model inference, aiming for better image quality and computational efficiency.