Xiaoming Liu

We propose TRIM, a post-training framework that accelerates 3D Gaussian diffusion by pruning trajectories and background tokens, enabling efficient inference-time scaling without sacrificing quality.