GaussianTrim3R: Controllable 3D Gaussians Pruning for Feedforward models

ICLR 2026 Conference Submission6732 Authors

Keywords: Gaussians, feed-forward

TL;DR: Feed forward pruning of 3D Gaussians in sparse view

Abstract: Feed-forward methods offer a promising paradigm for novel-view synthesis, eliminating computationally expensive per-scene optimization. However, current feed-forward approaches typically predict a fixed number of pixel-aligned Gaussian primitives, leading to significant redundancy. Naively pruning these Gaussians creates severe visual artifacts, necessitating fine-tuning that compromises the feed-forward nature. We introduce GaussianTrim3R, a novel framework for controllable and feed-forward 3D Gaussian representation method which gradually prunes 3D Gaussians and simultaneously adjusts the attributes of remaining Gaussians maintaining rendering quality, thus eliminating the need for finetuning 3D Gaussians post pruning. To achieve this, we construct SuperClusters by partitioning the 3D scene based on spatial and color attributes. By leveraging Discrete Wavelet Transform, we assign and rank texture complexity to these SuperClusters, enabling selective, texture-aware pruning. Doing so enables our method to directly predict attribute-adjusted Gaussians, thereby preserving scene integrity. Unlike existing methods, GaussianTrim3R offers an efficient, real-time solution with extensive experiments demonstrating superior trade-offs between quality and efficiency across diverse real world RealEstate10K, ACID and DTU datasets.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 6732