ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

ATSplat is a novel feed-forward 3D Gaussian Splatting (3DGS) framework that addresses the inefficiency of existing pixel-aligned methods by restoring scene-adaptive capacity allocation. Unlike prior approaches that regress dense, redundant Gaussians tied to input image grids, ATSplat utilizes Adaptive 3D Tokens. It first lifts coarse patch-level depth and camera cues into sparse 3D anchor tokens, forming a compact scene scaffold. Each token then regresses local Gaussians with learnable 3D offsets, decoupling primitive placement from input pixels. An Adaptive Token Expansion module predicts uncertainty scores to selectively expand high-uncertainty tokens, concentrating primitives in challenging regions. Experiments on RealEstate10K and DL3DV show ATSplat achieves state-of-the-art rendering quality, reducing Gaussians by over 5.7×. It reconstructs scenes from 12 images (512×960) in under a second, rendering at 1136 FPS with only 311K Gaussians.

Key takeaway

For Machine Learning Engineers developing real-time 3D reconstruction or novel-view synthesis applications, ATSplat offers a significant advancement by providing a compact and efficient feed-forward 3DGS solution. Its adaptive capacity allocation and reduced Gaussian count (over 5.7× fewer) mean you can achieve state-of-the-art rendering quality and high frame rates (1136 FPS) with lower computational and storage costs. Consider integrating ATSplat to improve performance and scalability, especially for high-resolution or resource-constrained deployments.

Key insights

ATSplat adaptively allocates 3D Gaussian primitives based on scene complexity, enabling compact, high-quality novel-view synthesis in a single forward pass.

Principles

• Sparse initialization from coarse 3D anchors.
• Free 3D primitive placement via local offsets.
• Adaptive capacity allocation through token expansion.

Method

ATSplat lifts patch-level depth into sparse 3D anchor tokens, decodes them into local Gaussians with 3D offsets, and uses an Adaptive Token Expansion module to selectively expand high-uncertainty tokens, supervised by rendering error maps.

In practice

• Achieve 1136 FPS rendering at 512×960 resolution.
• Reduce Gaussian count by over 5.7× for efficiency.
• Reconstruct high-quality scenes in under one second.

Topics

• 3D Gaussian Splatting
• Novel-View Synthesis
• Feed-forward Models
• Adaptive Token Expansion
• Compact Representation
• Real-time Rendering

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, Machine Learning Engineer

Related on AIssential

• ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion — ATSplat restores adaptive 3DGS primitive allocation in feed-forward models via sparse 3D tokens and uncertainty-driven expansion.
• GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens — GlobalSplat uses a global latent scene representation for efficient, compact, and consistent 3D Gaussian Splatting.
• Rendering-Aware Bayesian 3D Gaussian Splatting with Native Uncertainty and Adaptive Complexity Control — Bayesian 3DGS integrates native uncertainty and adaptive complexity for robust novel-view synthesis and active view selection.
• GrainGS: Gradient-Decoupled Gaussian Splatting for Efficient Dynamic Novel View Synthesis — Decoupling canonical geometry from dynamic deformation and appearance gradients improves 3D Gaussian Splatting stability and quality.
• Feed-Forward 3D Scene Modeling: A Problem-Driven Perspective — Feed-forward 3D reconstruction offers efficient, generalizable scene modeling by directly mapping 2D inputs to 3D representations.
• ZeroSplat: Generalized Referring Segmentation in 3D Gaussian Splatting — ZeroSplat enables generalized 3D referring segmentation by lifting 2D VLM priors to 3D without per-scene optimization or feature storage.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.