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

2026-07-22 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

ATSplat is a novel feed-forward 3D Gaussian Splatting (3DGS) framework designed to restore the adaptive primitive allocation capabilities often lost in existing feed-forward methods. Unlike pixel-aligned formulations that create dense, redundant Gaussian sets based on image resolution, ATSplat utilizes Adaptive 3D Tokens. It first converts coarse patch-level depth and camera data into sparse 3D anchor tokens, forming a compact scene scaffold. Each token then regresses into local Gaussians with learnable 3D offsets, decoupling primitive placement from input image grids. An Adaptive Token Expansion module predicts token-level uncertainty, guided by rendering error maps, to selectively expand high-uncertainty tokens. This sparse-to-adaptive approach concentrates primitives in challenging areas, maintaining a compact representation. Experiments on RealEstate10K and DL3DV datasets demonstrate state-of-the-art rendering quality, reducing Gaussians by over 5.7x compared to dense methods. ATSplat reconstructs scenes in under a second from 12 512x960 input images and renders novel views at 1136 FPS with only 311K Gaussians.

Key takeaway

For Computer Vision Engineers developing real-time 3D reconstruction and novel-view synthesis systems, ATSplat offers a significant advancement. You should consider integrating its adaptive token expansion approach to achieve high rendering quality with substantially fewer Gaussians. This method allows your systems to reconstruct complex scenes in under a second and render views at over 1000 FPS, optimizing both computational resources and visual fidelity for applications like virtual reality or autonomous navigation.

Key insights

ATSplat restores adaptive 3DGS primitive allocation in feed-forward models via sparse 3D tokens and uncertainty-driven expansion.

Principles

• Decouple primitive placement from input grids.
• Use uncertainty to guide adaptive resource allocation.
• Sparse scaffolds enable compact scene representations.

Method

ATSplat lifts patch-level cues to sparse 3D anchor tokens, regresses local Gaussians with offsets, then expands high-uncertainty tokens using rendering error maps.

In practice

• Implement sparse 3D anchor tokens for efficient scene scaffolds.
• Integrate uncertainty scores to adaptively densify primitives.
• Optimize 3D Gaussian Splatting for real-time novel view synthesis.

Topics

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

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

Related on AIssential

• ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion — ATSplat adaptively allocates 3D Gaussian primitives based on scene complexity, enabling compact, high-quality novel-view synthesis in a single forward pass.
• 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.
• InvSplat: Inverse Feed-Forward Scene Splatting — InvSplat predicts 3D Gaussian representations with intrinsic material attributes for inverse rendering in a single forward pass.
• 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.
• Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting — LGTM enables 4K novel view synthesis by decoupling geometry from resolution using textured Gaussian primitives.
• 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 Computer Vision and Pattern Recognition.