Spark3R: Asymmetric Token Reduction Makes Fast Feed-Forward 3D Reconstruction

TL;DR

Spark3R introduces a novel asymmetric token reduction method for faster 3D reconstruction models, significantly improving efficiency without retraining by decoupling token compression based on their roles.

Contribution

It proposes a training-free, asymmetric token reduction framework that adaptively compresses query and key-value tokens differently, enhancing speed while preserving quality.

Findings

Achieves up to 28x speedup on 1000-frame inputs.

Maintains competitive reconstruction quality across multiple models.

Integrates seamlessly with pretrained models without retraining.

Abstract

Feed-forward 3D reconstruction models based on Vision Transformers can directly estimate scene geometry and camera poses from a small set of input images, but scaling them to video inputs with hundreds or thousands of frames remains challenging due to the quadratic cost of global attention layers. Recent token-merging methods accelerate these models by compressing the token sequence within the global attention layers, but they apply a uniform reduction to query tokens and key-value tokens, ignoring their functionally distinct roles in 3D reconstruction. In this work, we identify a key property of feed-forward 3D reconstruction models: query tokens encode view-specific geometric requests and are sensitive to compression, while key-value tokens represent shared scene context and tolerate aggressive compression. Guided by this insight, we propose Spark3R, a training-free acceleration framework that decouples the compression of query tokens and key-value tokens by assigning distinct reduction factors, with intra-group token merging applied to query tokens and lightweight token pruning to key-value tokens. Additionally, Spark3R adaptively adjusts the key-value reduction factor across layers, further improving the quality-efficiency trade-off. As a plug-and-play framework requiring no retraining, Spark3R integrates directly into multiple pretrained feed-forward 3D reconstruction models, including VGGT, $\pi^3$, Depth-Anything-3, and VGGT-$\Omega$, and achieves up to $28\times$ speedup on 1,000-frame inputs while maintaining competitive reconstruction quality.