# STream3R: Scalable Sequential 3D Reconstruction with Causal Transformer

**Authors:** Yushi Lan, Yihang Luo, Fangzhou Hong, Shangchen Zhou, Honghua Chen, Zhaoyang Lyu, Shuai Yang, Bo Dai, Chen Change Loy, Xingang Pan  
**arXiv:** [2508.10893](https://arxiv.org/abs/2508.10893)  
**Date:** 2025-08-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2508.10893v1/

## Key points

- Authors: Yushi Lan, Yihang Luo, Fangzhou Hong, Shangchen Zhou, Honghua Chen, Zhaoyang Lyu, Shuai Yang, Bo Dai, Chen Change Loy, and Xingang Pan
- Approach: Introduces a groundbreaking approach to 3D reconstruction called STream3R
- Innovation: Reimagines pointmap prediction as a decoder-only problem using causal attention
- Generalization: Demonstrates strong generalization capabilities across diverse and challenging scenarios by leveraging geometric priors learned from extensive 3D datasets
- Performance: Outperforms prior approaches in both static and dynamic scene benchmarks
- Compatibility: Compatible with Large Language Model (LLM)-style training infrastructure for efficient large-scale pretraining and fine-tuning for various downstream 3D tasks
- Potential: Underscores the potential of causal models for real-time 3D perception in streaming environments
- Further Details: More information available on the project page at https://nirvanalan.github.io/projects/stream3r

## Summary

In their paper titled "STream3R: Scalable Sequential 3D Reconstruction with Causal Transformer," authors Yushi Lan, Yihang Luo, Fangzhou Hong, Shangchen Zhou, Honghua Chen, Zhaoyang Lyu, Shuai Yang, Bo Dai, Chen Change Loy, and Xingang Pan introduce a groundbreaking approach to 3D reconstruction. The  method reimagines pointmap prediction as a decoder-only  problem. It departs from existing techniques that rely on costly global optimization or limited memory mechanisms that struggle with longer sequences. The key innovation of  lies in its , which efficiently processes image sequences using causal attention inspired by advancements in modern language modeling. By leveraging geometric priors learned from extensive 3D datasets,  demonstrates strong generalization capabilities across diverse and challenging scenarios. Particularly noteworthy is its success in capturing dynamic scenes where traditional methods often fall short. Extensive experiments showcased the superior performance of  compared to prior approaches across both static and dynamic scene benchmarks. Furthermore, the method's compatibility with Large Language Model (LLM)-style training infrastructure enables efficient large-scale pretraining and fine-tuning for various downstream 3D tasks. The results presented in this study underscore the potential of causal  models for real-time 3D perception in streaming environments. The authors provide further details on their project page at https://nirvanalan.github.io/projects/stream3r. This research opens up exciting possibilities for advancing online 3D understanding and paves the way for enhanced capabilities in real-time 3D reconstruction applications.