Lucas Kuzma


InstantSplat: Revolutionizing 3D Scene Reconstruction with Speed and Precision


The landscape of Novel View Synthesis (NVS) is evolving rapidly, and InstantSplat is at the forefront of this transformation. Combining Coarse Geometric Initialization (CGI) for scene structure estimation with Fast 3D-Gaussian Optimization (F-3DGO) for pose optimization, InstantSplat sets a new benchmark for speed and accuracy in 3D scene reconstruction. This blog delves into the technical intricacies of InstantSplat, highlighting its groundbreaking contributions and methodology.

Why InstantSplat is a Game-Changer

Traditional methods for 3D scene reconstruction, such as COLMAP, struggle with sparse-view conditions where camera poses and intrinsics are unknown. InstantSplat addresses these challenges with a holistic solution that eliminates the need for pre-computed camera parameters, delivering state-of-the-art (SOTA) performance in record time.

Key Contributions

  1. Holistic Approach to Sparse-View Synthesis: InstantSplat integrates an explicit 3D Gaussian representation with pose priors from DUSt3R, an end-to-end dense stereo model. This obviates the need for pre-computed camera intrinsics and extrinsics, streamlining the reconstruction process.

  2. Speed and Efficiency: Capable of reconstructing large-scale scenes in under one minute on a modern GPU, InstantSplat outpaces traditional methods by a significant margin. For example, it accelerates optimization from approximately two hours (using Nope-NeRF) to just one minute.

  3. Performance Metrics: InstantSplat demonstrates substantial improvements in SSIM (Structural Similarity Index) and Absolute Trajectory Error (ATE) on datasets like Tanks & Temples and MVImgNet. Specifically, it boosts SSIM from 0.68 to 0.89 (a 32% improvement) and reduces ATE from 0.055 to 0.011.


InstantSplat’s methodology departs from the conventional sparse point cloud approach generated by Structure from Motion (SfM). Instead, it leverages DUSt3R to create dense point clouds and optimizes these through a novel graph-based process.

Detailed Breakdown

  1. Graph-Based Image Connection:

    • InstantSplat constructs a graph where nodes represent input images and edges represent shared visual content. This graph facilitates global alignment of point maps, ensuring consistency across images.
  2. Global Alignment:

    • The system refines point maps, transformation matrices, and scale factors for each image pair in the graph. This process minimizes differences between transformed point maps and a globally aligned point map, resulting in accurate camera pose estimation.
  3. Simultaneous Optimization:

    • InstantSplat optimizes camera extrinsics and the 3D model jointly. This involves balancing photometric loss (the difference between rendered and actual images) with the deviation from initial pose estimates.
  4. Weiszfeld Algorithm:

    • To calculate per-camera focal lengths, InstantSplat employs the Weiszfeld algorithm. By averaging per-frame focal lengths under the assumption of a single camera with a fixed focal length, it further refines the reconstruction process.

Technical Details

  • DUSt3R’s Role: DUSt3R estimates camera poses from image pairs, but individual pair estimations can lead to inconsistencies. InstantSplat addresses this by globally aligning these estimations through the graph-based approach.
  • Optimization Process: The system adjusts both camera poses and 3D Gaussians simultaneously, using a hand-tuned tradeoff between photometric loss and pose adjustment. This joint optimization ensures a robust and accurate 3D representation.


Why InstantSplat Matters

InstantSplat is a major leap forward in 3D scene reconstruction, offering freakishly fast and accurate results without relying on SfM. Its ability to handle unconstrained sparse-view synthesis with minimal preprocessing sets it apart from traditional methods.

Influences and Comparisons

InstantSplat builds on and surpasses previous work, including CF-3DGS, Nope-NeRF, and NeRFmm. However, DUSt3R is the pivotal contribution that underpins InstantSplat’s remarkable performance.

Hidden Insights

One aspect that InstantSplat does not elaborate on is the application of the Weiszfeld algorithm for focal length calculation. This detail, lifted from the DUSt3R paper, is a crucial component of the system’s precision but is not extensively covered in their documentation.


InstantSplat represents a significant advancement in the field of 3D scene reconstruction. Its combination of speed, accuracy, and innovative methodology makes it a valuable tool for applications requiring rapid and reliable 3D representations. While it comes with some caveats, such as the need for a high VRAM GPU and a non-commercial license, its benefits for previews and initial reconstructions are undeniable. Explore the potential of InstantSplat and revolutionize your approach to 3D scene reconstruction. For more information, visit the InstantSplat GitHub page and the arXiv paper.