QuerySplat
A feed-forward 3D Gaussian Splatting framework that decouples geometry and appearance for pose-free, high-fidelity novel-view synthesis.
Overview
Geometry first. Appearance, unconstrained.
QuerySplat predicts a renderable 3D Gaussian scene from images in a single feed-forward pass. It is built for novel-view synthesis when camera poses are not available in advance.
The model separates the information needed to understand where a scene is from the information needed to reproduce how it looks. This makes the representation more spatially flexible while preserving high-frequency appearance detail.
The official release includes custom-image preprocessing, 3D Gaussian prediction and rendering, VGGT-Omega camera and depth prediction, and optional test-time optimization.
Specifications
Key Capabilities
Pose-Free Geometry
A pretrained Vision Geometric Model provides spatial priors so the model can reason about scene geometry without camera poses supplied by the user.
Explicit Decoupling
Separate geometry and appearance branches reduce entanglement between spatial structure and visual detail during Gaussian prediction.
High-Frequency Detail
The dedicated appearance pathway recovers fine texture and color detail for sharper novel views than early query-based approaches.
Feed-Forward Inference
Run the released inference pipeline on custom images, with optional test-time optimization when additional reconstruction quality is needed.
Evaluation
State-of-the-art novel-view synthesis on DL3DV
On the challenging DL3DV benchmark, QuerySplat improves average PSNR over the strongest pose-free baseline by 2.30 dB and over the strongest pose-required baseline by 1.04 dB.
Average PSNR gain over the best pose-free baseline
Average PSNR gain over the best pose-required baseline
Method
A query decoder grounded in visual geometry
QuerySplat uses a dual-branch query-based decoder. The geometry branch reads spatially-aware memory from a pretrained Vision Geometric Model and defines the scene coordinate system. Geometry queries decode spatial Gaussian parameters, while appearance queries read RGB and Plücker memory to recover the remaining attributes. Differentiable splatting supervises the resulting Gaussians in the VGM-defined coordinate system.
Experience
Take spatial reconstruction beyond the paper
Explore InSpatio spatial experiences on the web or on iOS. These product entry points are designed for experiencing reconstructed scenes, while the full QuerySplat implementation and weights remain available in the research release.
Open Source
Run the official release
The repository contains inference code, configuration, preprocessing, rendering, and checkpoint instructions. QuerySplat and VGGT-Omega weights are downloaded separately from Hugging Face.
QuerySplat-authored code is provided under Apache-2.0. Vendored VGGT-Omega/DINOv3 components retain their original licenses.
Frequently Asked Questions
What is QuerySplat?
QuerySplat is a feed-forward 3D Gaussian Splatting framework that uses geometric priors and explicitly decouples geometry from appearance to predict renderable 3D Gaussian scenes from multi-view images.
Does QuerySplat require camera poses?
No precomputed camera poses are required. Its geometry branch uses a pretrained Vision Geometric Model for spatial understanding and predicts cameras and 3D Gaussians in the model-defined coordinate system.
How does QuerySplat perform on DL3DV?
QuerySplat achieves state-of-the-art novel-view synthesis performance on DL3DV, improving average PSNR by 2.30 dB over the best pose-free baseline and 1.04 dB over the best pose-required baseline.
Is QuerySplat open source?
Yes. The official GitHub repository provides the inference implementation, and QuerySplat model weights are published on Hugging Face. QuerySplat-authored code uses Apache-2.0, while third-party components retain their own licenses.
All Models
Browse the model libraryResearch
Explore publications