Abstract
Snap-Snap is a novel feed-forward framework designed to reconstruct high-quality 3D human models from only two images—front and back—in just 190 milliseconds. By bypassing the requirement for complex human mesh priors (such as SMPL) or expensive dense camera arrays, Snap-Snap enables high-fidelity 3D digital human creation using standard mobile devices. The system achieves state-of-the-art performance by combining point cloud prediction with Gaussian attribute regression, offering a practical solution for real-time applications in the Metaverse and mobile augmented reality.
Introduction
Creating a 3D digital double of a person has traditionally been a choice between two extremes: expensive professional setups utilizing dozens of synchronized cameras for dense-view reconstruction, or single-view methods that rely heavily on pre-defined human shape "priors." While dense views provide high fidelity, they are inaccessible to average users. Conversely, single-view methods often struggle with pose misalignment and uncontrollable textures.
Snap-Snap identifies a "Goldilocks" zone: two-view reconstruction. By capturing just the front and back of a subject, users can provide enough information for a complete model without specialized hardware. However, this presents a significant technical challenge: there is virtually zero overlap between the two images, making it difficult to establish 3D consistency and "hallucinate" the missing side geometry.
Methodology
Snap-Snap introduces a three-stage, purely feed-forward pipeline that transitions from 2D images to 3D Gaussians without the need for per-scene optimization.
1. Point Cloud Prediction
The first stage adapts the architecture of DUSt3R, a foundation model for 3D reconstruction. Given front and back images, the model predicts point clouds for four viewpoints: front, back, left, and right. Specialized "side heads" are trained to infer plausible geometry for the sides of the body, effectively learning the spatial relationship between the visible front and back surfaces.
2. Side-view Enhancement
Because the predicted side point clouds initially lack color, the framework employs a Nearest Neighbor Search (NNS) algorithm. This process "warps" color information from the known front and back views onto the side point clouds. These enhanced point clouds are then projected into "pseudo-views" to create a complete, colored four-view representation of the subject.
3. Gaussian Attribute Regression
In the final stage, a UNet-like network processes the colored point clouds and images to regress the attributes of 3D Gaussians—including position offsets, opacity, color, scale, and rotation. Utilizing 3D Gaussian Splatting (3DGS) allows for significantly higher rendering quality and faster inference speeds compared to traditional point clouds or meshes.
Key Insights
A core innovation of Snap-Snap is its geometry-over-templates approach. Most human reconstruction methods rely on the SMPL model, a "naked" human mesh template. Snap-Snap ignores this constraint, allowing it to capture complex geometry such as flowing dresses, jackets, and hair that SMPL-based methods typically fail to represent accurately. Furthermore, by training on extensive human datasets like THuman2.0 and 4D-Dress, the model learns a robust "side prior," enabling it to fill in occluded regions with high plausibility.
Results and Discussion
Snap-Snap demonstrates exceptional efficiency and accuracy. It reconstructs a full human model in approximately 190 ms on a single NVIDIA RTX 4090, achieving a processing speed of ~5 FPS.
In quantitative evaluations on the THuman2.0 dataset, Snap-Snap achieved a PSNR of 22.44, outperforming state-of-the-art methods such as GHG (21.79) and even GPS-Gaussian (20.70), despite the latter requiring five input views. Qualitative results further highlight the model's robustness to "in-the-wild" data captured via low-cost mobile phones, maintaining pose and texture consistency even with loose clothing.
Conclusion
Snap-Snap represents a significant step forward in accessible 3D human reconstruction. By leveraging a two-view input and a fast feed-forward Gaussian regression pipeline, it balances ease of capture with high-fidelity output. Future work will focus on addressing remaining challenges, such as filling "holes" in heavily occluded areas (e.g., armpits) and improving real-world scale estimation without camera metadata.
Summary updated