Reproduce “Splatshot: 3D Face Avatar Generation from a Single Unconstrained Photo” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hao Liang, Zhixuan Ge, Soumendu Majee, Joanna Li, Ashok Veeraraghavan, Guha Balakrishnan
SplatShot reconstructs photorealistic and geometrically consistent 3D face avatars from a single unconstrained photo by iteratively coupling 3D Gaussian Splatting with 2D diffusion models. This training-free framework achieves leading performance in identity fidelity, multi-view consistency, and perceptual quality compared to existing single-image and diffusion-based methods.
Key insights
- Effectively combining 2D diffusion models with explicit 3DGS within an iterative denoising framework can leverage their complementary strengths for both photorealism and multi-view consistency.
- Direct geometric feedback from 3DGS during the diffusion process resolves per-view inconsistencies, guiding the 2D outputs towards a coherent 3D avatar.
- A noise mixture mechanism is critical for maintaining stability during early denoising steps, preventing corrupted predictions from destabilizing 3DGS fitting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.