Reproduce “AutoDecoding Latent 3D Diffusion Models” 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
Evangelos Ntavelis, Aliaksandr Siarohin, Kyle Olszewski, Chaoyang Wang, Luc Van Gool, Sergey Tulyakov
A two-stage framework integrates a 3D autodecoder with a latent diffusion model to synthesize high-quality static and articulated 3D assets directly from 2D images or monocular videos. This method achieves superior performance on FID and KID metrics compared to previous GAN-based and diffusion-based approaches on various datasets and scales to diverse object categories.
Key insights
- Learning rich latent 3D spaces from readily available 2D observations (images or monocular videos) can circumvent the need for explicit 3D supervision, enabling more scalable 3D generation.
- An intermediate volumetric latent space within a 3D autodecoder, combined with robust normalization, provides a suitable compact representation for efficient and stable latent 3D diffusion training.
- Employing multi-frame training and foreground supervision significantly improves the quality of 3D reconstructions, especially for real-world video datasets with complex backgrounds and foregrounds.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.