Reproduce “Compress3D: a Compressed Latent Space for 3D Generation from a Single Image” 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
Bowen Zhang, Tianyu Yang, Yu Li, Lei Zhang, Xi Zhao
Compress3D introduces a two-stage diffusion framework for single-image 3D generation, predicting a 3D-aware shape embedding to condition the generation of compressed triplane representations. This approach generates high-quality 3D models in 7 seconds with a 10x smaller latent space, achieving a FID score of 53.21 while requiring significantly less training data than prior methods.
Key insights
- Efficient 3D generation from single images necessitates both highly compressed latent representations and enriched, 3D-aware conditioning.
- A learnable 3D-to-2D projection combined with 3D-aware cross-attention in the encoder is crucial for preventing information loss and enhancing the expressive power of compressed triplane representations.
- Dual conditioning on both image and predicted shape embeddings provides the generative model with richer structural priors, leading to more consistent and accurate 3D model generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.