Event
Paper proposes directly generating 3D scenes in 3D latent space
1 reports · 1 sources
A paper proposes directly generating 3D scenes in a high-dimensional implicit 3D latent space derived from 2D foundation models (such as DINOv2, SigLIP2, DA3), and modeling in that latent space with a diffusion Transformer to achieve 3D-grounded scene generation. The paper's authors claim that the method outperforms 2D-based methods in efficiency and spatial consistency, and exhibits scalability in representation and model capacity.
Report timeline
- Any 3D Scene is Worth 1K Tokens: A Paper on Scene Generation in 3D Latent Space
The paper proposes performing 3D scene generation directly in a high-dimensional implicit 3D latent space derived from 2D foundation models (such as DINOv2, SigLIP2, and DA3), using a diffusion Transformer to model within that 3D latent space and achieve 3D-grounded scene generation.
HuggingFace Daily Papers ·