3D Diffusion Models

3D diffusion models learn distributions over three-dimensional objects and environments, enabling new content generation, completion of partial observations, and editing of existing geometry. They diffuse over voxels, point clouds, meshes, Gaussian primitives, or compact latent tokens, using denoising networks, diffusion transformers, and flow-matching methods, often conditioned on text, images, sketches, measurements, or spatial layouts. Applications span object benchmarks such as ShapeNet and Objaverse, driving scenes and LiDAR, medical MRI and CT, CAD, and molecular structures. A central challenge is scaling to detailed, large scenes while preserving topology, physical plausibility, and consistency across views; compact latents and coarse-to-fine generation help, but trade-offs in fidelity, control, data, and inference cost remain.