Models combines sparse structures with powerful visual features, enabling efficient and detailed 3D modeling, what can be conditioned on text prompts or images, making it highly flexible for various applications.

3D Alchemy uses a two-step generation process. First, it creates the sparse structure, and then it fills in latent vectors for non-empty voxels. We use rectified flow transformers as the backbone, optimized to handle the sparsity.

Model places local latents on active voxels that intersect the object's surface. These latents are encoded by fusing image features from multiple rendered views of the 3D model, capturing fine geometric and visual details. This approach enhances the coarse structure provided by the voxels.