A sparse DiT applies transformer-based generative modeling to sparse latent tokens associated with occupied or active 3D locations.

DiT
Diffusion Transformer
Domain
Sparse 3D latents
Purpose
Generative denoising / flow prediction
Related system
TRELLIS.2

Why sparse computation

Most cells in a large 3D grid do not describe a surface. Sparse processing avoids treating all empty and active regions as equally expensive.

Structure before detail

A staged system can first establish sparse structure, then predict detailed shape and material latents conditioned on that structure and the input image.