Coarse3D
A coarse-stage model for image-to-3D generation that achieves direction-controllable and structurally faithful sparse voxel geometry generation through first-frame anchoring of a canonical space, a dual-tower DiT, and spatial-prior-guided attention.
Timeline (last 90 days)
Coarse3D is an image-to-3D generation model that, by improving the sparse structure generation stage, achieves direction-controllable voxel geometry highly consistent with the input image.
Coarse3D uses spatial priors to guide attention, regularizing cross-attention with pixel-to-voxel correspondences to further ensure consistency between the generated structure and the input image.
The paper's experiments show that Coarse3D leads by a large margin in sparse structure generation quality and can consistently enhance the final 3D generation result, achieving state-of-the-art performance.
Because Coarse3D can directly replace the coarse stage of existing pipelines, its improvement is plug-and-play.
In the future, Coarse3D technology may be extended to more complex scenes and dynamic object generation.
All Facts (5)
Coarse3D is an image-to-3D generation model that, by improving the sparse structure generation stage, achieves direction-controllable voxel geometry highly consistent with the input image.
50%UnverifiedCoarse3D uses spatial priors to guide attention, regularizing cross-attention with pixel-to-voxel correspondences to further ensure consistency between the generated structure and the input image.
50%UnverifiedThe paper's experiments show that Coarse3D leads by a large margin in sparse structure generation quality and can consistently enhance the final 3D generation result, achieving state-of-the-art performance.
50%UnverifiedBecause Coarse3D can directly replace the coarse stage of existing pipelines, its improvement is plug-and-play.
50%UnverifiedIn the future, Coarse3D technology may be extended to more complex scenes and dynamic object generation.
50%