This article discusses an innovative approach to 3D object generation using proxy-guided diffusion conditioning. The primary focus is on enhancing the interactive generation workflow alongside volume-conditioned reconstruction techniques. The authors detail their experimental setup, including the dataset used and the training parameters. Key evaluations compare proxy-based with image-based 3D generation methods, emphasizing the advantages of proxy-guided systems for controllable object generation. The findings suggest improved performance in generating 3D objects, with supplementary materials supporting the findings through implementation details and user studies.
In our experiment, we use the LVIS subset of Objaverse to train the model, which contains 28,000+ objects after a heuristic cleanup process.
We set up 16 image views with -30° pitch, evenly facing the object from 360°, enhancing the training view rendering.
The source code will be released upon acceptance of this paper, ensuring wider access and reproducibility for the research community.
The implementation details highlighted in the supplementary sections provide crucial insights into our training methods and experimental setup.
Collection
[
|
...
]