Our paper Luciddreamer is selected as Spotlight at CVPR 2024 and featured on Hugging Face Daily
We are thrilled to announce that our recent work, LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching, has been selected as a spotlight paper at the prestigious CVPR 2024 conference. This recognition is a testament to the innovative breakthroughs the paper introduces in the field of generative models.
LucidDreamer has been designed to transform textual descriptions directly into high-quality 3D models, overcoming previous challenges that limited the detail and accuracy of generated 3D content. The key to our approach lies in a novel method named Interval Score Matching (ISM), which ensures the generation of detailed, high-quality 3D models by effectively using pre-trained 2D diffusion models as a strong image prior.
Highlights of Our Work:
- Innovative Technique: Our Interval Score Matching significantly enhances the text-to-3D generation process, allowing for the creation of photorealistic 3D models with unprecedented detail and fidelity.
- Efficient and Effective: Compared to existing methods, LucidDreamer reduces training costs and complexity while delivering superior results, as evidenced by extensive experiments and qualitative assessments.
- Community and Industry Impact: Since its feature as a daily paper on Hugging Face, our framework has attracted attention from both the academic community and industry practitioners, opening up new possibilities for creative and commercial applications.
The code for LucidDreamer is available to the public under the EnVision-Research GitHub repository, enabling researchers and developers to build upon our work.
Authors:
- Yixun Liang, My student from HKUST(GZ)
- Xin Yang, My student from HKUST(GZ)
- Jiantao Lin, My student from HKUST(GZ)
- Haodong Li, My student from HKUST(GZ)
- Xiaogang Xu, Collaborator from Zhejiang University
- Ying-Cong Chen, HKUST(GZ)
