MIT researchers have developed a new technique called Score Distillation to generate realistic 3D shapes for virtual reality, filmmaking and engineering design applications.
The Score Distillation activities allow the creation of 3D shapes by using 2D image generation models that appear more lifelike and more realistic.
MIT researchers have achieved high-quality 3D shapes using Score Distillation without additional training or complex post-processing.
The researchers identify a mismatch between a formula that forms part of the SDS process and its counterpart in 2D diffusion models as the root cause of low-quality 3D models.
Rather than randomly sampling the noise term, the approximation technique infers it from the current 3D shape rendering to generate sharp and realistic 3D shapes.
The method relies on a pre-trained diffusion model and inherits the biases and shortcomings, making it prone to hallucinations and other failures.
Future work can help facilitate the process of creating more realistic 3D shapes by improving the mathematical understanding of Score Distillation and other related techniques.
This work is funded by the Toyota Research Institute, the US National Science Foundation, and various other organizations.