Core Concepts
DreamView enables customizable and consistent text-to-3D generation by adaptively injecting overall and view-specific text guidance into a diffusion model.
Abstract
The paper proposes DreamView, a text-to-image generation model that can be lifted to 3D object generation. DreamView enables viewpoint customization while maintaining instance-level consistency by collaborating view-specific text and overall text via an adaptive guidance injection module.
The key highlights are:
DreamView-2D: A text-to-image model that can generate customized images from different viewpoints by adaptively injecting overall and view-specific text guidance. This is achieved through an adaptive guidance injection module that dynamically selects the appropriate text guidance for each U-Net block.
DreamView-3D: The text-to-image model is lifted to 3D generation by using score distillation sampling, inheriting the customization and consistency capabilities from DreamView-2D.
Extensive experiments demonstrate the advancement of DreamView in text-to-3D generation, providing a versatile and personalized approach to producing consistent and customizable 3D assets.
A user study shows that 74.5% of users prefer DreamView-3D over other methods in terms of generating 3D assets consistent with text descriptions, and 67.9% of users like DreamView-3D the best.
Stats
Text-to-image generation models like SD-v2.1 and MVDream achieve Inception Scores of 15.3/15.6 and 13.2/13.1, respectively, while DreamView-2D achieves 14.5.
DreamView-3D takes around 55 minutes to generate a 3D asset on a single A100 GPU, while DreamFusion, Magic3D, and SJC take around 30 minutes, MVDream takes about 50 minutes, and ProlificDreamer takes ~180 minutes.
Quotes
"DreamView empowers artists to design 3D objects creatively, fostering the creation of more innovative and diverse 3D assets."
"DreamView enables viewpoint customization while maintaining instance-level consistency by collaborating view-specific text and overall text via an adaptive guidance injection module."