DreamReward: Text-to-3D Generation with Human Preference
핵심 개념
DreamReward proposes a comprehensive framework to improve text-to-3D models by learning from human preference feedback, resulting in high-fidelity and aligned 3D results.
초록
Introduction:
Recent advancements in diffusion models have accelerated automated 3D generation.
Two principal categories of 3D creation: inference-only native methods and optimization-based lifting methods.
DreamReward Framework:
DreamReward introduces a novel framework for text-to-3D generation aligned with human preferences.
Reward3D model is trained on a diverse annotated 3D dataset to evaluate the quality of generated 3D content effectively.
DreamFL Algorithm:
DreamFL optimizes multi-view diffusion models using a redefined scorer to enhance human preference alignment in 3D results.
Experiments:
Comparative experiments show that DreamReward outperforms baseline models across various evaluation criteria.
User studies demonstrate high alignment, quality, and consistency of DreamReward-generated 3D assets with human preferences.
DreamReward
통계
25k expert comparisons based on systematic annotation pipeline.
인용구
"RLHF has shown success in improving generative models."
"DreamReward successfully aligns text-to-3D generation with human intention."