DreamReward-X: Boosting High-Quality 3D Generation with Human Preference Alignment.

Area of Science:

• Computer Vision
• Artificial Intelligence
• 3D Graphics

Background:

• Large-scale diffusion models have advanced 3D content generation.
• Aligning 3D content with human preferences, especially in text-driven scenarios, remains a significant challenge.

Purpose of the Study:

• To propose DreamReward, a novel framework for improving text-driven 3D generation using human preference feedback.
• To develop Reward3D, a general-purpose text-to-3D human preference reward model.
• To introduce the Reward3D Feedback Learning (DreamFL) algorithm for optimizing 3D generation models.

Main Methods:

• Collected over 25,000 expert comparisons through a systematic annotation pipeline (filtering, rating, ranking).
• Developed Reward3D, a human preference reward model for text-to-3D generation.
• Implemented the Reward3D Feedback Learning (DreamFL) algorithm to align generation with user prompts.
• Extended the framework to 4D and image-to-3D generation (DreamReward-4D, DreamReward-img).
• Proposed DreamReward++ with a reward-aware noise sampling strategy to enhance diversity and preference alignment.

Main Results:

• Achieved significant improvements in prompt alignment and generation fidelity.
• Demonstrated low-cost yet effective extensions for 4D and image-to-3D generation.
• DreamReward++ successfully generates high-fidelity, diverse 3D results, addressing diversity limitations.
• The reward-aware noise sampling strategy enhances text-driven diversity while maintaining human preference alignment.

Conclusions:

• Learning from human feedback is a powerful approach to enhance 3D generation models.
• DreamReward and its extensions offer a promising direction for high-quality and diverse 3D content creation.
• The proposed methods show significant potential for future advancements in text-driven 3D generation.