May 29, 2024Open Access

Tuning-Free Alignment of Diffusion Models with Direct Noise Optimization

View Full Paper

Ask AI

Bookmark

Authors

ZTZhiwei TangUniversity of Electronic Science and Technology of China JPJiangweizhi Peng JTJiasheng TangAlibaba Group (United States)

Discussion

Member takes

Implication

Computational study demonstrates state-of-the-art reward alignment in diffusion models, highlighting an effective tuning-free method to prevent reward hacking.

Key Points

Direct noise optimization aligns diffusion models with continuous reward functions during sampling, producing state-of-the-art reward scores and high image quality.
Evaluation across human feedback reward functions optimizes the injected noise in an online fashion, incorporating probability regularization to prevent reward hacking.
Supports tuning-free and prompt-agnostic alignment within reasonable time budgets, while providing theoretical extensions for non-differentiable reward functions.

Cite This Study

Tang et al. (2024) studied this question.

synapsesocial.com/papers/68e67f72b6db643587608ff0 https://doi.org/10.48550/arxiv.2405.18881

View Full Paper

Ask AI

Bookmark

Implication

Computational study demonstrates state-of-the-art reward alignment in diffusion models, highlighting an effective tuning-free method to prevent reward hacking.

Tuning-Free Alignment of Diffusion Models with Direct Noise Optimization

Discussion

Member takes

Implication

Key Points

Cite This Study

Tuning-Free Alignment of Diffusion Models with Direct Noise Optimization

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider