February 26, 2024Open Access

Feedback Efficient Online Fine-Tuning of Diffusion Models

View Full Paper

Ask AI

Bookmark

Authors

MUMasatoshi UeharaChan Zuckerberg Initiative (United States)YZYulai ZhaoHefei University of Technology KBKevin BlackIdaho State University

Discussion

Member takes

Implication

Theoretical analysis demonstrates feedback-efficient online fine-tuning in diffusion models, highlighting provable regret guarantees across images, biological sequences, and molecules.

Key Points

A novel reinforcement learning procedure enables efficient exploration directly on the manifold of feasible samples, optimizing generative outputs without producing invalid data.
The proposed approach provides a theoretical regret guarantee and optimizes online fine-tuning across three benchmark domains, spanning images, biological sequences, and molecules.
Restricting exploration to feasible samples substantially lowers query demands on the reward function, which supports robust property maximization in complex diffusion models.

Cite This Study

Uehara et al. (2024) studied this question.

synapsesocial.com/papers/68e779ebb6db6435876eea9f https://doi.org/10.48550/arxiv.2402.16359

View Full Paper

Ask AI

Bookmark

Implication

Theoretical analysis demonstrates feedback-efficient online fine-tuning in diffusion models, highlighting provable regret guarantees across images, biological sequences, and molecules.

Feedback Efficient Online Fine-Tuning of Diffusion Models

Discussion

Member takes

Implication

Key Points

Cite This Study

Feedback Efficient Online Fine-Tuning of Diffusion Models

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider