This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Inversion-DPO: Precise and Efficient Post-Training for Diffusion Models
Created by
Haebom
Author
Zejian Li, Yize Li, Chenye Meng, Zhongni Liu, Yang Ling, Shengyuan Zhang, Guang Yang, Changyuan Yang, Zhiyuan Yang, Lingyun Sun
Outline
In this paper, we propose a new framework for alignment of diffusion models (DMs), Inversion-DPO. Existing methods have problems such as high computational cost for reward model learning and low model accuracy and training efficiency. Inversion-DPO does not require a reward model by reconstructing Direct Preference Optimization (DPO) using DDIM inversion. We propose a new post-training paradigm by replacing complex posterior probability sampling in DPO with deterministic inversion. This significantly improves accuracy and efficiency, and we experimentally demonstrate that it outperforms existing methods by applying it to text-to-image and composite image generation tasks. For post-training for composite image generation, we utilize a paired dataset with 11,140 complex structural annotations and comprehensive scores.
Takeaways, Limitations
•
Takeaways:
◦
A novel method for efficiently and accurately performing alignment of diffusion models without compensation model is presented.
◦
Reduce computational costs and improve model performance by reconstructing DPO using DDIM inversion
◦
Extending the applicability of diffusion models to complex generative tasks such as composite image generation.
◦
Improved ability to produce high-quality, complex, consistent images
•
Limitations:
◦
Further research is needed on the generalization performance of the proposed method.
◦
Need to verify applicability to various types of production tasks
◦
Limitations of the presented dataset and need for performance evaluation on other datasets
