Summary of Seppo: Semi-policy Preference Optimization For Diffusion Alignment, by Daoan Zhang et al.
SePPO: Semi-Policy Preference Optimization for Diffusion Alignment
by Daoan Zhang, Guangchen Lan, Dong-Jun Han, Wenlin Yao, Xiaoman Pan, Hongming Zhang, Mingxiao Li, Pengcheng Chen, Yu Dong, Christopher Brinton, Jiebo Luo
First submitted to arxiv on: 7 Oct 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Reinforcement learning from human feedback (RLHF) methods are being explored to fine-tune diffusion models (DMs) for visual generation tasks. However, on-policy strategies have limitations due to the generalization capability of the reward model, while off-policy approaches require large amounts of paired human-annotated data. To address these limitations, we propose a preference optimization method called Semi-Policy Preference Optimization (SePPO). SePPO leverages previous checkpoints as reference models and generates on-policy reference samples that replace “losing images” in preference pairs, allowing for optimization using only off-policy “winning images.” We also design a strategy for reference model selection to expand the exploration in the policy space. Our approach mitigates performance degradation caused by uncertainty in reference sample quality. We validate SePPO across text-to-image and text-to-video benchmarks, surpassing previous approaches on text-to-image tasks and demonstrating outstanding performance on text-to-video tasks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Imagine a way to teach computers to generate images or videos based on what people like or dislike. This is called reinforcement learning from human feedback (RLHF). But there’s a problem: existing methods have limitations. To solve this, researchers propose a new method called SePPO (Semi-Policy Preference Optimization). It uses old versions of the computer program as references and generates new images that are similar to what people like or dislike. This way, the computer can learn from what people prefer without needing lots of paired human-annotated data. The new method is tested on tasks such as generating images based on text descriptions and videos based on text prompts. It performs better than previous methods and shows great potential for future applications. |
Keywords
» Artificial intelligence » Generalization » Optimization » Reinforcement learning from human feedback » Rlhf
