Reproduce “Can We Generate Images with CoT? Let's Verify and Reinforce Image Generation Step by Step” on OpenResearch

About this paper

Ziyu Guo, Renrui Zhang, Chengzhuo Tong, Zhizheng Zhao, Rui Huang, Haoquan Zhang, Manyuan Zhang, Jiaming Liu and 4 more

This work comprehensively investigates applying Chain-of-Thought reasoning strategies to autoregressive image generation, demonstrating how verification and reinforcement mechanisms can significantly enhance image quality and text-to-image alignment. The proposed approach achieves a 24% improvement over the Show-o baseline on the GenEval benchmark, surpassing Stable Diffusion 3 by 15%.

Key insights

• Integrating CoT-inspired verification and reinforcement strategies during both training and inference can substantially improve autoregressive image generation quality and text alignment, presenting a new paradigm beyond architectural scaling.
• The development of specialized reward models, such as PARM, is crucial for effectively assessing and guiding image generation at intermediate steps, as generic reward models struggle with the unique characteristics of visual intermediates (e.g., blurriness).
• Combining preference alignment (via DPO) during training with reward-model-guided test-time verification and reflection yields greater performance gains than either strategy applied in isolation.

Summary by alphaXiv.