EAGLE-GRPO: Element-Aware Group Learning for E-Commerce Image Generation

A novel approach known as EAGLE-GRPO (Element-Aware Group Learning for E-Commerce Image Generation) has been introduced to enhance the quality of prompts utilized in generating images for e-commerce. This technique tackles the limitations of prompt quality found in contemporary image generation and editing systems. While vision-language models (VLMs) can create image-editing prompts based on product images and their metadata, improving their prompt generation necessitates post-training that incorporates feedback from the resulting images. Group Relative Policy Optimization (GRPO) serves as a suitable framework for optimizing rewards at the outcome level, but it only attributes credit at the full-prompt level, despite the fact that image quality is often influenced by specific design aspects like composition and background. Existing methods for fine-grained credit assignment typically demand step-level supervision or learned critics. To overcome this issue, EAGLE-GRPO breaks down the group-centered reward across predefined elements, framing element-level credit assignment as a multi-task learning challenge. This methodology is elaborated in a paper available on arXiv (arXiv:2608.00584), which was announced as a cross-type submission. The goal is to improve the performance of VLMs in producing high-quality e-commerce visuals by offering more detailed feedback during the training process.