CommerceGen-Studio: An End-to-End Framework for Commercial-Grade Main-Image Generation

TL;DR

CommerceGen-Studio is a production-ready system that generates high-quality e-commerce main images at scale using a three-stage pipeline: metadata extraction, semantic refinement, and conditional image generation. It leverages a closed-loop agent for prompt optimization, layout-aware policy training, and a specialized diffusion model trained with a progressive RDR strategy to preserve product fidelity and text accuracy.

Why It Matters

The system enables automated, scalable generation of commercial-grade product images, reducing reliance on manual photography and design. Its ability to maintain trademark integrity and accurate text rendering makes it suitable for direct deployment in live e-commerce platforms.

Key Findings

• The PERL-Agent iteratively refines SKU metadata into concise selling points using vision-language and large language models, guided by a reference-free evaluation metric.
• Layout Thinker Optimization improves alignment between caption quality and image output through domain-specific policy updates in an alternating training scheme.
• SKU-DiT, trained via Reinforce-Distill-Reinforce, achieves strong subject fidelity, trademark accuracy, and text precision when conditioned on product images, trademarks, and layout structures.