CommerceGen-Studio, an end-to-end framework described in a recent arXiv paper, targets commercial-grade main images for e-commerce SKUs. The authors organize the system as a three-stage pipeline: information extraction, semantic rewriting, and multi-condition generation. In the first stage, a Prompt-Eval-Refine-Loop Agent (PERL-Agent) iterates between VLM and LLM roles to distill heterogeneous SKU metadata into compact selling-point text lists, with a reference-free rubric scoring each iteration. The second stage, Layout Thinker Optimization (LTO), uses alternating domain-specific policy optimization to align the rewriter with both caption quality and rendered-image quality. Finally, SKU-DiT, trained with a Reinforce-Distill-Reinforce scheme, conditions jointly on the white-background product image, a trademark crop, and a structured layout, aiming to preserve subject fidelity, trademark integrity, and character-level text accuracy.

According to the paper, the framework runs in production, generating millions of main images per day with a latency of a few seconds per image. The authors also release CommerceGen-Bench, a benchmark of hard cases collected from live traffic across all product categories.

On CommerceGen-Bench, the framework achieves the highest overall scores among the evaluated open-source baselines on both editing tracks, with and without prompt enhancement. The authors report that it matches much larger closed-source models and beats several of them at Instruction Following and Product Integration, which they identify as the two properties that determine whether an image can be published. As the sole source, these claims come directly from the paper and are not independently verified.