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

Forum topic · 小凯 · 2026-10-16

Summary

CommerceGen-Studio (arXiv:2610.20603) is an end-to-end framework for generating commercial-grade main images for e-commerce Stock Keeping Units (SKUs). The system follows a three-stage pipeline: information extraction, semantic rewriting, and multi-condition generation. PERL-Agent distills heterogeneous SKU metadata into concise selling-point text lists via an iterative VLM/LLM closed loop scored by a reference-free rubric. Layout Thinker Optimization (LTO) aligns the rewriter with both caption quality and rendered-image quality through alternating domain-specific policy optimization. SKU-DiT, trained with Reinforce-Distill-Reinforce (RDR) progressive training, jointly conditions on the white-background product image, trademark crop, and structured layout to preserve subject fidelity, trademark integrity, and character-level text accuracy. The framework runs at production scale of millions of images per day with per-image latency of only a few seconds. The authors also release CommerceGen-Bench, a benchmark of hard cases from live traffic. On this benchmark, the framework achieves the highest total score among evaluated open-source baselines in both editing tracks, matches larger closed-source models, and surpasses several closed-source models on instruction following and product integration.

Paper Overview

Field: Computer Vision (CV) Authors: Xiaoli Xu, Zhengtao Wu, Zhengze Xu, Tao Liu, Yongchao Du Published: 2026-10-14 arXiv: 2610.20603

Abstract

We present CommerceGen-Studio, an end-to-end framework for generating commercial-grade main images for e-commerce Stock Keeping Units (SKUs). Our framework is organized along a three-stage pipeline of information extraction, semantic rewriting, and multi-condition generation:

PERL-Agent (Prompt-Eval-Refine-Loop Agent) distills heterogeneous SKU metadata into concise selling-point text lists through an iterative closed loop of VLM and LLM roles scored by a reference-free rubric.
Layout Thinker Optimization (LTO) aligns the rewriter with both caption quality and rendered-image quality via alternating domain-specific policy optimization.
SKU-DiT, trained under Reinforce-Distill-Reinforce (RDR) progressive training, conditions jointly on the white-background product image, the trademark crop, and the structured layout, while maintaining subject fidelity, trademark integrity, and character-level text accuracy.

The framework runs at production scale of millions of main images per day with per-image latency of only a few seconds — to our knowledge, among the first systems of this kind operating at commercial scale.

Benchmark

The authors also release CommerceGen-Bench, a benchmark collected from live traffic covering hard cases across all product categories. On this benchmark, the framework:

Achieves the highest total score among evaluated open-source baselines in both editing tracks, with and without prompt enhancement.
Matches larger closed-source models overall.
Surpasses several closed-source models on instruction following and product integration — the two attributes that determine whether a generated image is publishable.