ProductWebGen: Benchmarking Multimodal Product Webpage Generation

Researchers introduce ProductWebGen, a benchmark dataset and evaluation framework for assessing multimodal AI models' ability to generate e-commerce product webpages from images and textual instructions. The study compares two approaches—using separate image editing and language models versus unified multimodal models—and releases a 1,000-sample fine-tuning dataset to advance webpage generation capabilities.

ProductWebGen addresses a practical gap in multimodal AI evaluation by focusing on a real-world e-commerce use case that requires both strict visual consistency and precise instruction following. The benchmark's design reflects the growing sophistication of generative AI systems, which must now coordinate multiple modalities—image editing, HTML generation, and visual content creation—to produce coherent outputs. This work matters because product webpage generation directly impacts how businesses present merchandise online, making it a valuable testbed for production-grade AI systems.

The research reveals important trade-offs between different architectural approaches. Editing-based workflows, which separate HTML and image generation, show stronger performance in webpage structure and visual appeal, while unified multimodal models demonstrate advantages in interpreting visual content instructions. This finding suggests neither approach is universally superior, indicating that practitioners must choose based on specific use-case priorities. The construction of ProductWebGen-1k with real product images and LLM-generated HTML code provides a practical resource for fine-tuning open-source models, lowering barriers to adoption for smaller organizations.

For the AI industry, this benchmark contributes to standardizing multimodal evaluation in e-commerce contexts, an increasingly important domain as retailers seek automation solutions. The release of code and datasets democratizes access to these capabilities, accelerating development cycles for downstream applications. The study's systematic comparison methodology establishes a template for evaluating complex, multi-step generative tasks beyond simple image or text generation.

• →ProductWebGen benchmark tests multimodal AI systems on e-commerce webpage generation using 500 test samples across 13 product categories.
• →Editing-based workflows outperform unified models in HTML instruction following and visual appeal, while unified models excel at visual content interpretation.
• →The released ProductWebGen-1k dataset (1,000 samples) enables fine-tuning of open-source multimodal models for practical deployment.
• →The research demonstrates that different architectural approaches offer complementary strengths, with no universally optimal solution for all webpage generation requirements.
• →This benchmark advances standardization of multimodal AI evaluation in e-commerce, a domain increasingly adopting generative AI for automation.