Title: ProductWebGen: Evaluating Multimodal Generation of Product Webpages

Abstract:

The creation of product showcase webpages, derived from source images alongside specific layout and visual directives, offers substantial utility for e-commerce, advertising, and marketing sectors. This process requires rigorous visual consistency in product representation and precise adherence to instructions to produce functional HTML code. These demands for controllability and instruction compliance align closely with the capabilities of sophisticated multimodal generative systems, including unified models and image editing tools. To address this, we introduce ProductWebGen, a benchmark designed to systematically evaluate the product webpage generation abilities of such models.

ProductWebGen comprises 500 test cases spanning 13 distinct product categories. Each instance includes a source image, a visual content directive, and a webpage layout instruction. The objective is to generate a comprehensive product showcase page featuring multiple consistent images that align with both the source material and the provided instructions.

Due to the mixed-modality nature of the inputs and outputs, we established and compared two distinct evaluation workflows. The first, termed "editing-based," utilizes large language models (LLMs) and image editing models independently to generate HTML code and images, respectively. The second, "UM-based," employs a single unified model (UM) to generate both components, with image creation conditioned on prior multimodal context.

Our empirical findings indicate that editing-based methods excel in following webpage instructions and enhancing content appeal. Conversely, UM-based approaches demonstrate superior performance in meeting visual content instructions. Additionally, we developed ProductWebGen-1k, a supervised fine-tuning dataset consisting of 1,000 pairs of real product images and LLM-generated HTML code. We validated the efficacy of this dataset using the open-source UM BAGEL. The associated code and data are accessible at https://github.com/SJTU-DENG-Lab/ProductWebGen.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC