Ecommerce ops · Production

Enhancing eBay's Visual Shopping Experience With Automated Image Generation And Optimization For Themes and Categories

The problem

eBay category pages were represented by plain text rather than images; generating a category image required a curator or designer to manually create one—a time-intensive process requiring extensive resources that could not quickly adapt to emerging trends or scale to thousands of categories.

First attempt

In initial trials of the automated generation system, non-deterministic generative models produced problematic variations including distorted objects, unexpected random objects, and dark or blurred images that did not meet quality standards.

Workflow diagram · grounded in source
1
Sample popular item titles
trigger
“We start by sampling popular item titles that belong to each category or theme. Popularity is determined by the click through rate.”
2
LLM title simplification
ai_action
“Since item titles often include extraneous non-visual details (e.g. brand names, sizing information or model numbers) that could mislead the LLM in prompt creation, we utilize LLMs to simplify these titles. By filtering out non-visual el…”
3
LLM image prompt generation
ai_action
“we use an LLM to generate image prompts that capture the category's theme. Using the category name and simplified titles as input, the LLM produces detailed prompts specifying the subject, style, and background of the image to be generat…”
4
Vision model image generation
ai_action
“The LLM-generated prompts are passed to a Large Vision Model, which produces representative images for the specified category.”
5
Multimodal quality assessment
validation
“A novel multimodal LLM-driven evaluation framework automatically assesses each generated image to ensure high quality and thematic alignment before storage. Each image is scored in these categories and images that meet or exceed a predef…”
6
Iterative prompt refinement
feedback_loop
“In cases where an image does not meet the required threshold, the system uses a multimodal LLM to analyze the image, prompt and the evaluation results, to generate an improved prompt. This enhanced prompt is then used to create a new ima…”
Reported outcome

Human evaluators confirmed that 88% of AI-generated images were suitable for use on eBay, demonstrating a scalable approach to producing high-quality, category-specific visuals with minimal manual intervention and reduced resource demands compared to traditional curation.

Reported metrics
images suitable for use on eBay (human assessment)88%
Image quality improvementsignificantly reduces distortion, blurriness, and stylistic mismatches
Manual intervention in image creationminimal manual intervention
Resource demands of traditional image curationreducing the resource demands of traditional image curation
Reported stack
Multimodal Generative AImultimodal LLMsLarge Vision Model
Source
https://www.linkedin.com/pulse/enhancing-ebays-visual-shopping-experience-automated-image-galsurkar-9pgle/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Human evaluators confirmed that 88% of AI-generated images were suitable for use on eBay, demonstrating a scalable approach to producing high-quality, category-specific visuals with minimal manual intervention and red…

What tools did this team use?

Multimodal Generative AI, multimodal LLMs, Large Vision Model.

What results were reported?

images suitable for use on eBay (human assessment): 88%; Image quality improvement: significantly reduces distortion, blurriness, and stylistic mismatches; Manual intervention in image creation: minimal manual intervention; Resource demands of traditional image curation: reducing the resource demands of traditional image curation (source-reported, not independently verified).

What failed first in this deployment?

In initial trials of the automated generation system, non-deterministic generative models produced problematic variations including distorted objects, unexpected random objects, and dark or blurred images that did not…

How is this ecommerce ops AI workflow structured?

Sample popular item titles → LLM title simplification → LLM image prompt generation → Vision model image generation → Multimodal quality assessment → Iterative prompt refinement.

Related ecommerce ops cases