Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dynamic Product Image Generation and Recommendation at Scale for Personalized E-commerce

Ádám Tibor Czapp, Mátyás Jani, Bálint Domián, Balázs Hidasi

TL;DR

The paper tackles the scalability challenge of creating engaging, personalized product imagery for online retargeting by marrying latent diffusion-based background generation with contextual bandits. It leverages Stable Diffusion for environment backgrounds, ControlNet to constrain generation around the product edges, and LinUCB for category-specific prompt personalization, forming a production-ready pipeline with caching and latency considerations. The key contributions include a robust background-generation workflow, artifact-reduction techniques, and an empirically validated online A/B program showing CTR uplifts and downstream conversion benefits. The work demonstrates tangible, scalable improvements in user engagement for dynamic product ads in e-commerce, enabling personalized visuals at catalog scale.

Abstract

Coupling latent diffusion based image generation with contextual bandits enables the creation of eye-catching personalized product images at scale that was previously either impossible or too expensive. In this paper we showcase how we utilized these technologies to increase user engagement with recommendations in online retargeting campaigns for e-commerce.

Dynamic Product Image Generation and Recommendation at Scale for Personalized E-commerce

TL;DR

The paper tackles the scalability challenge of creating engaging, personalized product imagery for online retargeting by marrying latent diffusion-based background generation with contextual bandits. It leverages Stable Diffusion for environment backgrounds, ControlNet to constrain generation around the product edges, and LinUCB for category-specific prompt personalization, forming a production-ready pipeline with caching and latency considerations. The key contributions include a robust background-generation workflow, artifact-reduction techniques, and an empirically validated online A/B program showing CTR uplifts and downstream conversion benefits. The work demonstrates tangible, scalable improvements in user engagement for dynamic product ads in e-commerce, enabling personalized visuals at catalog scale.

Abstract

Coupling latent diffusion based image generation with contextual bandits enables the creation of eye-catching personalized product images at scale that was previously either impossible or too expensive. In this paper we showcase how we utilized these technologies to increase user engagement with recommendations in online retargeting campaigns for e-commerce.
Paper Structure (3 sections, 4 figures)

This paper contains 3 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Generating product images
  3. Online tests

Figures (4)

  • Figure 1: Mild and extreme artifacts produced by inpainting.
  • Figure 2: Examples of color variations through conditioning.
  • Figure 3: Main steps of the background generation pipeline.
  • Figure 4: Relative CTR gains. (top left) phase I: positioned product on white & generated background vs. original image; (top right) phase II: improved vs. old pipeline; (bottom left) phase II: generated image (improved pipeline) vs. original image; (bottom right) phase III: personalized vs. non-personalized background.