Automated Image Generation for E-commerce Categories Using Multimodal LLMs

ebay · E-commerce · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: E-commerce

eBay needed to generate high-quality category and theme images across thousands of categories without relying on slow manual curation by designers.

Constraints:

• category breadcrumbs contain little visual context
• item titles contain non-visual metadata that can mislead image generation
• text-to-image outputs are non-deterministic and may contain distortions or random elements
• production images must follow eBay's visual standards and responsible AI principles

Architecture

eBay built a three-stage multimodal generation pipeline. The preprocessing stage samples popular item titles by click-through rate and uses LLMs to simplify titles, preserving visual theme details while removing brand, size, and model noise. A prompt-generation LLM combines category names and simplified titles into product-photography prompts, and a Large Vision Model creates category images. A multimodal LLM evaluator scores each image for prompt fidelity, detail clarity, and product-photography style. Images above threshold are kept; failures enter a refinement loop where the image, prompt, and rubric results are analyzed to produce an enhanced prompt until quality passes or a retry limit is reached.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  category-taxonomy[Category taxonomy] --> item-sampler[Item sampler]
  item-sampler --> title-simplification-llm[Title simplification LLM]
  category-taxonomy --> prompt-generation-llm[Prompt generation LLM]
  title-simplification-llm --> prompt-generation-llm
  prompt-generation-llm --> large-vision-model[Large Vision Model]
  large-vision-model --> multimodal-evaluator[Multimodal evaluator]
  multimodal-evaluator --> prompt-refinement-loop[Prompt refinement loop]
  prompt-refinement-loop --> prompt-generation-llm
  prompt-refinement-loop --> large-vision-model
  human-evaluation-sample[Human evaluation sample] --> multimodal-evaluator

Components

Data flow

For a category or theme, the pipeline selects popular item titles, simplifies them into visual descriptors, and builds a detailed image prompt from the category plus simplified titles. The vision model generates an image, then a multimodal evaluator scores it against fidelity, clarity, and photography-style criteria. Passing images are stored for use; failing images feed an optimization loop that analyzes the image, original prompt, and rubric failures to create a better prompt and regenerate the image until it passes or hits the retry limit.

Why this architecture

The pipeline separates context extraction, prompt creation, image synthesis, and quality control because category text alone is too sparse and generative images are variable. Item-title simplification removes noisy non-visual details before prompt generation. The evaluator/rubric layer acknowledges that image generation cannot be perfectly constrained, while the refinement loop turns failed outputs into targeted prompt improvements instead of forcing manual review for every category. Human sampling validates the evaluator rather than replacing it entirely.

Models used: multimodal LLM · Large Vision Model

Scale

• The system targets thousands of eBay categories.
  “To scale high-quality, thematically specific images for thousands of categories on eBay, we developed an automated image generation system”article, stated
• Human evaluation found 88% of automatically approved images suitable for eBay use.
  “In our assessment, human evaluators confirmed that 88% of the images were suitable for use on eBay, meaning they met all quality and thematic criteria required for accurate category representation.”article, stated
• The system uses click-through rate as the popularity signal when sampling category items.
  “We start by sampling popular item titles that belong to each category or theme. Popularity is determined by the click through rate.”article, stated

Alternatives considered (roads not taken)

Manual category-image design

Rejected because: Manual curation was time-intensive, resource-heavy, slow to adapt, and not scalable across thousands of categories.

Raw item titles as image-generation context

Rejected because: Item titles include non-visual metadata such as brands, sizes, and model numbers that could mislead the LLM.

Single-pass image generation

Rejected because: Single-pass systems do not use evaluator feedback to repair distortions, blurriness, or style mismatch.

Patterns

prompt-chaining (generation-workflow)

The system chains title simplification, prompt generation, image generation, evaluation, and prompt refinement.

llm-as-judge (evaluation)

A multimodal LLM scores generated images against rubric criteria before they are kept.

eval-driven-development (quality-control)

Evaluation results drive the refinement loop and future rubric improvements from false positives and false negatives.

guardrails-io-filtering (safety)

Threshold-based rubric screening excludes images that fail fidelity, clarity, or style expectations before production use.

Tradeoffs

Automated visual generation

Chose multimodal LLM and vision-model pipeline over manual designer-created category images.

Rationale: Automation can cover thousands of categories and update quickly, while manual curation is slow and resource-intensive.

Rubric thresholding before storage

Chose automated multimodal evaluation with pass/fail threshold over keeping every generated image or manually reviewing all images.

Rationale: The threshold blocks known failure modes while keeping the pipeline scalable.

Retry limit on refinement

Chose iterative prompt/image refinement with a predefined stopping point over unbounded retries until an image passes.

Rationale: The loop improves failed images but must avoid infinite compute consumption on difficult categories.

Gotchas (what bit them)

Raw category text lacked enough visual context.

Root cause: Categories are represented as simple breadcrumb text.

Resolution: The team enriched prompts with popular item titles sampled from the category.

Item titles contained misleading non-visual metadata.

Root cause: E-commerce titles include brand names, sizing, model numbers, and other details not useful for visual generation.

Resolution: LLMs simplify item titles by filtering non-visual elements before prompt generation.

Generated images can contain distortions and unrelated objects.

Root cause: Generative image models are non-deterministic and produce variable-quality outputs from the same prompt.

Resolution: The rubric-based multimodal evaluator filters bad images and triggers prompt refinement for retries.

Evaluation

eBay evaluates images with a multimodal rubric over fidelity, detail sharpness and clarity, and product-photography style, then validates automated approvals with human evaluators and studies false positives/false negatives for rubric improvement.

Metrics: fidelity to the prompt · detail sharpness and clarity · adherence to product photography style · 88% of automatically approved images suitable for eBay use

Outcomes

• The system enables scalable category and theme image creation across thousands of categories.
• Human evaluators judged 88% of automatically approved images suitable for eBay.
• The pipeline reduces manual curation effort while improving merchandising flexibility and visual consistency.

The case targets thousands of eBay categories and relies on automated evaluation plus selective human calibration rather than full manual review.