AI‑Driven Aesthetic Evaluation for E‑commerce Image Generation

This article describes a systematic approach to define, train, and apply aesthetic standards for evaluating AI‑generated images in the e‑commerce domain, especially for Taobao.

Four main steps:

Define aesthetic standards – combine AI‑generated image characteristics with traditional image‑quality criteria (composition, lighting, texture, realism, etc.).

Train aesthetic models – use the defined standards to build a scoring model that can automatically assign aesthetic scores (1‑5) and identify abnormal attributes such as poor background‑subject fusion, hand anomalies, facial defects, or body distortions.

Apply the models – guide Taobao’s AI image‑generation pipeline to improve the “good‑image” rate and provide fine‑grained feedback for model refinement.

Upgrade Taobao style models – create a style‑model library based on the style standards, enabling merchants to select from diverse, brand‑consistent visual styles.

The article also details the technical framework:

Use multimodal pre‑training and multi‑task fine‑tuning to predict aesthetic scores and abnormal‑attribute tags.

Leverage CLIP distillation for human‑model image evaluation, employing L1 and binary‑cross‑entropy losses.

Implement a 5‑point rating system with 19 detailed criteria, covering image composition and AI‑specific traits.

Training involves multiple validation rounds: human‑average scoring, AI‑human alignment checks, and iterative calibration to ensure consistency.

Testing phases assess both model accuracy and robustness across internal (QianNiu) and third‑party systems, using a balanced test set (easy, medium, hard) of 1,200 images.

Future plans include publishing the aesthetic standards in collaboration with the China Academy of Art, finalizing style standards, and launching an AI‑PaaS product to provide these capabilities to internal and external developers.