Applying AIGC in E‑commerce: Product Copy and Image Generation with Large Language Models

The presentation introduces recent AIGC (AI‑generated content) work applied to e‑commerce, beginning with an overview of the 2022 AI breakthroughs such as OpenAI's ChatGPT and the rapid development of large language models.

It then examines product copy generation, explaining why automated copy is essential for attracting shoppers, reducing manual effort, and maintaining SEO quality. The problem is defined as generating titles and descriptions from known product highlights. Early attempts used template‑based pipelines with knowledge graphs, which were rigid. A newer approach leverages GPT‑style generative models, fine‑tuned on e‑commerce data, achieving more fluent and diverse copy. The article also reviews GPT fundamentals, including next‑token prediction, transformer architecture, and the evolution from GPT‑2 to GPT‑4, as well as training stages (SFT, reward modeling, RLHF).

Experimental phases are described: Phase 1 (template‑based), Phase 2 (GPT‑2 fine‑tuning), and Phase 3 (two‑stage fine‑tuning with data generation and ranking). Case analyses show that the refined models better capture distinctive selling points and align with human evaluation.

The second major part focuses on product image generation. It outlines the need for AI‑generated images to cut photography costs and accelerate iteration. Milestones such as DALL‑E 1/2, Stable Diffusion, and LoRA are highlighted. Technical details of diffusion models, latent diffusion, and the integration of CLIP with cross‑attention are explained, emphasizing how these architectures enable high‑quality image synthesis on consumer‑grade hardware.

Subsequent advancements (Textual Inversion, DreamBooth, ControlNet, LoRA) are mentioned, followed by practical e‑commerce applications: generating product images from keyword prompts, reducing reliance on photographers, and streamlining the production pipeline. Limitations for high‑resolution images and quality control are acknowledged.

The conclusion recaps the surveyed AI advances, summarizes the company's GPT‑based copy generation and Stable Diffusion‑based image generation experiments, and looks ahead to further model iterations and broader industry impact.

A Q&A section addresses common concerns, including model size versus capability, incorporating stylistic prompts, handling hallucinations, data quality importance, hardware requirements for >100 B‑parameter models, Chinese pre‑trained models, and potential e‑commerce pain points solvable by AIGC.