uses Style Transfer, Few-Shot and LoRA for product description

This post records an experiment in e-commerce product description.

The goal is to input product images and produce product descriptions that match the tone of the brand.

Three different control methods

The most direct way is to describe the desired style and tone in the prompt words.

This method is the cheapest and easiest to start, but the abstract concept of “brand tone” is difficult to fully convey through words alone.

The second method is Few-Shot Prompting.

The method is to put multiple examples into the prompt words and let the model imitate the sentence patterns, word usage and narrative style from the examples.

Generally, the more examples there are, the easier it is for the model to capture the target style, but each execution must be accompanied by these examples, and the inference cost will also increase.

The third method is LoRA Fine-Tuning.

LoRA does not directly modify the entire base model, but instead trains an additional set of parameter layers to allow the model to produce stable changes in behavior on a specific task or style.

It requires a relatively tedious data preparation and training process, but after the training is completed, the Adapter Layers can be saved and reused later.

Experimental Observation

The original experiment used Everlane’s product information to fine-tune Llama 3.2 and compare the output before and after fine-tuning.

The prompt words in the post were fairly simple, but in the Zero-Shot scenario, there was a noticeable difference in the fine-tuned description.

The original post also attached a Google Colab example, allowing readers to load the fine-tuned LoRA model, read a product from the test data, and compare the responses before and after fine-tuning.

The value of this saved content is not just to show a certain model output, but to compare “prompt word control”, “example control” and “model fine-tuning” in the same experimental context.

In actual use, you should still recheck the correctness of the model version, training data, product information, and whether the brand allows the use of generative content.