FashionDPO: Fashion Image Generation with Direct Preference Optimization

This repository provides FashionDPO models, including checkpoints fine-tuned on the iFashion and Polyvore datasets, and the trained VBPR model. FashionDPO utilizes a novel approach to fashion image generation by incorporating direct preference optimization. This allows for the generation of high-quality fashion images that align with user preferences. The model is based on this paper.

Usage:

The FashionDPO pipeline can be used to generate fashion images. Refer to the Github repository for detailed usage instructions, including how to generate initial recommendations and incorporate feedback from multiple experts. This includes instructions for running sample.py for image generation and multiple_evaluate.py for feedback generation.

Model Checkpoints:

• checkpoint_ifashion: Fine-tuned on the iFashion dataset.
• checkpoint_polyvore: Fine-tuned on the Polyvore dataset.

These checkpoints are available in the repository.