metadata

license: mit
title: The Other Tiger
sdk: gradio
emoji: 👁
colorFrom: blue
colorTo: purple
pinned: true

The Other Tiger

tl;dr

Train on embeddings of media preferred by a specific user -> produce embeddings of media they may enjoy.

In our case here, we take the ECLIPSE text embedding -> image embedding prior (https://arxiv.org/abs/2312.04655) and finetune it to become a preferred image embeddings -> heldout image embedding prior.

Related work:

Patron et al. models preference using a diffusion prior and condition on user ids with ratings: https://arxiv.org/abs/2502.18477

Wang et al. models preference using a generator conditioned on averaged CLIP embeddings of users: https://arxiv.org/abs/2304.03516

My previous work based on Collaborative Filtering with CLIP embeddings: https://github.com/rynmurdock/generative_recommender