A modality-agnostic recommender learns one tokenizer from image and text data

_reachsumit · x · 2026-08-04

A modality-agnostic generative recommender learns one tokenizer for images and text

This paper proposes a modality-agnostic generative recommendation method that can learn from paired, image-only, and text-only item data.

The model builds a shared semantic-ID tokenizer across different modalities.
The goal is to unify recommendation learning even when training data is incomplete or unpaired.
The result is a generative recommender that can exploit heterogeneous item sources without requiring every example to have both image and text.

More from Multimodal