TUNA: Taming Unified Visual Representations for Native Unified Multimodal Models
Zhiheng Liu, Weiming Ren, Haozhe Liu, Zijian Zhou, Shoufa Chen, Haonan Qiu, Xiaoke Huang, Zhaochong An, Fanny Yang, Aditya Patel, Viktar Atliha, Tony Ng, Xiao Han, Chuyan Zhu, Chenyang Zhang, Ding Liu, Juan-Manuel Perez-Rua, Sen He, Jürgen Schmidhuber, Wenhu Chen, Ping Luo, Wei Liu
2025-12-02
Summary
This paper introduces TUNA, a new type of artificial intelligence model designed to understand and create content using different types of data, like images and videos, all at the same time.
What's the problem?
Existing multimodal models, which try to handle multiple types of data, often treat images and videos separately. They use different methods to process each type, creating a disconnect and making it harder for the model to truly understand the relationship between them. This separation leads to less effective performance in tasks that require combining information from both sources.
What's the solution?
The researchers created TUNA, which uses a single, unified way to represent visual information from both images and videos. It does this by first compressing the visual data into a smaller, more manageable form using a technique called a VAE, and then further refining that representation with another encoder. Because everything is processed in the same way, there are no mismatches in how the model 'sees' the data, allowing for smoother and more accurate understanding and generation.
Why it matters?
TUNA’s unified approach significantly improves performance on a variety of tasks, including understanding what’s happening in images and videos, creating new images and videos, and even editing existing ones. It also shows that using a better initial representation of the visual data leads to better results overall, and that training the model to both understand *and* create content helps it learn more effectively.
Abstract
Unified multimodal models (UMMs) aim to jointly perform multimodal understanding and generation within a single framework. We present TUNA, a native UMM that builds a unified continuous visual representation by cascading a VAE encoder with a representation encoder. This unified representation space allows end-to-end processing of images and videos for both understanding and generation tasks. Compared to prior UMMs with decoupled representations, TUNA's unified visual space avoids representation format mismatches introduced by separate encoders, outperforming decoupled alternatives in both understanding and generation. Moreover, we observe that stronger pretrained representation encoders consistently yield better performance across all multimodal tasks, highlighting the importance of the representation encoder. Finally, in this unified setting, jointly training on both understanding and generation data allows the two tasks to benefit from each other rather than interfere. Our extensive experiments on multimodal understanding and generation benchmarks show that TUNA achieves state-of-the-art results in image and video understanding, image and video generation, and image editing, demonstrating the effectiveness and scalability of its unified representation design.