UniModel: A Visual-Only Framework for Unified Multimodal Understanding and Generation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhang, Chi, Wang, Jiepeng, Wang, Youming, Liang, Yuanzhi, Yang, Xiaoyan, Li, Zuoxin, Huang, Haibin, Li, Xuelong
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918212582506496
author Zhang, Chi
Wang, Jiepeng
Wang, Youming
Liang, Yuanzhi
Yang, Xiaoyan
Li, Zuoxin
Huang, Haibin
Li, Xuelong
author_facet Zhang, Chi
Wang, Jiepeng
Wang, Youming
Liang, Yuanzhi
Yang, Xiaoyan
Li, Zuoxin
Huang, Haibin
Li, Xuelong
contents We present UniModel, a unified generative model that jointly supports visual understanding and visual generation within a single pixel-to-pixel diffusion framework. Our goal is to achieve unification along three axes: the model, the tasks, and the representations. At the representation level, we eliminate modality discrepancies by mapping both text and images into a shared visual space: textual prompts are rendered as painted text images on a clean canvas, and all inputs and outputs are treated purely as RGB pixels. This yields a fully vision-native formulation of multimodal learning. At the task level, a broad range of vision-language problems are cast as pixel-to-pixel transformations in this visual space. For understanding tasks, the model takes an RGB image and produces a painted text image that visually encodes the semantic prediction. For generation tasks, painted text images serve as visual conditions that guide realistic and semantically aligned image synthesis. Captioning and text-to-image generation thus become different directions of the same underlying visual translation process. At the model level, we instantiate a single Unified Diffusion Transformer trained with rectified flow in pixel space. A shared backbone jointly learns bidirectional mappings between natural images and painted text images, with lightweight task embeddings to specify the desired direction. Experiments on text-to-image synthesis and image-to-text understanding demonstrate strong cross-modal alignment and emergent controllability such as cycle-consistent image-caption-image loops. Our initial exploration suggests that unifying model, tasks, and representations in a single visual space is a promising paradigm for general-purpose multimodal intelligence.
format Preprint
id arxiv_https___arxiv_org_abs_2511_16917
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle UniModel: A Visual-Only Framework for Unified Multimodal Understanding and Generation
Zhang, Chi
Wang, Jiepeng
Wang, Youming
Liang, Yuanzhi
Yang, Xiaoyan
Li, Zuoxin
Huang, Haibin
Li, Xuelong
Computer Vision and Pattern Recognition
We present UniModel, a unified generative model that jointly supports visual understanding and visual generation within a single pixel-to-pixel diffusion framework. Our goal is to achieve unification along three axes: the model, the tasks, and the representations. At the representation level, we eliminate modality discrepancies by mapping both text and images into a shared visual space: textual prompts are rendered as painted text images on a clean canvas, and all inputs and outputs are treated purely as RGB pixels. This yields a fully vision-native formulation of multimodal learning. At the task level, a broad range of vision-language problems are cast as pixel-to-pixel transformations in this visual space. For understanding tasks, the model takes an RGB image and produces a painted text image that visually encodes the semantic prediction. For generation tasks, painted text images serve as visual conditions that guide realistic and semantically aligned image synthesis. Captioning and text-to-image generation thus become different directions of the same underlying visual translation process. At the model level, we instantiate a single Unified Diffusion Transformer trained with rectified flow in pixel space. A shared backbone jointly learns bidirectional mappings between natural images and painted text images, with lightweight task embeddings to specify the desired direction. Experiments on text-to-image synthesis and image-to-text understanding demonstrate strong cross-modal alignment and emergent controllability such as cycle-consistent image-caption-image loops. Our initial exploration suggests that unifying model, tasks, and representations in a single visual space is a promising paradigm for general-purpose multimodal intelligence.
title UniModel: A Visual-Only Framework for Unified Multimodal Understanding and Generation
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2511.16917