Uni-Edit: Intelligent Editing Is A General Task For Unified Model Tuning

Problem
This paper addresses the limitations of current approaches to enhancing Unified Multimodal Models (UMMs) through mixed multi-task training, which often leads to performance trade-offs due to task conflicts. The authors argue that existing methods require complex multi-stage pipelines and extensive data mixing, which do not facilitate true mutual reinforcement among image understanding, generation, and editing capabilities. The work is presented as a preprint, indicating it has not yet undergone peer review.

Method
The core technical contribution is the introduction of Uni-Edit, an intelligent image editing task designed as a unified approach for tuning UMMs. Uni-Edit simplifies the training process by focusing on a single task, utilizing a single training stage, and employing a single dataset. The authors propose a novel automated data synthesis pipeline that transforms existing Visual Question Answering (VQA) data into complex editing instructions, incorporating embedded questions and nested logic. This results in the creation of the Uni-Edit-148k dataset, which pairs diverse reasoning-intensive instructions with high-quality edited images. The architecture and specific training compute details are not disclosed, but the methodology emphasizes the efficiency of using a singular task to enhance multiple capabilities simultaneously.

Results
The authors conducted extensive experiments on the BAGEL and Janus-Pro benchmarks. Tuning solely on the Uni-Edit task resulted in significant performance improvements across image understanding, generation, and editing capabilities. While specific numerical results are not provided in the summary, the authors claim comprehensive enhancements compared to baseline models that utilize traditional multi-task training approaches. The effectiveness of Uni-Edit is underscored by its ability to outperform existing methods without the need for auxiliary operations.

Limitations
The authors acknowledge that the reliance on a single task may limit the exploration of other potential tasks that could further enhance UMM capabilities. Additionally, the automated data synthesis pipeline, while innovative, may introduce biases inherent in the VQA data it transforms. The paper does not address the scalability of the approach to other domains outside of image editing or the potential for overfitting due to the focused nature of the training data.

Why it matters
The introduction of Uni-Edit as a general task for UMM tuning has significant implications for the field of multimodal AI. By demonstrating that a single, well-defined task can enhance multiple capabilities simultaneously, this work challenges the prevailing paradigm of complex multi-task training. It opens avenues for future research to explore other unified tasks that could similarly streamline the training of UMMs. Furthermore, the automated data synthesis approach may inspire new methodologies for generating training data across various domains, potentially leading to more efficient model training processes.

Authors: Dian Zheng, Manyuan Zhang, Hongyu Li, Hongbo Liu, Kai Zou, Kaituo Feng, Hongsheng Li
Source: arXiv:2605.21487
URL: https://arxiv.org/abs/2605.21487v1