DIVA: Harnessing the Representation Divergence in Unified Multimodal Models for Mutual Reinforcement

Unified Multimodal models (UMMs) built on a single architecture have shown impressive performance in both understanding and generation. We identify a fundamental challenge lies in inductive biases induced by distinct supervision signals: generation branch prefers high-fidelity, fine-grained representations capable of reconstruction, while the understanding favours semantically discriminative embeddings that remain invariant to task-irrelevant factors. Consequently, optimizing these complementary but non-equivalent objectives within a monolithic backbone leads to mutual impairment instead of enhancement. In this paper, we first analyze the root cause of this interference in unified backbones and reveal a complementary structure in their internal representations. Motivated by the observation, we propose DIVA, a self-improved post-training framework that transforms the representation divergence into interior synergy. By explicitly factorizing the visual representation into shared and unique components based on two complementary information flow, DIVA enables both the understanding and generation branches to achieve beneficial transferring while preserving the integrity of unique information from cross-flow interference via mutual information estimation. Despite its generality, our method consistently achieves improvements across visual understanding (+7.82%) and generation (+8.46%). The official code is available at: https://anonymous.4open.science/r/DIVA-D225.

Modern AI systems are increasingly expected to both understand images and generate new ones, but combining these abilities in a single model is difficult. Image understanding often needs the model to focus on high-level meaning, while image generation needs it to preserve fine visual details. When both abilities are trained together inside one shared model, they can interfere with each other instead of helping each other. In this paper, we propose DIVA, a training method that turns this conflict into cooperation. The key idea is to separate the model’s internal visual information into two parts: information that should be shared between understanding and generation, and information that should remain specific to each task. This allows the two abilities to exchange useful signals while avoiding harmful interference. Our results show that DIVA improves both visual understanding and image generation across several benchmarks. This suggests that unified multimodal models do not always need larger or more complicated architectures to improve. Instead, better use of their internal representations can help them become more capable and balanced.