Research map / Vision-language and multimodal / Human and embodied modeling
DIVA: Harnessing the Representation Divergence in Unified Multimodal Models for Mutual Reinforcement
Renjie Lu, Xulong Zhang, Xiaoyang Qu, Jianzong Wang, Shangfei Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DIVA post-training | Nexus-Gen · MMMU | accuracy: 43.5 -> 49.4 | Improved |
| add DIVA post-training | Show-o · DPG-Bench | score: 69.81 -> 76.03 | Improved |
| add DIVA post-training | Liquid · GenEval | score: 0.70 -> 0.81 | Improved |
| plain SFT on the same post-training data instead of DIVA (ablation) | Show-o · MMMU | accuracy: 26.3 -> 26.8 | No clear effect |
| plain SFT on same data (ablation) | Show-o · GenEval | score: 0.69 -> 0.67 | Hurt |
| add the unique-information regularizer Iuni (ablation) | DIVA · Show-o MMMU | accuracy: 28.3 -> 32.4 | Improved |
| add stop-gradient in shared MI alignment (ablation) | DIVA · Show-o DPG-Bench | score: 74.92 -> 76.03 | Improved |
| use gated MLP encoders instead of Linear+LN (ablation) | DIVA · Show-o MMMU | accuracy: 29.4 -> 32.4 | Improved |
| use Transformer encoders instead of gated MLP (ablation) | DIVA · Show-o | MMMU/POPE/DPG-Bench: 32.4 -> 32.1 | No clear effect |
| use random masks instead of contiguous block mask (ablation) | DIVA · Show-o MMMU | accuracy: 24.7 -> 32.4 | Improved |
| vary middle-layer range (9-17, 7-17, 7-19 vs 8-18) (ablation) | DIVA · Show-o | MMMU/GenEval/DPG-Bench: GenEval 0.72-0.75 | No clear effect |
| increase post-training data from 70K to 200K (ablation) | DIVA · Show-o DPG-Bench | score: 75.83 -> 76.03 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Show-o — Show-o: One Single Transformer to Unify Multimodal Understanding and Generation (ICLR 2025) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- UVU: Improving Multimodal Understanding via Vision-Language Unified Autoregressive Paradigm (CVPR 2026 · similarity 0.89)
- Generation Enhances Understanding in Unified Multimodal Models via Multi-Representation Generation (ICML 2026 · similarity 0.89)
- Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation (CVPR 2025 · similarity 0.89)
- ParaUni: Enhance Generation in Unified Multimodal Model with Reinforcement-driven Hierarchical Parallel Information Interaction (CVPR 2026 · similarity 0.88)
- RealUnify: Do Unified Models Truly Benefit from Unification? A Comprehensive Benchmark (CVPR 2026 · similarity 0.88)
- InfMasking: Unleashing Synergistic Information by Contrastive Multimodal Interactions (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
