Research map / Vision-language and multimodal / Multimodal LLMs
Transferability Between Understanding and Generation in Unified Multimodal Models
Jiwon Kang, Heeji Yoon, Jaewoo Jung, Jaewon Min, Minkyeong Jeon, Biyeon Hwang, Sangwon Jung, Seungryong Kim
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 |
|---|---|---|---|
| train on counting understanding only (und->gen transfer) (ablation) | Lumina-DiMOO · Counting generation (GenEval-style) | accuracy (%): 48.0 -> 57.0 | Improved |
| train on counting understanding only (und->gen) (ablation) | Janus-Pro · Counting generation (GenEval-style) | accuracy (%): 38.0 -> 32.0 | Hurt |
| train on counting understanding only (und->gen) (ablation) | BAGEL · Counting generation (GenEval-style) | accuracy (%): 42.0 -> 47.0 | Improved |
| train on counting understanding only (und->gen) (ablation) | BLIP3-o · Counting generation (GenEval-style) | accuracy (%): 31.0 -> 37.0 | Improved |
| train on counting generation only (gen->und) (ablation) | Lumina-DiMOO · Counting understanding (PixMo-Count) | accuracy (%): 22.0 -> 30.0 | Improved |
| train on counting generation only (gen->und) (ablation) | BLIP3-o · Counting understanding (PixMo-Count) | accuracy (%): 63.0 -> 62.0 | No clear effect |
| train counting via understanding instead of direct generation fine-tuning | Lumina-DiMOO · Counting generation (GenEval-style) | FID vs baseline images: 52.51 -> 31.47 | Improved |
| train counting via understanding instead of direct generation fine-tuning | Lumina-DiMOO · Counting generation (GenEval-style) | Inception Score: 15.29 -> 17.55 | Improved |
| train counting via understanding instead of direct generation fine-tuning | Lumina-DiMOO · Counting generation (GenEval-style) | MAD: 0.91 -> 0.83 | Improved |
| train spatial relation via understanding instead of generation | Lumina-DiMOO · Spatial relation generation (diagonal, synthetic) | accuracy (%): 80.0 -> 74.0 | Hurt |
| train spatial relation via understanding instead of generation | Lumina-DiMOO · Spatial relation generation (diagonal, synthetic) | FID vs baseline images: 32.28 -> 30.95 | Improved |
| train text via understanding instead of generation | Lumina-DiMOO · Text generation (rendered Markdown) | BLEU: 0.354 -> 0.274 | Hurt |
Closest papers by meaning
- DuoGen: Towards Autonomous Interleaved Multimodal Generation (CVPR 2026 · similarity 0.90)
- Generation Enhances Understanding in Unified Multimodal Models via Multi-Representation Generation (ICML 2026 · similarity 0.90)
- Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models (CVPR 2026 · similarity 0.90)
- Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation (CVPR 2025 · similarity 0.89)
- Breaking Dual Bottlenecks: Evolving Unified Multimodal Models into Self-Adaptive Interleaved Visual Reasoners (ICML 2026 · similarity 0.89)
- RealUnify: Do Unified Models Truly Benefit from Unification? A Comprehensive Benchmark (CVPR 2026 · similarity 0.89)
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.
