Research map / Vision-language and multimodal / Multimodal LLMs
Semantic Generative Tuning for Unified Multimodal Models
Songsong Yu, Yuxin Chen, Ying Shan, Yanwei Li
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 segmentation as generative proxy (SGT) alongside SFT | BAGEL · CV-Bench | accuracy: 73.21 -> 79.23 | Improved |
| add segmentation as generative proxy (SGT) alongside SFT | OmniGen2 · MMVP | accuracy: 65.00 -> 68.33 | Improved |
| add segmentation proxy vs pixel reconstruction proxy (ablation) | BAGEL · CV-Bench | accuracy: 75.23 -> 79.23 | Improved |
| add segmentation proxy vs edge-detection proxy (ablation) | OmniGen2 · MMStar | accuracy: 55.53 -> 57.07 | Improved |
| use SFT data alone without generative proxy (ablation) | BAGEL · GenEval | score: 78.21 -> 77.18 | Hurt |
| add any generative proxy (edge, reconstruction, segmentation) to SFT (ablation) | BAGEL · GenEval | score: 77.18 -> 80.95 | Improved |
| add segmentation proxy to SFT (ablation) | BAGEL · GEdit-Bench-En | score: 6.49 -> 6.94 | Improved |
| add segmentation proxy to SFT (ablation) | BAGEL · MMVP | accuracy: 83.00 -> 83.33 | No clear effect |
| mix panoptic, reconstruction and edge tasks instead of segmentation alone (ablation) | BAGEL · Understanding average (12 benchmarks) | overall average: 75.8 -> 74.9 | Hurt |
| generative tuning with low-level edge or derain tasks only (ablation) | BAGEL · Understanding average (12 benchmarks) | overall average: 74.8 -> 74.4 | No clear effect |
| generative tuning with proxy tasks (no VQA) (ablation) | BAGEL · OCRBench / MathVista | accuracy: MathVista 72.2 -> 69.8 (+Semantic) | Hurt |
| set segmentation-to-VQA mixing ratio to 2:1 (ablation) | BAGEL and OmniGen2 · aggregate understanding score | normalized score | 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.
- OmniGen2 (Base) — OmniGen2: Towards Instruction-Aligned Multimodal Generation (CVPR 2026) · 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
- Generation Enhances Understanding in Unified Multimodal Models via Multi-Representation Generation (ICML 2026 · similarity 0.91)
- Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models (CVPR 2026 · similarity 0.90)
- UVU: Improving Multimodal Understanding via Vision-Language Unified Autoregressive Paradigm (CVPR 2026 · similarity 0.90)
- Janus-Pro-R1: Advancing Collaborative Visual Comprehension and Generation via Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Vision as a Dialect: Unifying Visual Understanding and Generation via Text-Aligned Representations (NeurIPS 2025 · similarity 0.89)
- Generative Multimodal Pretraining with Discrete Diffusion Timestep Tokens (CVPR 2025 · 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.
