Research map / Vision-language and multimodal / Cross-modal learning
Enhancing Alignment for Unified Multimodal Models via Semantically-Grounded Supervision
Jiyeong Kim, Yerim So, Hyesong Choi, Uiwon Hwang, Dongbo Min
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 |
|---|---|---|---|
| replace SFT with SeGroS fine-tuning | Harmon-1.5B (BLIP3o-60k) · GenEval overall | overall score: 85.0 -> 88.66 | Improved |
| replace Reca with SeGroS fine-tuning | OpenUni-3.6B · GenEval overall | overall score: 74.1 -> 75.37 | Improved |
| replace SFT with SeGroS fine-tuning | OpenUni-3.6B · GenEval Position | position accuracy: 25.0 -> 46.8 | Improved |
| replace Reca with SeGroS fine-tuning | Show-o-512 · GenEval overall | overall score: 71.63 -> 70.69 | Hurt |
| add visual hints from top 30% grounded patches (ablation) | Harmon-0.5B · GenEval overall (Harmon-0.5B, MidjourneyV6 ablation) | GenEval overall: 72.1 -> 81.5 | Improved |
| use full image (100%) instead of top 30% patches as hint (ablation) | SeGroS Harmon-0.5B · GenEval overall (Harmon-0.5B, MidjourneyV6 ablation) | GenEval overall: 81.5 -> 79.2 | Hurt |
| add intra- and inter-modal affinity text token filtering (ablation) | SeGroS Harmon-0.5B · GenEval overall (Harmon-0.5B, MidjourneyV6 ablation) | GenEval overall: 79.2 -> 81.5 | Improved |
| add intra-modal affinity only text filtering (ablation) | SeGroS Harmon-0.5B · GenEval overall (Harmon-0.5B, MidjourneyV6 ablation) | GenEval overall: 79.2 -> 79.8 | Improved |
| use bottom patches as unmasked context instead of top patches (ablation) | SeGroS Harmon-0.5B · GenEval overall (Harmon-0.5B, MidjourneyV6 ablation) | GenEval overall: 79.9 -> 81.5 | Improved |
| use top rather than bottom patches as visual hints (ablation) | SeGroS Harmon-0.5B · GenEval overall (Harmon-0.5B, MidjourneyV6 ablation) | GenEval overall: 80.4 -> 81.5 | Improved |
| restrict reconstruction loss to top-30% grounded masked targets (ablation) | SeGroS Harmon-0.5B · GenEval overall (Harmon-0.5B, MidjourneyV6 ablation) | GenEval / DPGBench: 81.5 -> 82.1 | Mixed |
| add noise scale 0.5 to the visual grounding map (ablation) | SeGroS Harmon-1.5B (BLIP3o-60k) · GenEval overall | overall score: 87.9 -> 88.7 | Improved |
Closest papers by meaning
- UNIALIGN: Scaling Multimodal Alignment within One Unified Model (CVPR 2025 · similarity 0.88)
- Reconstruction Alignment Improves Unified Multimodal Models (ICLR 2026 · similarity 0.88)
- Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models (CVPR 2026 · similarity 0.88)
- Implicit Multimodal Alignment: On the Generalization of Frozen LLMs to Multimodal Inputs (NeurIPS 2024 · similarity 0.87)
- Emu: Generative Pretraining in Multimodality (ICLR 2024 · similarity 0.87)
- SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation (CVPR 2025 · similarity 0.87)
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.
