Research map / Image generation / Text-to-image diffusion
Scone: Bridging Composition and Distinction in Subject-Driven Image Generation via Unified Understanding-Generation Modeling
Yuran Wang, Bohan Zeng, Chengzhuo Tong, Wenxuan Liu, Yang Shi, Xiaochen Ma, Hao Liang, Yuanxing Zhang, Wentao Zhang
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 |
|---|---|---|---|
| Fine-tune BAGEL on 70K base single-candidate data (stage I step 1) (ablation) | BAGEL · OmniContext | overall score: 6.03 -> 7.95 | Improved |
| Add a second epoch on 22K refined single-candidate data (ablation) | Scone stage I · OmniContext | overall score: 7.95 -> 8.02 | Improved |
| Use two-step training (understanding expert first, then both) instead of direct fine-tuning (ablation) | Scone stage II · SconeEval | overall score: 7.94 -> 8.43 | Improved |
| Add understanding bridge (semantic attention mask) in stage II (ablation) | Scone · SconeEval | overall score: 8.43 -> 8.50 | Improved |
| Add understanding bridge (ablation) | Scone · SconeEval | distinction score: 8.70 -> 8.79 | Improved |
| Raise mask threshold from 0.82 to 0.88 (ablation) | Scone · SconeEval | overall score: 8.46 -> 8.50 | Improved |
| Apply stage II after stage I (ablation) | Scone · SconeEval | overall score: 7.86 -> 8.50 | Improved |
| Unified understanding-generation backbone over pure generation models for distinction | OmniGen2 vs Qwen-Image-Edit-2509 · SconeEval | average distinction score: 7.65 -> 7.81 | Improved |
| Scone vs Echo-4o on composition in single-subject (SconeEval) | Scone · SconeEval Single COM | composition score: 8.58 vs 8.52 | No clear effect |
Closest papers by meaning
- Instruct-Imagen: Image Generation with Multi-modal Instruction (CVPR 2024 · similarity 0.86)
- ConceptMix: A Compositional Image Generation Benchmark with Controllable Difficulty (NeurIPS 2024 · similarity 0.86)
- Compositional Scene Understanding through Inverse Generative Modeling (ICML 2025 · similarity 0.85)
- WEAVE: Unleashing and Benchmarking the In-context Interleaved Comprehension and Generation (CVPR 2026 · similarity 0.85)
- Non-confusing Generation of Customized Concepts in Diffusion Models (ICML 2024 · similarity 0.85)
- Progressive Compositionality in Text-to-Image Generative Models (ICLR 2025 · similarity 0.85)
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.
