Research map / Image generation / Text-to-image diffusion
MosaicFusion: Diffusion Models as Data Augmenters for Large Vocabulary Instance Segmentation
Jiahao Xie, Wei Li, Xiangtai Li, Ziwei Liu, Yew-Soon Ong, Chen Change Loy
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 |
|---|---|---|---|
| increase generated objects per image from 1 to 4 (ablation) | MosaicFusion Mask R-CNN · LVIS v1.0 val (Mask R-CNN R50-FPN ablations) | mask AP_r: 14.0 -> 15.2 | Improved |
| add synthetic MosaicFusion data to Mask R-CNN | Mask R-CNN R50 · LVIS v1.0 val (long-tailed) | mask AP_r: 9.6 -> 15.2 | Improved |
| add MosaicFusion data to Box-Supervised CenterNet2 Swin-B | CenterNet2 · LVIS v1.0 val (long-tailed) | mask AP_r: 35.9 -> 39.0 | Improved |
| add MosaicFusion to F-VLM R50x64 on open-vocabulary LVIS | F-VLM · LVIS open-vocabulary val | AP novel: 31.7 -> 34.9 | Improved |
| add MosaicFusion to Co-DETR Swin-L | Co-DETR · LVIS val | box AP_r: 61.4 -> 64.4 | Improved |
| add MosaicFusion to F-VLM R50x64 on V3Det | F-VLM · V3Det open-vocabulary val | box AP novel: 7.0 -> 14.5 | Improved |
| generate multi-object images directly instead of Mosaic augmentation of single-object images (ablation) | MosaicFusion · LVIS v1.0 val (Mask R-CNN R50-FPN ablations) | mask AP_r: 14.0 -> 15.2 | Improved |
| add category definition to text prompt (ablation) | MosaicFusion · LVIS v1.0 val (Mask R-CNN R50-FPN ablations) | mask AP_r: 13.5 -> 15.2 | Improved |
| moderate region overlap (64,48) instead of none (ablation) | MosaicFusion · LVIS v1.0 val (Mask R-CNN R50-FPN ablations) | mask AP_r: 14.2 -> 15.2 | Improved |
| larger region overlap (128,96) (ablation) | MosaicFusion · LVIS v1.0 val (Mask R-CNN R50-FPN ablations) | mask AP_r: 15.2 -> 13.2 | Hurt |
| generate all categories instead of rare only (ablation) | MosaicFusion · LVIS v1.0 val (Mask R-CNN R50-FPN ablations) | mask AP_r: 14.2 -> 15.2 | Improved |
| generate 50 instead of 25 images per category (ablation) | MosaicFusion · LVIS v1.0 val (Mask R-CNN R50-FPN ablations) | mask AP_r: 15.2 -> 14.7 | No clear effect |
Closest papers by meaning
- Open-Vocabulary Attention Maps with Token Optimization for Semantic Segmentation in Diffusion Models (CVPR 2024 · similarity 0.87)
- FreeGen: Bridging Visual-Linguistic Discrepancies Towards Diffusion-based Pixel-level Data Synthesis (AAAI 2025 · similarity 0.87)
- DiverGen: Improving Instance Segmentation by Learning Wider Data Distribution with More Diverse Generative Data (CVPR 2024 · similarity 0.86)
- Seg4Diff: Unveiling Open-Vocabulary Semantic Segmentation in Text-to-Image Diffusion Transformers (NeurIPS 2025 · similarity 0.86)
- Mosaic3D: Foundation Dataset and Model for Open-Vocabulary 3D Segmentation (CVPR 2025 · similarity 0.86)
- SeeDiff: Off-the-Shelf Seeded Mask Generation from Diffusion Models (AAAI 2025 · similarity 0.86)
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.
