Research map / Image generation / Text-to-image diffusion
Training-free Composite Scene Generation for Layout-to-Image Synthesis
Jiaqi Liu, Tao Huang, Chang Xu
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 inter-token constraint to intra-token loss only (ablation) | CSG · COCO 2014 (2 objects) | AP50: +7.9 AP50, +3.7 AP | Improved |
| add inter-token constraint (ablation) | CSG · COCO 2014 (2 objects) | CLIP score: 0.3202 -> 0.3164 | Hurt |
| add self-attention constraint (ablation) | CSG · COCO 2014 (2 objects) | AP50: 40.2 -> 50.3 (full) | Improved |
| add attention redistribution (ablation) | CSG · COCO 2014 (2 objects) | AP50: 46.8 -> 50.3 | Improved |
| add selective sampling to intra+inter (ablation) | CSG · COCO 2014 (2 objects) | AP50: 39.1 -> 40.2 | Improved |
| set inter-token margin to 0.1 rather than 0.05 or 0.2 (ablation) | CSG · COCO 2014 (2 objects) | AP50: 48.8 / 46.8 -> 50.3 | Improved |
| use selective sampling K=80% rather than 100% or 40% (ablation) | CSG · COCO 2014 (2 objects) | AP50: 46.6 (100%) / 37.2 (40%) -> 50.3 | Improved |
| keep M=50% of selected attention values (ablation) | CSG · COCO 2014 (2 objects) | AP50: 48.2 (100%) / 45.7 (25%) -> 50.3 | Improved |
| increase refinement steps T_R from 1 to 5 (ablation) | CSG · COCO 2014 (2 objects) | AP50: 27.5 -> 50.3 | Improved |
| increase T_R from 5 to 7 (ablation) | CSG · COCO 2014 (2 objects) | AP50: 50.3 -> 51.1 AP50, time 16.5 -> 23.4 s | No clear effect |
| apply CSG on top of GLIGEN | GLIGEN · COCO 2014 (5 objects) | AP50: 54.2 -> 60.6 | Improved |
| use CSG instead of BoxDiff/MultiDiffusion/Layout-control | SD 1.4 layout-to-image · COCO 2014 (3 objects) | FID: 57.08 vs 56.00 (BoxDiff) | Mixed |
Closest papers by meaning
- R&B: Region and Boundary Aware Zero-shot Grounded Text-to-image Generation (ICLR 2024 · similarity 0.92)
- Adversarial Supervision Makes Layout-to-Image Diffusion Models Thrive (ICLR 2024 · similarity 0.91)
- SSMG: Spatial-Semantic Map Guided Diffusion Model for Free-Form Layout-to-Image Generation (AAAI 2024 · similarity 0.90)
- Check Locate Rectify: A Training-Free Layout Calibration System for Text-to-Image Generation (CVPR 2024 · similarity 0.90)
- Build-A-Scene: Interactive 3D Layout Control for Diffusion-Based Image Generation (ICLR 2025 · similarity 0.90)
- Control and Realism: Best of Both Worlds in Layout-to-Image without Training (ICML 2025 · similarity 0.90)
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.
