Research map / Image generation / Text-to-image diffusion
Label-free Neural Semantic Image Synthesis
Jiayi Wang, Kevin A Laube, Yumeng Li, Jan Hendrik Metzen, Shin-I Cheng, Julio Borges, Anna Khoreva
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 |
|---|---|---|---|
| use neural layout (20 PCA) instead of ground-truth semantic segmentation as ControlNet conditioning | ControlNet · COCO-Stuff | mIoU: 43.3 -> 52.9 | Improved |
| use SD features rather than CLIP for neural layout (ablation) | LUMEN · COCO-Stuff | mIoU: 41.7 -> 51.4 | Improved |
| increase number of PCA components from 1 to 100 (ablation) | LUMEN (SD features) · COCO-Stuff | mIoU: 45.0 -> 54.6 | Improved |
| increase number of PCA components from 1 to 100 (ablation) | LUMEN (SD features) · COCO-Stuff | LPIPS diversity: 0.53 -> 0.25 | Hurt |
| increase number of PCA components from 1 to 100 (ablation) | LUMEN (SD features) · COCO-Stuff | TIFA: 0.87 -> 0.47 | Hurt |
| add unlabeled Cityscapes images (L + U) when training the generator with neural layout (ablation) | LUMEN · Cityscapes | mIoU: 57.6 -> 58.2 | Improved |
| add unlabeled images using pseudo-labels for Sem. Seg. conditioning (ablation) | ControlNet Sem. Seg. · Cityscapes | mIoU: 57.9 -> 53.3 | Hurt |
| add LUMEN synthetic data to real NYUv2 training data | multi-task ResNet-50 · NYUv2 sem. seg. | mIoU: 44.6 -> 47.3 | Improved |
| add LUMEN synthetic data to real NYUv2 training data | multi-task ResNet-50 · NYUv2 depth | AbsErr: 1.17 -> 1.25 | Hurt |
| augment Cityscapes training with LUMEN prompt-edited images | SegFormer · Cityscapes to ACDC | Avg. mIoU: 47.0 -> 52.9 | Improved |
| use PnP-Diffusion prompt editing for augmentation | SegFormer · Cityscapes to ACDC | Avg. mIoU: 47.0 -> 48.7 | Improved |
| use neural layout instead of Canny edges, depth or predicted Sem. Seg. | ControlNet · COCO-Stuff | TIFA: 0.88 -> 0.66 | Hurt |
Closest papers by meaning
- PLACE: Adaptive Layout-Semantic Fusion for Semantic Image Synthesis (CVPR 2024 · similarity 0.88)
- Unlocking Pre-trained Image Backbones for Semantic Image Synthesis (CVPR 2024 · similarity 0.87)
- Adversarial Supervision Makes Layout-to-Image Diffusion Models Thrive (ICLR 2024 · similarity 0.87)
- Universal Few-shot Spatial Control for Diffusion Models (NeurIPS 2025 · similarity 0.87)
- SSMG: Spatial-Semantic Map Guided Diffusion Model for Free-Form Layout-to-Image Generation (AAAI 2024 · similarity 0.87)
- Grounded Text-to-Image Synthesis with Attention Refocusing (CVPR 2024 · 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.
