Research map / Detection and segmentation / Weakly supervised detection and segmentation
Controlling Style and Semantics in Weakly-Supervised Image Generation
Dario Pavllo, Aurelien Lucchi, Thomas Hofmann
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 |
|---|---|---|---|
| use sparse detector masks instead of dense semantic segmentation masks (ablation) | 1-step weakly-supervised model · COCO-val | FID: 23.97 -> 20.02 | Improved |
| use sparse detector masks instead of ground-truth masks (SPADE) (ablation) | 1-step model · COCO-val | FID: 22.64 -> 20.02 | Improved |
| train on extra unlabeled images (VG+ augmented) (ablation) | 1-step sparse mask model · COCO-val | FID: 20.02 -> 18.93 | Improved |
| add stuff classes to the mask classes (vs COCO things only) (ablation) | SPADE ground-truth masks · COCO-val | FID: 32.31 (+9.67) | Improved |
| add ground-truth attributes to the discriminator (ablation) | text-conditioned 1-step model · COCO-val | FID: 20.63 -> 20.44 (-0.19) | No clear effect |
| draw BERT embeddings from second-to-last layer (ablation) | text-conditioned 1-step model · COCO-val | FID: 20.63 -> 19.77 (-0.86) | Improved |
| use 6 attention heads instead of 12 with second-to-last layer (ablation) | text-conditioned 1-step model · COCO-val | FID: 19.77 -> 19.65 | Improved |
| feed foreground information to S_avg block of G1 (ablation) | 2-step attribute model · COCO-val | FID: 25.16 -> 20.83 (+4.33 when removed) | Improved |
| randomize style attributes at inference (ablation) | 2-step attribute model · COCO-val | FID: 20.83 -> 20.64 (-0.19) | No clear effect |
| two-step background/foreground decomposition instead of one step (ablation) | text-conditioned model · COCO-val | FID: 20.63 -> 20.64 | No clear effect |
| two-step decomposition instead of one step (ablation) | attribute-conditioned model · COCO-val / VG-val | FID: 21.13 -> 20.83 (COCO), 15.12 -> 14.88 (VG) | Improved |
| add text style input (6h, L_{n-1}) to sparse mask model (ablation) | 1-step sparse model · COCO-val | FID: 20.02 -> 19.65 | No clear effect |
Closest papers by meaning
- Semantic Browsing: Controllable Diversity for Image Generation (ECCV 2026 · similarity 0.88)
- Combining Bottom-Up, Top-Down, and Smoothness Cues for Weakly Supervised Image Segmentation (CVPR 2017 · similarity 0.86)
- SSMG: Spatial-Semantic Map Guided Diffusion Model for Free-Form Layout-to-Image Generation (AAAI 2024 · similarity 0.86)
- Looking Beyond Single Images for Weakly Supervised Semantic Segmentation Learning (TPAMI 2022 · similarity 0.86)
- CtrLoRA: An Extensible and Efficient Framework for Controllable Image Generation (ICLR 2025 · similarity 0.86)
- Learning Canonical Representations for Scene Graph to Image Generation (ECCV 2020 · 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.
