Research map / Image generation / Fast diffusion sampling
Foreground-Background Separation through Concept Distillation from Generative Image Foundation Models
Mischa Dombrowski, Hadrien Reynaud, Matthew Baugh, Bernhard Kainz
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 |
|---|---|---|---|
| train a U-Net on refined masks of original data plus synthetic data D'_s (ablation) | Concept distillation pipeline · CUB | IoU: 75.1 -> 77.2 | Improved |
| train a U-Net on preliminary masks instead of using them directly (ablation) | Preliminary masks · CUB | IoU: 29.0 -> 66.5 | Improved |
| replace simple crop inpainting with fine-tuned diffusion inpainting for mask refinement (ablation) | Mask refinement · CUB | IoU: 30.7 -> 63.6 | Improved |
| use refined masks directly instead of U-Net trained on preliminary masks (ablation) | Pipeline · CUB | IoU: 66.5 -> 63.6 | No clear effect |
| average attention over more reverse diffusion steps up to T0 about 40 (ablation) | Mean attention maps · CUB | AUC-ROC | Improved |
| include higher (noisier) diffusion steps beyond T0 about 40 (ablation) | Mean attention maps · CUB | AUC-ROC | Hurt |
| instance-wise normalization of attention maps (ablation) | Mean attention maps · CUB | AUC-ROC: 97.08 (no normalization) vs 97.1 | No clear effect |
| Monte Carlo averaging over repeated diffusion samples N (ablation) | Preliminary masks · CUB | AUC-ROC | No clear effect |
| increase classifier-free guidance scale toward foreground (ablation) | Fine-tuned LDM · CUB | FID: 9.8 -> 22.3 | Hurt |
| use object token only or list of objects as prompt rather than 'a photo of a bird' (ablation) | Preliminary masks · CUB | AUC-ROC | Hurt |
| replace Layered GAN with diffusion concept distillation | Foreground-background synthesis · CUB / Stanford Cars | FID: 12.9 -> 9.8; 19.0 -> 13.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- OneGAN — OneGAN: Simultaneous Unsupervised Learning of Conditional Image Generation, Foreground Segmentation, and Fine-Grained Clustering (ECCV 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Emergence of Object Segmentation in Perturbed Generative Models (NeurIPS 2019 · similarity 0.88)
- SeeDiff: Off-the-Shelf Seeded Mask Generation from Diffusion Models (AAAI 2025 · similarity 0.87)
- Explore In-Context Segmentation via Latent Diffusion Models (AAAI 2025 · similarity 0.87)
- Using Diffusion Priors for Video Amodal Segmentation (CVPR 2025 · similarity 0.86)
- FOCUS: Towards Universal Foreground Segmentation (AAAI 2025 · similarity 0.86)
- Segment Anyword: Mask Prompt Inversion for Open-Set Grounded Segmentation (ICML 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.
