Research map / Image generation / GANs
Compositional GAN: Learning Image-Conditional Binary Composition
Samaneh Azadi, Deepak Pathak, Sayna Ebrahimi, Trevor Darrell
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 example-specific meta-refinement (ESMR) at test time (ablation) | CoDe (Compositional GAN) · Chair-Table (synthetic) | AMT preference after vs before refinement: 71.3% prefer after refinement | Improved |
| add example-specific meta-refinement (ESMR) at test time (ablation) | CoDe (Compositional GAN) · Basket-Bottle (synthetic) | AMT preference after vs before refinement: 64.2% prefer after refinement | Improved |
| train with paired data instead of unpaired data | CoDe (Compositional GAN) · Chair-Table and Basket-Bottle (synthetic) | AMT preference paired vs unpaired: 57% and 57% prefer paired | No clear effect |
| use CoDe instead of ST-GAN baseline (paired training) | Compositional GAN · Face-Sunglasses (real) | AMT preference vs ST-GAN: 84% prefer ours over ST-GAN | Improved |
| use CoDe instead of ST-GAN baseline (unpaired training) | Compositional GAN · Face-Sunglasses (real) | AMT preference vs ST-GAN: 73% prefer ours over ST-GAN | Improved |
| use CoDe instead of ST-GAN baseline (unpaired training) | Compositional GAN · Street Scene-Car (real) | AMT preference vs ST-GAN: 61% prefer ours over ST-GAN | Improved |
| add inpainting network when training with unpaired data (ablation) | CoDe (unpaired) · Chair-Table and Basket-Bottle (synthetic, qualitative) | occlusion correctness (qualitative) | Improved |
| add decomposition generator and self-consistency loss (ablation) | CoDe · Basket-Bottle (qualitative) | object preservation (qualitative) | Improved |
| add cross-entropy mask prediction loss (ablation) | CoDe · Basket-Bottle (qualitative) | object appearance and occlusion (qualitative) | Improved |
| add pixel reconstruction loss on the composite image (ablation) | CoDe · Basket-Bottle (qualitative) | color and occlusion (qualitative) | Improved |
| add GAN loss in training and inference (ablation) | CoDe · Basket-Bottle (qualitative) | color fidelity and quality (qualitative) | Improved |
Closest papers by meaning
- Multi-view Consistent Generative Adversarial Networks for Compositional 3D-Aware Image Synthesis (IJCV 2023 · similarity 0.90)
- Generative Compositional Augmentations for Scene Graph Prediction (ICCV 2021 · similarity 0.90)
- Compositional Transformers for Scene Generation (NeurIPS 2021 · similarity 0.90)
- Using latent space regression to analyze and leverage compositionality in GANs (ICLR 2021 · similarity 0.90)
- Generative Modeling of Infinite Occluded Objects for Compositional Scene Representation (ICML 2019 · similarity 0.88)
- Compositional Scene Understanding through Inverse Generative Modeling (ICML 2025 · 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.
