Research map / Image generation / Autoregressive and latent generation
PICS: Pairwise Image Compositing with Spatial Interactions
Hang Zhou, Xinxin Zuo, Sen Wang, Li cheng
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 |
|---|---|---|---|
| replace MLP object-fusion with the interaction transformer block (ablation) | PICS · In-the-wild test set | FID: 173.1 -> 165.2 | Improved |
| replace MLP object-fusion with the interaction transformer block (ablation) | PICS · In-the-wild test set | CLIP-score: 74.6 -> 76.3 | Improved |
| add in-plane rotation augmentation (ablation) | PICS · In-the-wild test set | FID: 165.2 -> 162.5 | Improved |
| add in-plane rotation augmentation (ablation) | PICS · In-the-wild test set | CLIP-score: 76.3 -> 74.9 | Hurt |
| train on combined 1M-image data instead of LVIS only (ablation) | PICS · In-the-wild test set | FID: 162.5 -> 158.2 | Improved |
| train on combined 1M-image data instead of LVIS only (ablation) | PICS · In-the-wild test set | CLIP-score: 74.9 -> 77.3 | Improved |
| add multi-view shape prior (ablation) | PICS · In-the-wild test set | FID: 158.2 -> 151.3 | Improved |
| add multi-view shape prior (ablation) | PICS · In-the-wild test set | CLIP-score: 77.3 -> 79.1 | Improved |
| use PICS parallel compositing instead of AnyDoor sequential compositing | PICS · LVIS val (recompositing) | FID: 27.17 -> 24.99 | Improved |
| use PICS instead of AnyDoor | PICS · LVIS val (recompositing) | DINOv2-score: 0.6089 -> 0.5713 | Hurt |
| use PICS instead of OmniPaint | PICS · DreamBooth | DreamSim: 0.2632 -> 0.3054 | Hurt |
| use PICS instead of ObjectStitch | PICS · DreamBooth | FID: 260.4 -> 255.5 | 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.
- OmniPaint — OmniPaint: Mastering Object-Oriented Editing via Disentangled Insertion-Removal Inpainting (ICCV 2025) · better on 3 of 5 shared results
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
- $\textit{Bifr\"ost}$: 3D-Aware Image Compositing with Language Instructions (NeurIPS 2024 · similarity 0.87)
- HOComp: Interaction-Aware Human-Object Composition (NeurIPS 2025 · similarity 0.86)
- MV-CoLight: Efficient Object Compositing with Consistent Lighting and Shadow Generation (NeurIPS 2025 · similarity 0.86)
- PS-Diffusion: Photorealistic Subject-Driven Image Editing with Disentangled Control and Attention (CVPR 2025 · similarity 0.85)
- Learning to Generate Multiple Objects from Dense and Occluded Layouts (NeurIPS 2026 · similarity 0.85)
- ORIDa: Object-centric Real-world Image Composition Dataset (CVPR 2025 · similarity 0.85)
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.
