CatalogStitch: Dimension-Aware and Occlusion-Preserving Object Compositing for Catalog Image Generation
Sanyam Jain, Pragya Kandari, Manit Singhal, He Zhang, Soo Ye Kim
Adobe
CVPR 2026 — HiGen Workshop (Human-Interactive Generation and Editing)
[Project Page] [Paper] [arXiv]
This dataset is released under the Adobe Research License for noncommercial research purposes only. See dataset/LICENSE.md for the full license text and third-party image licensing details.
A 58-example evaluation benchmark for catalog image compositing:
- 35 dimension-mismatch scenes — products with significantly different aspect ratios
- 23 occlusion scenes — products partially occluded by 1–2 foreground elements
| Component | Path | Description |
|---|---|---|
| Background images | dataset/images/backgrounds/ |
High-resolution scene images with target product regions |
| Product images | dataset/images/products/ |
Replacement product images for compositing |
| Masks | dataset/masks/ |
Freeform, bounding box, and dimension-aware masks per example |
| Model outputs | dataset/results/ |
Composited outputs from ObjectStitch, OmniPaint, and InsertAnything |
| Metadata | dataset/dataset_metadata.json |
Source URLs, licensing information, and quantitative metrics |
| PDF summaries | additional_results_*.pdf |
Full visual results for all 58 benchmark examples |
| Interactive viewers | dataset/results/*/index.html |
Side-by-side comparison browsers |
@inproceedings{catalogstitch,
title = {CatalogStitch: Dimension-Aware and Occlusion-Preserving Object Compositing for Catalog Image Generation},
author = {Sanyam Jain and Pragya Kandari and Manit Singhal and He Zhang and Soo Ye Kim},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
workshop = {HiGen: Human-Interactive Generation and Editing},
year = {2026}
}- Sanyam Jain — sanyjain@adobe.com
- Pragya Kandari — pkandari@adobe.com
- Manit Singhal — manits@adobe.com
- He Zhang — hezhan@adobe.com
- Soo Ye Kim — sooyek@adobe.com