amazon/ABO-Edit

DiTailed: Ensuring Visual Object Consistency in Text-Image-to-Image Flow Matching Models

We present ABO-Edit, a curated dataset for training and evaluating generative models on Visual Object Consistency. ABO-Edit addresses the challenging task of transforming “Lifestyle” images (depicting products in complex real-world usage scenarios) into studio-quality representations: the same product isolated on a white background with realistic shadow, rotated and tilted to a precisely specified angle.

Each sample comprises a triplet $⟨x_{src},p_{src\to trg},x_{trg}⟩\backslash langle\; x\_\{src\},\backslash \; p\_\{src\; \backslash rightarrow\; trg\},\backslash \; x\_\{trg\}\; \backslash rangle$, where:

1. $x_{src}x\_\{src\}$ denotes a source lifestyle image
2. $p_{src\to trg}p\_\{src\; \backslash rightarrow\; trg\}$ represents a detailed editing prompt including fine-grained rotation angles
3. $x_{trg}x\_\{trg\}$ is the corresponding ground-truth target image rendered from a 3D asset

Attribution & License

• Original data credit: Images and 3D assets are © Amazon.com.
• Modifications: This dataset was constructed on top of ABO by rendering studio-quality target images from ABO 3D assets and pairing them with lifestyle sources and VLM-generated editing prompts; see our paper for full details.

In accordance with CC BY 4.0, ABO-Edit is derived from Amazon Berkeley Objects (ABO) and it is distributed under the same CC BY 4.0 license, and no additional restrictions are applied.

Citation

If you use this dataset, please cite both our work and the original ABO dataset.

@misc{taioli2026ABO-Edit,
  title={DiTailed: Ensuring Visual Object Consistency in Text-Image-to-Image Flow Matching Models}, 
  author={Francesco Taioli and Daniel Coelho and Iaroslav Melekhov and Roberto Alcover-Couso and Jose Miguel Grande Saiz and Virginia Fernandez Arguedas and Artur Bekasov},
  year={2026},
  eprint={2607.12539},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  ]url={https://arxiv.org/abs/2607.12539}, 
}

@article{collins2022abo,
  title={ABO: Dataset and Benchmarks for Real-World 3D Object Understanding},
  author={Collins, Jasmine and Goel, Shubham and Deng, Kenan and Luthra, Achleshwar and Xu, Leon and Gundogdu, Erhan and Zhang, Xi and Yago Vicente, Tomas F and Dideriksen, Thomas and Arora, Himanshu and Guillaumin, Matthieu and Malik, Jitendra},
  journal={CVPR},
  year={2022}
}