DL3DV-2K

📖Paper | 🏠Homepage | 🤗ETCHR-FLUX.2-klein-9B Model | 🤗ETCHR SFT-400K Dataset | 🤗ETCHR GRPO-10K Dataset | 🤗DL3DV-2K Benchmark

DL3DV-2K is a benchmark constructed from the DL3DV dataset for evaluating the viewpoint transformation capability of large models in spatial reasoning tasks, comprising 2K samples in total. Each sample contains: images (original images), aux_images (transformed images provided for human reference only and not used as question input), question, candidates, and answer. The model needs to imagine the viewpoint of the aux_images from the images in order to effectively answer the question.

✒️Citation

If you find this project useful, please kindly cite:

@article{zhang2026etchr,
  title={ETCHR: Editing To Clarify and Harness Reasoning},
  author={Beichen Zhang, Yuhong Liu, Jinsong Li, Yuhang Zang, Jiaqi Wang, Dahua Lin},
  journal={arXiv preprint arXiv:2605.23897},
  year={2026}
}