DSH-Bench: A Difficulty- and Scenario-Aware Benchmark with Hierarchical Subject Taxonomy for Subject-Driven Text-to-Image Generation

Code and Data for DSH-Bench: A Difficulty- and Scenario-Aware Benchmark with Hierarchical Subject Taxonomy for Subject-Driven Text-to-Image Generation

Arxiv: https://arxiv.org/pdf/2603.08090

📊 Data

DSH-Bench Data

The DSH-Bench dataset we collected can be obtained from the following: DSH-bench_data/images

SICS model

The model we trained used SICS method can be obtained from the following link: https://huggingface.co/mapels/DSH-Bench/tree/main

🧑‍💻 Code

1. We have made publicly available the code for all computational metrics used in our experiments, as listed in the directory below: ./code/metric_calculation/
2. The SICS method is implemented based on Llama Factory (https://github.com/hiyouga/LLaMA-Factory).
3. All baseline code we used in our experiment is obtained from official version:
   • Textual Inversion: https://github.com/rinongal/textual_inversion
   • Custom diffusion: https://github.com/adobe-research/custom-diffusion
   • DreamBooth: https://github.com/google/dreambooth
   • λ-Eclipse: http://github.com/eclipse-t2i/lambda-eclipse-inference
   • HiPer: https://github.com/HiPer0/HiPer
   • NeTI: https://github.com/NeuralTextualInversion/NeTI
   • BLIP-Diffusion: https://github.com/salesforce/LAVIS/tree/main/projects/blip-diffusion
   • IP-Adapter: https://github.com/tencent-ailab/IP-Adapter
   • MS-Diffusion: https://github.com/MS-Diffusion/MS-Diffusion
   • OminiControl: https://github.com/Yuanshi9815/OminiControl
   • SSR-Encoder: https://github.com/Xiaojiu-z/SSR_Encoder
   • UNO: https://github.com/bytedance/UNO
   • Emu2: https://github.com/baaivision/Emu
   • RealCustom++: https://github.com/bytedance/RealCustom
   • OmniGen: https://github.com/VectorSpaceLab/OmniGen

Citation

@article{hu2026dsh,
  title={DSH-Bench: A Difficulty-and Scenario-Aware Benchmark with Hierarchical Subject Taxonomy for Subject-Driven Text-to-Image Generation},
  author={Hu, Zhenyu and Wang, Qing and Cao, Te and Liao, Luo and Lu, Longfei and Liu, Liqun and Li, Shuang and Chen, Hang and Xue, Mengge and Chen, Yuan and others},
  journal={arXiv preprint arXiv:2603.08090},
  year={2026}
}