GigaBench

Benchmarking Learnable Representations for Gigapixel Image Compression and Super-Resolution

WACV 2027

Learnable image representations are usually tested on images of a few megapixels. GigaBench is the first benchmark to cross the gigapixel line: ten images of more than a billion pixels each, from five domains.

Pick one to explore it at full resolution.

    Domain
    Resolution
    Gigapixels
    Raw size
    License
    Credit

    GigaBench on image compression

    Eighteen configurations of eight methods, each fitted to all ten images. Every point is an average over the benchmark, placed by how good, how small and how fast it is.

    Quality

    Time

    Size

    Raw image bytes divided by model bytes.

    Drag to rotate. Hover a point for its numbers, or pick a family or method below to highlight and label its points.

    Show all numbers

    GigaBench on super-resolution

    Each method is fitted on the image downsampled ×4 or ×8, then queried at full resolution and compared with the original. Classical interpolation is the reference to beat.

    Quality

    Upsampling

    Show all numbers

    Citation

    If you use GigaBench, please cite:

    @inproceedings{fusconi2027benchmarking,
  title     = {{Benchmarking Learnable Representations for Gigapixel Image Compression and Super-Resolution}},
  author    = {Fusconi, Matteo and Costanzino, Alex and Zama Ramirez, Pierluigi and Di Stefano, Luigi},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year      = {2027}
}

    If you use the histopathology images, please also cite their source datasets:

    @article{brancati2022bracs,
  title     = {{BRACS}: A Dataset for {BReAst} Carcinoma Subtyping in {H\&E} Histology Images},
  author    = {Brancati, Nadia and Anniciello, Anna Maria and Pati, Pushpak and Riccio, Daniel and Scognamiglio, Giosu{\`e} and Jaume, Guillaume and De Pietro, Giuseppe and Di Bonito, Maurizio and Foncubierta, Antonio and Botti, Gerardo and others},
  journal   = {Database},
  publisher = {Oxford University Press},
  volume    = {2022},
  year      = {2022}
}

@article{deng2025kpis2024challengeadvancing,
  title   = {{KPIs} 2024 Challenge: Advancing Glomerular Segmentation from Patch- to Slide-Level},
  author  = {Deng, Ruining and Yao, Tianyuan and Tang, Yucheng and Guo, Junlin and Lu, Siqi and Xiong, Juming and Yu, Lining and Cap, Quan Huu and Cai, Pengzhou and Lan, Libin and Zhao, Ze and Galdran, Adrian and Kumar, Amit and Deotale, Gunjan and Das, Dev Kumar and Paik, Inyoung and Lee, Joonho and Lee, Geongyu and Chen, Yujia and Li, Wangkai and Li, Zhaoyang and Hou, Xuege and Wu, Zeyuan and Wang, Shengjin and Fischer, Maximilian and Kramer, Lars and Du, Anghong and Zhang, Le and Sanchez Sanchez, Maria and Sanchez Ulloa, Helena and Ribalta Heredia, David and Perez de Arenaza Garcia, Carlos and Xu, Shuoyu and He, Bingdou and Cheng, Xinping and Wang, Tao and Moreau, Noemie and Bozek, Katarzyna and Innani, Shubham and Baid, Ujjwal and Kefas, Kaura Solomon and Landman, Bennett A. and Wang, Yu and Zhao, Shilin and Yin, Mengmeng and Yang, Haichun and Huo, Yuankai},
  journal = {Medical Image Analysis},
  year    = {2025}
}

@article{tang2026holohisto,
  title   = {{HoloHisto}: End-to-end Gigapixel {WSI} Segmentation with {4K} Resolution Sequential Tokenization},
  author  = {Tang, Yucheng and He, Yufan and Nath, Vishwesh and Guo, Pengfeig and Deng, Ruining and Yao, Tianyuan and Liu, Quan and Cui, Can and Yang, Yuechen and Yin, Mengmeng and Xu, Ziyue and Roth, Holger and Xu, Daguang and Yang, Haichun and Huo, Yuankai},
  journal = {Electronic Imaging},
  year    = {2026}
}