GigaBench

Benchmarking Learnable Representations for Gigapixel Image Compression and Super-Resolution (WACV 2027)

Matteo Fusconi¹, Alex Costanzino¹, Pierluigi Zama Ramirez², Luigi Di Stefano¹
¹ CVLab, University of Bologna · ² Ca' Foscari University of Venice

Project page · Code (coming soon) · Paper (coming soon)

TL;DR

GigaBench is a benchmark of 10 lossless gigapixel images (1.00–1.19 GPx each) from five domains, used to test whether Learnable Representations (LRs) — Implicit, Hybrid and Explicit — work at gigapixel scale for image compression and super-resolution on a single consumer GPU.

• LRs can reach the gigapixel regime, but only with dedicated adaptation (tiling, gradient accumulation), and fitting cost becomes a selection criterion in its own right.
• Compression: Hybrid methods give the best quality at moderate compression ratios, with Instant-NGP offering the best quality vs. compression-time trade-off. At extreme ratios, Implicit methods and SmartSplat remain viable; SmartSplat is competitive across the whole range and decodes fastest.
• Super-resolution: classical interpolation is still preferred by humans at ×4; at ×8, Hybrid methods match it.
• PSNR and human preference disagree, so human assessment should be a primary signal, not a supplementary one.

Dataset

Image	File	Domain	Resolution (H × W)	GPx	Raw size	License	Attribution
Venice	Venice.tif	Panorama	20,000 × 64,000	1.19	3.58 GB	CC BY-NC 4.0	Jeffrey Martin
Statuario	Statuario.tif	Industrial	34,015 × 33,750	1.07	3.21 GB	CC0 1.0	Anonymous
Calacatta Viola	CalacattaViola.tif	Industrial	50,040 × 22,677	1.06	3.17 GB	CC0 1.0	Anonymous
Breast Cancer	BreastCancer.tif	Scientific	29,474 × 36,430	1.00	3.00 GB	CC BY-NC 4.0	BRACS
Kidney	Kidney.tif	Scientific	55,967 × 20,800	1.08	3.25 GB	CC BY 4.0	KPIs Challenge 2024
Munich	Munich.tif	Satellite	37,454 × 32,765	1.14	3.43 GB	CC BY 4.0	Bayerische Vermessungsverwaltung
Andromeda	Andromeda.tif	Satellite	22,230 × 55,000	1.14	3.42 GB	CC BY 4.0	ESA/Hubble
Italy	Italy.tif	Satellite	30,654 × 35,040	1.00	3.00 GB	Copernicus open data	Contains modified Copernicus Sentinel data (2025)
Lynxsdesign	Lynxsdesign.tif	Synthetic	24,588 × 43,712	1.00	3.00 GB	CC BY 4.0	Lynxsdesign
House of Chores	HouseOfChores.tif	Synthetic	24,588 × 43,712	1.00	3.00 GB	CC BY 4.0	Blender Studio

Selection criteria:

• Gigapixel resolution: every image meets or marginally exceeds 1 GPx (1.00–1.19).
• Domain diversity: panoramas, industrial scans, histopathology slides, satellite and astronomical imagery, synthetic renders.
• Lossless appearance: no JPEG-compressed sources, so that artifacts of the representations are not mixed with pre-existing compression artifacts.
• Licensing: only images whose original license permits research use and redistribution.

Format: RGB, 8 bits per channel, LZW-compressed TIFF (images/*.tif). metadata.jsonl lists, for every image, its file, name, original license, attribution, source URL, height, width and gigapixels. GPx and GB are computed in powers of 2 (2³⁰ pixels, 2³⁰ bytes of raw RGB).

Download

The files are too large for the dataset viewer and for most image libraries' default limits. Download them directly and read them with tifffile or pyvips:

from huggingface_hub import snapshot_download
import tifffile

root = snapshot_download("Fusco6/GigaBench", repo_type="dataset")
venice = tifffile.imread(f"{root}/images/Venice.tif")   # (20000, 64000, 3) uint8, ~3.6 GB in RAM

Tasks and metrics

Compression. Each method is fitted to the full-resolution image and decoded back at the same resolution.

Metric	Measures
PSNR (dB)	Pixel-level fidelity
SSIM	Structural similarity
Compression Ratio (CR)	Raw image bytes / stored representation bytes: H·W·3·sizeof(uint8) / (#Params·sizeof(float32))
Compression Time (CT)	Time to fit the representation to the image
Inference Time (IT)	Time to decode the full image from the fitted representation

Super-resolution. Images are downsampled ×4 and ×8 (Lanczos); each method is fitted on the low-resolution image and queried on the original high-resolution grid, then compared with the original.

Metric	Measures
PSNR (dB), SSIM	Distortion with respect to the original
LPIPS	Learned perceptual distance
Human preference (Elo)	Ranking study with 170 participants on 2048² crops, scored with the Plackett–Luce model and mapped to Elo

PSNR, SSIM and CR are intrinsic to the representation. CT and IT are runtime measurements of each method's reference implementation, all on the same machine: one NVIDIA RTX 5090 (32 GB), 128 GB RAM.

Methods benchmarked

Family	Method	Configurations
Implicit	SIREN	(L, F) = (5, 256) and (8, 1024) hidden layers and features; CR ≈ 2600 and ≈ 100
Implicit	FINER	as SIREN
Implicit	FDHO	as SIREN
Hybrid	Instant-NGP	hash-table size T = 2²² and 2²⁴; CR ≈ 15 and ≈ 5
Hybrid	NeuRBF	as Instant-NGP
Hybrid	MetricGrids	as Instant-NGP
Explicit	GaussianImage	number of Gaussians set for CR = 15, 100 and 3000
Explicit	SmartSplat	as GaussianImage
Classical (super-resolution only)	Nearest neighbour, bicubic, Lanczos	—

Gaussian splatting at gigapixel scale does not fit in GPU memory, so Explicit methods are trained on tiles with gradient accumulation, keeping the same theoretical result as full-image training.

License

The dataset bundle is released under CC BY-NC 4.0 (see LICENSE.txt), with attribution to GigaBench. Each image also keeps its original license and attribution, listed above: when you use an image, credit its original author as well.

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}
}