GigaBench: Benchmarking Learnable Representations for Gigapixel Image Compression and Super-Resolution

Matteo Fusconi; Alex Costanzino; Pierluigi Zama Ramirez; Luigi Di Stefano

cs.CVOct 22, 2026

GigaBench: Benchmarking Learnable Representations for Gigapixel Image Compression and Super-Resolution

Authors: Matteo Fusconi, Alex Costanzino, Pierluigi Zama Ramirez, Luigi Di Stefano

Organizations: CVLab, University of Bologna · Ca’ Foscari University of Venice

Abstract

Learnable Representations (LRs) can fit parametric families of functions to diverse kinds of signals. These emergent neural-native data formats are gathering interest in the scientific community due to their compactness, resolution-agnosticism and amenability to direct neural processing. On the other hand, gigapixel images are becoming increasingly widespread across a variety of scientific, cultural, and industrial domains. In this paper, we present GigaBench, the first benchmark and experimental evaluation focused on deploying LRs with gigapixel imagery. GigaBench includes images collected, acquired, and synthetically rendered across diverse domains, and addresses image compression and super-resolution. We adapt a representative set of LRs, spanning Implicit, Hybrid, and Explicit methods, so as to operate reliably at the gigapixel scale on consumer hardware. Through systematic evaluation across reconstruction quality, model size, compression and inference times, we identify consistent trends across methods and suggest actionable guidance for future investigation.

Figures & tables

Figure 1: The images in GigaBench . (1) Venice , (2) Kidney , (3) Calacatta Viola , (4) Andromeda , (5) House Of Chores , (6) Lynxsdesign , (7) Italy , (8) Munich , (9) Breast Cancer , (10) Statuario .

Name	Domain	Resolution (H,W)	GPx	Size [GB]	Sub-Licence	Attribution
Venice	Panorama	$(20000,64000)$	$1.19$	$3.58$	CC BY-NC 4.0	Jeffrey Martin [ 24 ]
Statuario	Industrial	$(34015,33750)$	$1.07$	$3.21$	CC0 1.0	Anonymous
Calacatta Viola	Industrial	$(50040,22677)$	$1.06$	$3.17$	CC0 1.0	Anonymous
Breast Cancer	Scientific	$(29474,36430)$	$1.00$	$3.00$	CC BY-NC 4.0	BRACS [ 7 ]
Kidney	Scientific	$(55967,20800)$	$1.08$	$3.25$	CC BY 4.0	KPIs [ 14 , 47 ]
Munich	Satellite	$(37454,32765)$	$1.14$	$3.43$	CC BY 4.0	BVV [ 4 ]

Table 1: GigaBench Overview. List of the gigapixel images, together with their typology, resolution, raw size in GigaBytes, and original license and attribution. More in the Supplementary.

			Venice	Statuario	Calacatta	Br. Can.	Kidney	Munich	Androm.	Italy	Lynx	Chores	Avg
Implicit Representations	SIREN (5,256)	PSNR	21.55	31.26	22.39	16.91	16.98	14.12	13.16	22.55	20.66	23.09	20.27
		CR	2904.70	2605.17	2575.10	2436.63	2641.72	2784.84	2774.55	2437.49	2439.02	2439.02	2603.82
		CT	17.27	20.59	19.21	18.09	15.72	16.75	20.83	18.07	18.06	18.17	18.28
		IT	44.06	40.85	38.81	37.46	40.20	42.81	42.85	36.54	36.40	36.58	39.66
	FINER (5,256)	PSNR	14.17	32.01	22.86	17.28	14.88	13.00	13.19	22.20	20.73	21.48	19.18
		CR	2904.70	2605.17	2575.10	2436.63	2641.72	2784.84	2774.55	2437.49	2439.02	2439.02	2603.82

Table 2: Compression Performance on GigaBench according to the PSNR (dB), CT (min), IT (s) and CR metrics. Best results in bold , runner-ups underlined . Best PSNRs within each family are highlighted by different colours.

			Venice	Statuario	Calacatta	Br. Can.	Kidney	Munich	Androm.	Italy	Lynx	Chores	Avg
Upsample $\mathbf{\times 4}$	Implicit	SIREN (5,256)	21.72	32.12	26.95	17.35	16.88	14.13	13.19	22.71	20.84	24.05	20.99
		FINER (5,256)	20.55	27.68	24.10	16.39	16.00	13.63	13.13	22.06	19.61	21.25	19.44
		FDHO (5,256)	20.24	27.30	22.79	15.85	14.89	12.98	13.10	21.56	19.28	21.18	18.92
		SIREN (8,1024)	23.00	32.27	27.58	18.43	20.83	16.41	13.22	22.87	21.09	27.37	22.31
		FINER (8,1024)	23.45	33.26	28.44	18.86	22.42	17.03	13.24	23.23	21.20	27.74	22.89
		FDHO (8,1024)	20.79	25.56	23.43	16.42	16.48	13.36	13.14	22.02	19.76	22.38	19.34

Table 3: Super-Resolution Performance on GigaBench . Best results in bold , runner-ups underlined . Best PSNRs within each family are highlighted by different colours.

Figure 4: Elo Rating of the Human Evaluation Study. Top:

\times 4

, Bottom:

\times 8

Panel of Figure 5 — Figure 5: Compression on GigaBench : qualitative results. We show $1024\times 1024$ crops from full-resolution images and consider the best method for each family, as highlighted by the different colours along the columns of Tab. 2 .

Panel of Figure 6 — Figure 6: Super-Resolution on GigaBench : qualitative results. We show $1024\times 1024$ crops from full-resolution images and, for each family, we consider the same method as in the Human Evaluation.

	Spectral			Edge		Entropy		Texture
	$\beta$	$R^{2}$	$E_{\mathrm{HF}}$	$\overline{\lVert\nabla I\rVert}$	$\rho_{\mathrm{edge}}$	$H$ [bits/px]	JXL [bpp]	Contrast	Homog.
Venice	2.03	1.00	0.085	0.022	0.059	7.51	7.25	2.34	0.73
Statuario	2.66	0.99	0.007	0.002	0.001	3.78	2.46	0.10	0.96
Calacatta Viola	2.69	0.99	0.007	0.006	0.008	6.15	6.02	0.27	0.92
Breast Cancer	2.30	1.00	0.082	0.019	0.071	5.34	9.28	2.62	0.81
Kidney ‡	4.28	0.98	0.003	0.015	0.010	7.07	5.74	0.51	0.82

Table 4: Complexity statistics of the GIGABENCH images .

\beta

R^{2}

: power-law fit of the radial PSD;

E_{\mathrm{HF}}

: fraction of spectral energy above

0.25\,f_{\mathrm{Ny}}

;

H

: Shannon entropy of the luminance histogram;

\overline{\lVert\nabla I\rVert}

: mean Scharr gradient;

\rho_{\mathrm{edge}}

: Canny edge density; Contrast, Homog.: GLCM statistics; JXL: lossless JPEG XL rate (RGB). † Non–power-law spectrum (

R^{2}{=}0.74

), see Fig. 7 . ‡ Band-limited by the scanner optics: refitting below/above

0.2\,f_{\mathrm{Ny}}

yields

\beta{=}3.47/5.68

, so the single slope overstates the decay.

Figure 7: Radially averaged power spectra of the ten images (log–log; Hann-windowed, mean-removed full-resolution luminance). Dashed lines: power-law fits over

[0.02,0.80]\,f_{\mathrm{Ny}}

; dotted vertical line: the

E_{\mathrm{HF}}

cutoff

0.25\,f_{\mathrm{Ny}}

. The spectra differ not only in slope but in shape: most images follow a power law closely, Andromeda departs from it with a broad excess of mid-to-high-frequency energy, and Kidney rolls off sharply toward Nyquist, consistent with a band-limited acquisition.

			Venice	Statuario	Calacatta	Br. Can.	Kidney	Munich	Androm.	Italy	Lynx	Chores	Avg
Compression	Implicit	SIREN (5,256)	0.450	0.926	0.823	0.645	0.353	0.143	0.048	0.560	0.218	0.354	0.452
		FINER (5,256)	0.181	0.927	0.825	0.646	0.301	0.114	0.048	0.558	0.220	0.335	0.415
		FDHO (5,256)	0.253	0.923	0.838	0.643	0.296	0.099	0.047	0.558	0.207	0.333	0.420
		SIREN (8,1024)	0.492	0.927	0.839	0.649	0.593	0.227	0.049	0.561	0.230	0.419	0.499
		FINER (8,1024)	0.495	0.926	0.843	0.657	0.608	0.190	0.051	0.565	0.222	0.385	0.494
		FDHO (8,1024)	0.491	0.925	0.834	0.646	0.596	0.160	0.049	0.561	0.218	0.375	0.486

Table 5: SSIM Performance on GigaBench for Compression. Best SSIMs within each family are highlighted by different colours; the best overall is in bold , and runner-ups are underlined .

			Venice	Statuario	Calacatta	Br. Can.	Kidney	Munich	Androm.	Italy	Lynx	Chores	Avg
Upsample $\mathbf{\times 4}$	Implicit	SIREN (5,256)	0.450	0.927	0.838	0.643	0.350	0.142	0.048	0.560	0.221	0.365	0.455
		FINER (5,256)	0.429	0.923	0.828	0.643	0.329	0.129	0.048	0.558	0.210	0.327	0.442
		FDHO (5,256)	0.427	0.923	0.824	0.642	0.311	0.114	0.047	0.558	0.208	0.330	0.438
		SIREN (8,1024)	0.471	0.927	0.840	0.648	0.490	0.198	0.049	0.561	0.227	0.417	0.483
		FINER (8,1024)	0.480	0.929	0.844	0.659	0.547	0.215	0.049	0.565	0.235	0.428	0.495
		FDHO (8,1024)	0.428	0.922	0.828	0.642	0.348	0.123	0.048	0.558	0.213	0.351	0.446

Table 6: SSIM Performance on GigaBench for SR

\times 4

and

\times 8

. Best SSIMs within each family are highlighted by different colours; the best overall is in bold , and runner-ups are underlined .

Figure 8: SSIM vs. Compression Ratio in image compression.

Figure 9: Super-Resolution SSIM performance on GigaBench (left: SR

\times 4

, right: SR

\times 8

			Venice	Statuario	Calacatta	Br. Can.	Kidney	Munich	Androm.	Italy	Lynx	Chores	Avg
Upsample $\mathbf{\times 4}$	Implicit	SIREN (5,256)	0.621	0.119	0.265	0.509	0.775	0.975	0.937	0.576	0.787	0.434	0.600
		FINER (5,256)	0.643	0.121	0.274	0.492	0.801	0.986	0.927	0.586	0.801	0.549	0.618
		FDHO (5,256)	0.617	0.121	0.265	0.479	0.756	0.902	0.931	0.570	0.789	0.486	0.592
		SIREN (8,1024)	0.589	0.120	0.262	0.490	0.519	0.892	0.971	0.566	0.773	0.306	0.549
		FINER (8,1024)	0.567	0.120	0.259	0.458	0.362	0.783	1.056	0.560	0.779	0.253	0.520
		FDHO (8,1024)	0.656	0.122	0.267	0.500	0.738	0.960	0.965	0.593	0.882	0.479	0.616

Table 7: LPIPS Performance on GigaBench for SR

\times 4

and

\times 8

(lower is better). Best LPIPS within each family are highlighted by different colours; the best overall is in bold , and runner-ups are underlined .

Panel of Figure 10 — Figure 10: Motivation for the exclusion of Instant-NGP from the human evaluation, in favor of the MetricGrids model. We show the $\times 4$ super-resolution qualitatives on a crop of Venice, performed by I-NGP (T=24), MetricGrids (T=24) and Nearest Neighbor.

	PSNR (dB)			SSIM			CT (min)			IT (s)
	Mean ± Std	Median ± IQR	Min-Max	Mean ± Std	Median ± IQR	Min-Max	Mean ± Std	Median ± IQR	Min-Max	Mean ± Std	Median ± IQR	Min-Max
SIREN (5,256)	20.27 ± 5.01	21.10 ± 5.58	13.16 - 31.26	0.452 ± 0.273	0.402 ± 0.372	0.048 - 0.926	18.28 ± 1.51	18.08 ± 1.48	15.72 - 20.83	39.66 ± 2.77	39.51 ± 5.52	36.40 - 44.06
FINER (5,256)	19.18 ± 5.62	19.01 ± 7.67	13.00 - 32.01	0.415 ± 0.290	0.318 ± 0.433	0.048 - 0.927	24.02 ± 1.76	23.69 ± 1.28	21.43 - 27.82	71.67 ± 5.42	71.48 ± 8.16	65.03 - 80.69
FDHO (5,256)	18.81 ± 5.14	17.62 ± 7.23	12.60 - 27.16	0.420 ± 0.289	0.315 ± 0.403	0.047 - 0.923	24.53 ± 1.47	23.80 ± 1.28	22.78 - 27.26	30.24 ± 1.77	30.31 ± 2.81	28.11 - 33.15
SIREN (8,1024)	22.71 ± 5.32	23.33 ± 7.67	13.21 - 32.29	0.499 ± 0.263	0.527 ± 0.358	0.049 - 0.927	175.57 ± 11.11	173.55 ± 19.98	163.28 - 192.88	431.35 ± 27.46	428.27 ± 46.33	401.50 - 482.18
FINER (8,1024)	22.67 ± 5.25	23.67 ± 6.12	13.28 - 31.66	0.494 ± 0.270	0.530 ± 0.382	0.051 - 0.926	203.61 ± 12.81	203.23 ± 20.54	189.15 - 229.20	620.27 ± 40.29	614.79 ± 69.78	574.99 - 698.06

Table 8: Dispersion statistics of PSNR, SSIM, Compression Time (CT) and Inference Time (IT) across the benchmark images for the task of Image Compression.

	$\mathbf{\times 4}$						$\mathbf{\times 8}$
	PSNR (dB)			SSIM			PSNR (dB)			SSIM
	Mean ± Std	Median ± IQR	Min-Max	Mean ± Std	Median ± IQR	Min-Max	Mean ± Std	Median ± IQR	Min-Max	Mean ± Std	Median ± IQR	Min-Max
SIREN (5,256)	20.99 ± 5.57	21.28 ± 6.71	13.19 - 32.12	0.455 ± 0.274	0.408 ± 0.369	0.048 - 0.927	20.85 ± 5.53	21.10 ± 6.62	13.19 - 32.07	0.453 ± 0.275	0.404 ± 0.370	0.048 - 0.927
FINER (5,256)	19.44 ± 4.42	20.08 ± 5.76	13.13 - 27.68	0.442 ± 0.277	0.379 ± 0.383	0.048 - 0.923	20.07 ± 4.79	20.48 ± 6.15	13.17 - 28.95	0.447 ± 0.276	0.392 ± 0.378	0.048 - 0.924
FDHO (5,256)	18.92 ± 4.40	19.76 ± 6.34	12.98 - 27.30	0.438 ± 0.279	0.378 ± 0.388	0.047 - 0.923	19.73 ± 5.01	20.06 ± 6.68	13.12 - 29.69	0.443 ± 0.278	0.386 ± 0.385	0.048 - 0.925
SIREN (8,1024)	22.31 ± 5.38	21.98 ± 7.25	13.22 - 32.27	0.483 ± 0.264	0.481 ± 0.352	0.049 - 0.927	21.76 ± 5.55	21.51 ± 7.21	13.21 - 32.43	0.469 ± 0.270	0.448 ± 0.354	0.048 - 0.928

Table 9: Dispersion statistics of PSNR and SSIM across the benchmark images for the task of Image Super Resolution

\times 4

and

\times 8

Figure 11: Win Rates of the Human Evaluation Study (i.e. how many participants voted a certain model as the first preference). The label displays the absolute number of voters.

			Venice	Statuario	Calacatta	Br. Can.	Kidney	Munich	Italy	Lynx	Chores	Overall
Win Rate	$\times 4$	Implicit	1	13	0	2	2	2	0	6	54	80
		Hybrid	22	74	99	88	86	27	66	66	65	593
		Explicit	2	4	4	3	36	32	4	5	0	90
		Classic	145	79	67	77	46	109	100	93	51	767
	$\times 8$	Implicit	1	1	2	5	10	0	0	15	32	66
		Hybrid	41	74	105	88	84	54	80	37	81	644

Table 10: Win Rates and Elo scores of the human evaluation

Panel of Figure 12 — Figure 12: Additional fitting qualitatives results on GigaBench . We show $1228\times 2048$ crops from the full resolution images.

Panel of Figure 13 — Figure 13: Additional fitting qualitatives results on GigaBench . We show $1228\times 2048$ crops from the full resolution images.

Panel of Figure 14 — Figure 14: Additional super resolution $\times 4$ qualitatives results on GigaBench . We show $1228\times 2048$ crops from the full resolution images.

Panel of Figure 15 — Figure 15: Additional super resolution $\times 4$ qualitatives results on GigaBench . We show $1228\times 2048$ crops from the full resolution images.

Panel of Figure 16 — Figure 16: Additional super resolution $\times 8$ qualitatives results on GigaBench . We show $1228\times 2048$ crops from the full resolution images.

Panel of Figure 17 — Figure 17: Additional super resolution $\times 8$ qualitatives results on GigaBench . We show $1228\times 2048$ crops from the full resolution images.

Panel of Figure 18 — Figure 18: Additional fitting qualitative results on GigaBench . We show $2048\times 2048$ crops from the reconstruction of Venice, for all the method variants.

Learned Image Compression

Learned image compression uses neural encoders, quantization, and entropy models to reduce image bitrates while balancing fidelity, perceptual quality, and task utility.

See all 120 papers

Control Your View: High-Resolution Global Semantic Manipulation in Learned Image Compression
May 9, 2026Jiaming Liang, Chi-Man Pun, Weisi Lin +1Learned Image Compression Adversarial Attacks
University of Macau Macau, China · Nanyang Technological University Singapore
GB-LSR: A Fast Local Spectral Image Representation with a Single Global Bandwidth for Continuous Reconstruction and Super-Resolution
Jun 17, 2026Max Shad, Naeem KhoshnevisImplicit Neural Representations Image Reconstruction
Kempner Institute for the Study of Natural and Artificial Intelligence Harvard University
Efficient Generative Image Compression with Quantized Latents Refinement
Oct 21, 2026Hao Cao, Yinhuan Huang, Wenqi Guo +2Generative Image Compression Learned Image Compression
Department of Electronic Engineering, Tsinghua University · Department of Automation, Tsinghua University · State Key Laboratory of Space Network and Communications +1

May 9, 2026J/K move · Enter open · S save

Control Your View: High-Resolution Global Semantic Manipulation in Learned Image Compression

Jiaming Liang, Chi-Man Pun, Weisi Lin, Greta Seng Peng Mok

University of Macau Macau, China · Nanyang Technological University Singapore

Learned Image Compression Adversarial Attacks Gradient Descent Adversarial Examples

Open paper AskPDF arXiv

Learned image compression (LIC) integrates deep neural networks (DNNs) to map high-dimensional images into compact latent representations, reducing redundancy and achieving superior rate-distortion (RD) performance in benign settings. Unfortunately, due to inherent vulnerabilities in DNNs, LIC systems are susceptible to adversarial perturbations that lead to downstream deterioration, compression rate degradation, untargeted distortion, and both local semantic manipulation (LSM) and low-resolution (

3\times28\times28

) global semantic manipulation (GSM). However, high-resolution GSM remains unexplored due to its intractability. Notably, the existing project gradient descent (PGD) method achieves near-perfect white-box attacks for classification, segmentation, and other tasks, yet fails to generalize to high-resolution GSM. Our theoretical and empirical analyses reveal that well-performing GSM drives adversarial examples from the Identity Region to the Amplification Region through the Lazying-Oscillating-Refining stages. General

\ell_{\infty}

-bounded attacks fail on high-resolution GSM because their step-size schedules cannot accommodate both the Oscillating and Refining stages. Based on this, we propose the Periodic Geometric Decay schedule that enables

\ell_{\infty}

-bounded high-resolution GSM. To verify our approach, we integrate it with PGD, yielding a minimal variant, PGD

^{2}

-GSM. Extensive experiments on the Kodak

(3\times768\times512)

demonstrate that our PGD

^{2}

-GSM is the first to stably achieve high-resolution GSM, thereby exposing a novel threat to LIC systems. Code is available at https://github.com/chinaliangjiaming/PGD2-GSM.

Abstract

Learned Image Compression

Explore similar work

Control Your View: High-Resolution Global Semantic Manipulation in Learned Image Compression

GB-LSR: A Fast Local Spectral Image Representation with a Single Global Bandwidth for Continuous Reconstruction and Super-Resolution

Efficient Generative Image Compression with Quantized Latents Refinement