InfScene-SR: Seamless Super-Resolution of Arbitrarily Large Remote-Sensing Scenes via Variance-Preserving Joint Denoising

Shoukun Sun; Zhe Wang; Xiang Que; Jiyin Zhang; Xiaogang Ma

cs.CVFeb 23, 2026

InfScene-SR: Seamless Super-Resolution of Arbitrarily Large Remote-Sensing Scenes via Variance-Preserving Joint Denoising

Authors: Shoukun Sun, Zhe Wang, Xiang Que, Jiyin Zhang, Xiaogang Ma

Organizations: University of Idaho, Moscow, Idaho, USA · NCEAS, UC Santa Barbara, Santa Barbara, California, USA · Fujian Agriculture and Forestry University, Fuzhou, China

Abstract

Diffusion models now give the best perceptual quality in super-resolution (SR), but their architecture and training confine them to small fixed crops. Operational remote sensing needs seamless scenes orders of magnitude larger. Joint denoising fuses overlapping tiles at every reverse step and lets text-to-image diffusion generate beyond its training crop, but it assumes deterministic ODE samplers. With the stochastic sampler of SR models such as SR3, the averaging also partly cancels independent per-tile noise. This known variance erosion blurs the detail SR should recover and must be corrected. We carry variance-corrected fusion to conditional SR and derive Spatially-Decoupled Variance Correction (SDVC), an exact reformulation that replaces per-step global normalization with independent per-tile contributions accumulated in one additive pass. SDVC turns the coupled per-step computation into independent tile-local work, so the resulting pipeline, InfScene-SR, runs in parallel across GPUs and makes SR of arbitrarily large scenes feasible. On a 5 $\times$ SR task built from NAIP aerial imagery, we evaluate whole scenes with fidelity, perceptual, seam-continuity, and faithfulness metrics. Under one backbone, InfScene-SR is the only fusion strategy that is seamless and sharp at once, the closest to the low-resolution observation among those that synthesize detail, and within 0.003 IoU of native high-resolution imagery on downstream invasive-plant segmentation. Code is available at https://github.com/TitorX/infscene-sr.

Figures & tables

Figure 1: 5 $\times$ SR of the 21-megapixel NAIP test scene. Stitching independently denoised tiles leaves seams and per-tile drift (left); InfScene-SR denoises 143 overlapping tiles jointly into one seamless, sharp sample (center). Right: detail at a four-tile corner—bicubic is blurry, stitching breaks, ours is sharp and continuous, matching native HR.

Panel of Figure 2 — Figure 2: Reverse step of InfScene-SR. The scene canvas $y_{t}$ is cropped into overlapping views; per-device workers run one sampler step and emit a tile-local SDVC contribution $C_{t}^{(i)}$ (Eq. 7 ), accumulated additively, into the canvas for step $t{-}1$ . Repeated for $t=T,\dots,1$ , the scene evolves as one consistent sample; independent per-tile denoising with stitching (bottom right) leaves seams and per-tile drift.

Figure 3: Variance erosion in the dense-overlap regime. Left: naive averaging (Eq. 3 ) collapses stochastic detail into an over-smoothed reconstruction. Center: variance-corrected fusion (Eq. 5 ) restores sharp texture. Right: ground truth.

Panel of Figure 4 — Figure 4: Left: study area—2024 NAIP over 15 coastal California counties; Santa Barbara County held out. Right: expert iceplant annotations (green) on the 2024 test scene.

Method	Stride	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$ ( ${\times}10^{-2}$ )	LR-PSNR $\uparrow$	SGI $\to 0$
Bicubic	—	28.77	0.715	0.569	90.13	7.20	42.03	$\mathbf{-0.01}$
Independent + stitching	512	17.32	0.583	0.352	65.97	4.02	17.85	4.25
Independent + blending	384	17.66	0.617	0.342	65.95	4.07	18.14	0.88
Joint + naive averaging (Eq. 3 )	384	21.50	0.480	0.649	121.30	9.78	25.30	$-0.03$
InfScene-SR (ours, Eq. 7 )	384	19.55	0.634	0.283	40.78	1.82	20.37	0.03
+ LF projection (Eq. 9 )	384	27.07	0.632	0.217	30.73	0.95	45.10	0.03

Table 1: Whole-scene comparison on the

4329{\times}4818

NAIP test scene (5

\times

SR). All learned rows share one SR3 checkpoint and seed and differ only in the fusion strategy. Bold: best per column (closest to 0 for SGI, Sec. 4.2 ; bicubic is untiled and marks its zero point). The italic row post-processes our output using only the observed LR input (Sec. 5.2 ); it is a diagnostic and is not ranked.

Figure 5: Whole-scene comparison under the identical backbone: reconstruction, per-pixel $|$ error $|$ , and local texture energy relative to HR.

Input	Acc. $\uparrow$	Prec. $\uparrow$	Rec. $\uparrow$	F1 $\uparrow$	IoU $\uparrow$
HR (native)	0.911	0.837	0.889	0.862	0.758
Bicubic	0.895	0.797	0.889	0.841	0.725
Indep. + stitching	0.896	0.837	0.828	0.833	0.713
Indep. + blending	0.906	0.838	0.864	0.851	0.741
Joint + naive avg.	0.883	0.903	0.700	0.788	0.651
InfScene-SR	0.914	0.873	0.849	0.861	0.755

Table 2: Downstream iceplant segmentation (HR-trained U-Net, 2024 test scene). Bold: best per column among reconstructions (HR is the reference, not ranked).

Figure 6: Resident GPU memory of the fusion layer per reverse step (float32; backbone excluded). Centralized VCF (Eq. 5 ) must hold every tile’s state and mean, so its footprint grows with scene area; SDVC holds one batch of tiles at any size (

7\times

less on the 143-tile test scene). Dotted: the

\mathcal{O}(HW)

canvas and accumulator, streamed from host or disk;

\star

: the full 3.4-Gpx NAIP raster of the test county.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Stride	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$ ( ${\times}10^{-2}$ )	LR-PSNR $\uparrow$	SGI $\to 0$
Bicubic	—	28.77	0.715	0.569	90.13	7.20	42.03	$-0.01$
Independent + stitching	512	17.32	0.583	0.352	65.97	4.02	17.85	4.25
Independent + blending	384	17.66	0.617	0.342	65.95	4.07	18.14	0.88
Joint + naive averaging	384	21.50	0.480	0.649	121.30	9.78	25.30	$-0.03$
InfScene-SR	384	19.55	0.634	0.283	40.78	1.82	20.37	0.03
with LF projection (main paper Eq. 9):

Appendix

Table S1 : Complete whole-scene metric suite on the

4329{\times}4818

NAIP test scene. The upper block is main paper Table 1 without its projected row; the lower block applies the LF projection to every strategy. Bold marks the best per column within each block. SGI is a diagnostic and is not ranked (target

0

; negative values mean boundaries smoother than native imagery). After projection every strategy agrees with the observation by construction, so LR-PSNR no longer separates the rows.

Figure S1 : Crop-level comparison across three land-cover types. Rows: ruins and corrals, iceplant mosaic, surf zone. Each inset magnifies the marked square $2.4\times$ .

Figure S2 : The dense-overlap collapse at one further location in each controlled window. Same controlled pairs as main paper Fig. 3, two $1024{\times}1024$ crops at patch $512$ and stride $128$ (up to $22$ tiles per pixel), one checkpoint and one seed ( $42$ ), with avg against sdvc fusion as the only difference and no LF projection. W5 is the rail crossing, W3 the farmstead.

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 248 papers

SlimDiffSR: Toward Lightweight and Efficient Remote Sensing Image Super-Resolution via Diffusion Model Distillation
May 4, 2026Ce Wang, Zhenyu Hu, Wanjie SunDiffusion Model Distillation Remote Sensing Image Super-Resolution
School of Remote Sensing and Information Engineering, Wuhan University; Wuhan 430079, China
Degradation-Aware and Structure-Preserving Diffusion for Real-World Image Super-Resolution
Apr 13, 2026Yang Ji, Zonghao Chen, Zhihao Xue +1Degradation-Aware Image Restoration Image Super-Resolution
KUNBYTE Hangzhou, Zhejiang Province, China · GOLDMYE Xuhui District, Shanghai, China
Single Image Super-Resolution via Bivariate `A Trous Wavelet Diffusion
Mar 7, 2026Maryam Heidari, Nantheera Anantrasirichai, Alin AchimImage Super-Resolution Diffusion Models for Image Restoration
Visual Information Laboratory, University of Bristol, Bristol, UK

May 4, 2026J/K move · Enter open · S save

SlimDiffSR: Toward Lightweight and Efficient Remote Sensing Image Super-Resolution via Diffusion Model Distillation

Ce Wang, Zhenyu Hu, Wanjie Sun

School of Remote Sensing and Information Engineering, Wuhan University; Wuhan 430079, China

Diffusion Model Distillation Remote Sensing Image Super-Resolution Structured Pruning Diffusion Model Inference Acceleration Image Super-Resolution

Open paper AskPDF arXiv

Diffusion models have recently achieved remarkable performance in image super-resolution (SR), but their high computational cost limits practical deployment in remote sensing applications. To address this issue, we propose SlimDiffSR, a lightweight and efficient diffusion-based framework for real-world remote sensing image super-resolution. Unlike existing single-step diffusion methods that rely on fixed timesteps, we first introduce an uncertainty-guided timestep assignment strategy to construct a stronger single-step teacher model, where reconstruction difficulty is explicitly linked to diffusion timesteps, enabling adaptive generative strength. Building upon this teacher, we further present a structured pruning strategy tailored to remote sensing imagery, which systematically removes redundant semantic modules and replaces standard operations with lightweight designs, including frequency-separable convolution, direction-separable convolution, and a query-driven global aggregation module. These components explicitly exploit the unique characteristics of remote sensing data, such as sparse high-frequency details, strong directional patterns, and long-range spatial dependencies. To enhance knowledge transfer, we incorporate Maximum Mean Discrepancy (MMD) into the distillation process to align feature distributions between the teacher and student models. Extensive experiments on multiple remote sensing benchmarks demonstrate that SlimDiffSR achieves a favorable balance between efficiency and reconstruction quality. In particular, it attains up to

200\times

inference acceleration and a

20\times

reduction in model parameters compared with multi-step diffusion models, while achieving competitive perceptual quality and clearly outperforming existing lightweight diffusion baselines in efficiency. The code is available at: https://github.com/wwangcece/SlimDiffSR.

Abstract

Image Super-Resolution

Explore similar work

SlimDiffSR: Toward Lightweight and Efficient Remote Sensing Image Super-Resolution via Diffusion Model Distillation

Degradation-Aware and Structure-Preserving Diffusion for Real-World Image Super-Resolution

Single Image Super-Resolution via Bivariate `A Trous Wavelet Diffusion