CAFlow: Adaptive-Depth Single-Step Flow Matching for Efficient Histopathology Super-Resolution

Elad Yoshai; Ariel D. Yoshai; Natan T. Shaked

cs.CVMar 19, 2026

CAFlow: Adaptive-Depth Single-Step Flow Matching for Efficient Histopathology Super-Resolution

Authors: Elad Yoshai, Ariel D. Yoshai, Natan T. Shaked

Organizations: School of Biomedical Engineering, Tel Aviv University, Tel Aviv, Israel

Abstract

In digital pathology, whole-slide images routinely exceed gigapixel resolution, making computationally intensive generative super-resolution (SR) impractical for routine deployment. We introduce CAFlow, an adaptive-depth single-step flow-matching framework that routes each image tile to the shallowest network exit that preserves reconstruction quality. CAFlow performs flow matching in pixel-unshuffled rearranged space, reducing spatial computation by 16x while enabling direct inference. We show that dedicating half of training to exact t=0 samples is essential for single-step quality (-1.5 dB without it). The backbone, FlowResNet (1.90M parameters), mixes convolution and window self-attention blocks across four early exits spanning 3.1 to 13.3 GFLOPs. A lightweight exit classifier (~6K parameters) achieves 33% compute savings at only 0.12 dB cost. On multi-organ histopathology x4 SR, adaptive routing achieves 31.72 dB PSNR versus 31.84 dB at full depth, while the shallowest exit exceeds bicubic by +1.9 dB at 2.8x less compute than SwinIR-light. The method generalizes to held-out colon tissue with minimal quality loss (-0.02 dB), and at x8 upscaling it outperforms all comparable-compute baselines while remaining competitive with the much larger SwinIR-Medium model. Downstream nuclei segmentation confirms preservation of clinically relevant structure. The model trains in under 5 hours on a single GPU, and adaptive routing can reduce whole-slide inference from minutes to seconds.

Figures & tables

Panel of Figure 1 — Figure 1: CAFlow architecture overview. (a) The pipeline operates entirely in pixel-unshuffled rearranged space at $\frac{H}{4}{\times}\frac{W}{4}$ resolution, reducing spatial compute by $16{\times}$ . (b) FlowResNet backbone (1.90M parameters): 16 blocks grouped into 4 exit segments with increasing hybrid (attention) block density, creating a non-linear cost gradient ( $4.3{\times}$ ratio). The ExitClassifier ( ${\sim}$ 6K parameters) predicts the optimal exit from E0 features. (c) Internal structure of FiLMResBlock (Eq. 5 ) and HybridFiLMBlock (Eq. 6 ). Dashed arrows denote residual connections; HybridFiLMBlock augments the convolutional path with W-MSA/SW-MSA ( $w{=}8$ , 8 heads) and a feed-forward MLP.

Method	Params	GFLOPs	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Bicubic	—	—	29.24	0.8096	0.2495
EDSR	1.52M	16.25	31.44	0.8316	0.2003
SwinIR-light	0.93M	8.72	31.76	0.8494	0.1953
SwinIR-Medium	11.90M	107.11	31.84	0.8790	0.1958
SRFormer-light	0.87M	8.25	31.55	0.8718	0.1955
SR3	2.94M	4376	31.39	0.8672	0.1917

Table 1: Quantitative comparison on multi-organ histopathology

\times

4 SR (343 validation patches). Best in bold , second-best underlined . SR3 GFLOPs reflect total cost of 40 forward passes (10 DDIM steps

\times

4 averaged samples).

Method	Params	GFLOPs	ms	img/s	VRAM
Bicubic	—	—	0.9	1,125	2
EDSR	1.52M	16.25	20.8	48	30
SwinIR-light	0.93M	8.72	60.1	17	99
SwinIR-Medium	11.90M	107.11	84.1	12	129
SRFormer-light	0.87M	8.25	44.9	22	106
SR3	2.94M	4,376	1,082.1	1	145

Table 2: Inference efficiency on a single RTX 5070 Ti (16 GB), bfloat16, batch 1,

64{\times}64

LR input (

\times

4 SR). Latency averaged over 200 runs after 20 warmup passes. Best learned method per metric in bold .

Exit	Blocks	Hybrid	GFLOPs	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
E0	0–3	0	3.10	31.17	0.8625	0.2003
E1	0–7	1	6.07	31.60	0.8717	0.1969
E2	0–11	3	9.39	31.72	0.8740	0.1971
E3	0–15	6	13.34	31.84	0.8797	0.1961
Adaptive	—	—	8.92	31.72	0.8737	0.1967

Table 3: Per-exit quality and compute breakdown for CAFlow on

\times

4 SR. Hybrid shows the number of HybridFiLMBlocks processed at each exit.

Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$\Delta$ PSNR
Full CAFlow	31.72	0.8737	0.1967	–
w/o early exit	31.65	0.8742	0.1959	$-0.07$
w/o consistency	31.55	0.8717	0.1975	$-0.17$
w/o $t{=}0$ mixing	30.17	0.8414	0.2175	$-1.54$
w/o SSIM loss	31.41	0.8680	0.1978	$-0.30$
Fewer blocks (8)	31.41	0.8677	0.1968	$-0.31$

Table 4: Ablation study on

\times

4 SR. Each row removes one component from the full CAFlow model.

\Delta

PSNR shows the drop from the full model. Ablations trained for 300 epochs; full model for 700 epochs. Best per metric in bold .

Figure 2: Compute-quality Pareto frontier. CAFlow at different operating points (Exit 0 through Exit 3 and Adaptive) compared against fixed baselines including SwinIR-Medium. The adaptive strategy achieves near-full-depth quality at reduced compute. SwinIR-Medium reaches similar PSNR to CAFlow Exit 3 only at substantially higher compute, while the CAFlow frontier remains stronger in the practical 3.1 to 13.3 GFLOPs regime.

Figure 3: Exit distribution across the 343 validation images under adaptive routing. The classifier routes a mix of images to each exit, with many easy images handled at E0/E1 and harder images pushed to E2/E3.

Figure 4: Spatial exit assignment on a held-out TCGA-BRCA frozen tissue slide (TCGA-A8-A08C, not used during training) tiled at 40 $\times$ into 1024 $\times$ 1024 patches. After tissue detection (Otsu thresholding), 301 tissue tiles are retained and colored by their router-assigned exit: blue (E1, 6.07 GFLOPs), orange (E2, 9.39 GFLOPs), or red (E3, 13.34 GFLOPs). The router assigns 33% to E1 and 61% to E2, with only 6% routed to E3, averaging 8.55 GFLOPs (a 36% reduction versus full-depth processing).

Figure 5: Qualitative comparison on $\times$ 4 histopathology SR across breast, kidney, and lung tissue. All examples are taken from the deterministic validation split, using organ-specific patches with highlighted ROIs. Colored boxes mark structures where CAFlow preserves clearer or competitive nuclear separation, lumen definition, or alveolar/stromal continuity than competing methods. Across the highlighted ROIs, CAFlow matches SwinIR-Medium in mean PSNR within 0.03 dB (25.61 vs. 25.63 dB) while achieving higher mean SSIM (0.760 vs. 0.751). Although SwinIR-Medium and SR3 can be visually competitive, they are roughly one to two orders of magnitude heavier in compute and memory.

	Breast			Kidney			Lung
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Bicubic	30.64 $\pm$ 4.41	0.8331 $\pm$ 0.0928	0.2293 $\pm$ 0.0523	27.35 $\pm$ 5.76	0.7777 $\pm$ 0.0957	0.2765 $\pm$ 0.0686	27.80 $\pm$ 3.92	0.7932 $\pm$ 0.0867	0.2639 $\pm$ 0.0527
EDSR	32.97 $\pm$ 4.42	0.8834 $\pm$ 0.0727	0.1837 $\pm$ 0.0506	29.57 $\pm$ 5.30	0.8499 $\pm$ 0.0734	0.2204 $\pm$ 0.0642	30.13 $\pm$ 4.06	0.8574 $\pm$ 0.0659	0.2138 $\pm$ 0.0525
SwinIR-light	33.23 $\pm$ 4.40	0.8883 $\pm$ 0.0701	0.1808 $\pm$ 0.0496	30.05 $\pm$ 5.37	0.8611 $\pm$ 0.0675	0.2148 $\pm$ 0.0586	30.46 $\pm$ 3.97	0.8650 $\pm$ 0.0627	0.2099 $\pm$ 0.0511
SwinIR-Medium	33.39 $\pm$ 4.40	0.8913 $\pm$ 0.0687	0.1789 $\pm$ 0.0497	30.27 $\pm$ 5.33	0.8665 $\pm$ 0.0647	0.2165 $\pm$ 0.0579	30.66 $\pm$ 4.04	0.8691 $\pm$ 0.0611	0.2099 $\pm$ 0.0513
SRFormer-light	33.04 $\pm$ 4.36	0.8852 $\pm$ 0.0714	0.1812 $\pm$ 0.0494	29.84 $\pm$ 5.39	0.8559 $\pm$ 0.0706	0.2125 $\pm$ 0.0605	30.26 $\pm$ 3.92	0.8611 $\pm$ 0.0639	0.2088 $\pm$ 0.0510

Table 5: Per-organ quantitative comparison on multi-organ histopathology

\times

4 SR. Values are mean

\pm

std across validation patches per organ. Best in bold , second-best underlined .

Method	Precision $\uparrow$	Recall $\uparrow$	Detection F1 $\uparrow$	Count $r$ $\uparrow$	Mean IoU $\uparrow$
Bicubic	0.863 $\pm$ 0.152	0.886 $\pm$ 0.132	0.869 $\pm$ 0.132	0.9799	0.847 $\pm$ 0.080
EDSR	0.898 $\pm$ 0.127	0.926 $\pm$ 0.095	0.905 $\pm$ 0.109	0.9878	0.879 $\pm$ 0.065
SwinIR-light	0.906 $\pm$ 0.122	0.928 $\pm$ 0.099	0.911 $\pm$ 0.108	0.9886	0.883 $\pm$ 0.064
SwinIR-Medium	0.895 $\pm$ 0.126	0.936 $\pm$ 0.102	0.909 $\pm$ 0.111	0.9868	0.884 $\pm$ 0.068
SRFormer-light	0.913 $\pm$ 0.111	0.916 $\pm$ 0.111	0.910 $\pm$ 0.101	0.9895	0.885 $\pm$ 0.042
SR3	0.901 $\pm$ 0.122	0.924 $\pm$ 0.100	0.906 $\pm$ 0.106	0.9866	0.880 $\pm$ 0.065

Table 6: Downstream nuclei detection on SR outputs. StarDist ( Schmidt et al., 2018 ) “2D_versatile_he” detects nuclei on each method’s SR output; metrics compare against detections on HR ground truth (IoU threshold = 0.5). All learned SR methods significantly outperform bicubic interpolation; differences among learned methods are not statistically significant (paired

t

-test,

p>0.05

), indicating that downstream clinical utility has saturated while computational cost varies widely (Table 2 ).

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Bicubic	29.58 $\pm$ 4.72	0.8104 $\pm$ 0.0815	0.2469 $\pm$ 0.0595
EDSR	31.49 $\pm$ 4.44	0.8620 $\pm$ 0.0630	0.2030 $\pm$ 0.0579
SwinIR-light	31.75 $\pm$ 4.34	0.8678 $\pm$ 0.0610	0.1994 $\pm$ 0.0564
SwinIR-Medium	31.90 $\pm$ 4.40	0.8709 $\pm$ 0.0602	0.1991 $\pm$ 0.0574
SRFormer-light	31.60 $\pm$ 4.30	0.8650 $\pm$ 0.0617	0.1987 $\pm$ 0.0557
SR3	31.42 $\pm$ 4.54	0.8592 $\pm$ 0.0655	0.1931 $\pm$ 0.0536

Table 7: Zero-shot cross-organ generalization on held-out colon tissue (

\times

4 SR). Models trained on breast/kidney/lung, evaluated on entirely unseen colon organ. Best in bold , second-best underlined .

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Bicubic	24.13 $\pm$ 4.58	0.6149 $\pm$ 0.1538	0.4303 $\pm$ 0.0733
EDSR	25.43 $\pm$ 4.45	0.6684 $\pm$ 0.1410	0.3800 $\pm$ 0.0697
SwinIR-light	25.30 $\pm$ 4.46	0.6616 $\pm$ 0.1399	0.3785 $\pm$ 0.0680
SwinIR-Medium	26.52 $\pm$ 4.61	0.7105 $\pm$ 0.1316	0.3555 $\pm$ 0.0711
SRFormer-light	25.23 $\pm$ 4.44	0.6584 $\pm$ 0.1406	0.3820 $\pm$ 0.0678
SR3	25.57 $\pm$ 4.61	0.6728 $\pm$ 0.1430	0.3504 $\pm$ 0.0659

Table 8: Quantitative comparison on multi-organ histopathology

\times

8 SR. Best in bold , second-best underlined .

Figure 6: Qualitative comparison on $\times$ 8 histopathology SR across breast, kidney, and lung tissue. All examples are taken from the deterministic validation split, using organ-specific patches with highlighted ROIs. Colored boxes mark structures where CAFlow preserves clearer or competitive nuclear separation and boundary continuity than competing methods at this harder scale. Across the highlighted ROIs, CAFlow matches SwinIR-Medium in mean PSNR within 0.03 dB (23.55 vs. 23.57 dB) while achieving higher mean SSIM (0.624 vs. 0.612). Although SwinIR-Medium and SR3 can be visually competitive, they are roughly one to two orders of magnitude heavier in compute and memory.

Medical Image Super-Resolution

Medical image super-resolution reconstructs fine spatial detail from low-resolution scans to support clearer anatomy and more useful clinical measurements.

See all 57 papers

Morphology-Aware Implicit Super-Resolution Network for Pathological Images
Aug 4, 2026Jiaming Liang, QiHui Han, Haolin Chen +5Computational Pathology Implicit Neural Representations
School of Computer Science and Engineering, South China University of Technology, Guangzhou, China · School of Software Engineering, South China University of Technology, Guangzhou, China · School of Future Technology, South China University of Technology, Guangzhou, China +1
Fast Image Super-Resolution via Consistency Rectified Flow
May 12, 2026Jiaqi Xu, Wenbo Li, Haoze Sun +8Rectified Flow Consistency Distillation
The Chinese University of Hong Kong · Huawei Noah’s Ark Lab · HKUST (GZ) +1
BigCell: Generating Gigapixel Whole-Slide Images
Oct 9, 2026Srikar Yellapragada, Alexandros Graikos, Zilinghan Li +10Image Generation Medical Image Synthesis
Stony Brook University · Argonne National Laboratory · University of Chicago +2

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

Morphology-Aware Implicit Super-Resolution Network for Pathological Images

Jiaming Liang, QiHui Han, Haolin Chen, Chengxin Ye+4

School of Computer Science and Engineering, South China University of Technology, Guangzhou, China · School of Software Engineering, South China University of Technology, Guangzhou, China · School of Future Technology, South China University of Technology, Guangzhou, China+1

Computational Pathology Implicit Neural Representations Medical Image Super-Resolution Image Super-Resolution

Open paper AskPDF arXiv

Accurate diagnosis in Digital Pathology (DP) relies on high-resolution whole-slide images, yet clinical deployment is often limited by hardware costs. Super-Resolution (SR) offers a promising alternative by computationally enhancing low-resolution acquisitions. However, existing SR methods frequently struggle to preserve fine-grained cellular morphology, leading to texture oversmoothing and blurred structural boundaries under complex tissue variability. To address this issue, we propose Morph-ISR, a morphology-aware implicit super-resolution framework for DP that restores diagnostically relevant details with sub-pixel precision. Morph-ISR reformulates SR as a continuous coordinate-based reconstruction problem and integrates an Implicit Position-aware Kernel Generator (IPKG) to adaptively model spatially varying tissue morphology. To further enhance structural fidelity, a Morphological Fidelity Prior (MFP) is introduced, leveraging semantic guidance from a pre-trained cell segmentation network to enforce boundary-preserving and region-aware reconstruction, thereby improving the representation of critical cellular boundaries and nuclear textures. Experiments on TCGA and SurGen datasets show that Morph-ISR achieves the best LPIPS and ST-LPIPS among the evaluated methods, reducing them by up to 38.37% and 39.55%, respectively, over the second-best methods while maintaining strong PSNR and SSIM. These results demonstrate superior preservation of diagnostically relevant cellular boundaries and nuclear textures, while compact parameterization and high throughput support efficient edge deployment. Code and trained models will be released upon publication.

Abstract

Medical Image Super-Resolution

Explore similar work

Morphology-Aware Implicit Super-Resolution Network for Pathological Images

Fast Image Super-Resolution via Consistency Rectified Flow

BigCell: Generating Gigapixel Whole-Slide Images