We introduce UnSPAD, an uncertainty-aware burst photography framework capable of fusing and denoising arbitrarily long sequences of binary single photon avalanche diode (SPAD) measurements. Unlike prior methods that require fixed burst lengths or rely on powerful generative priors, UnSPAD combines learned motion estimation with uncertainty-informed measurement fusion and denoising to produce temporally consistent, hallucination-free reconstructions. UnSPAD first forms an initial scene estimate by averaging flow-aligned binary frames. UnSPAD then uses per-pixel uncertainty maps to perform uncertainty-aware denoising; the denoiser is informed about which regions accumulated the fewest observations (and thus contain the most relative shot noise) due to motion, occlusion, or alignment errors. This two-step reconstruction pipeline helps mitigate hallucinations---outside of the structure implicitly introduced by the flow estimates, the initial scenes estimates are entirely prior-free. Experiments on real and simulated SPAD and quantum image sensor (QIS) benchmarks demonstrate state-of-the-art reconstruction quality, improved temporal consistency, and superior color fidelity relative to the state of the art.
Figures & tables
Figure 2: UnSPAD qualitative results on simulated SPAD data. UnSPAD’s robust temporal aggregation with spatial denoising recovers low-blur and low-noise images.
Figure 3: Method overview. UnSPAD aligns 1 -bit SPAD frames guided by optical flow computed on low-bit proxy frames generated by a light-weight denoiser. The aligned binary burst can be merged to produce a high-quality, temporally denoised result. However, due to scene/camera motion or optical flow errors, some regions may be under-observed. In the example, the camera is pan-tilting, causing the outer image edges to have higher uncertainty than the center. A final denoiser can leverage the uncertainty map to guide its strength.
Figure 4: Intermediate reconstructions. Binary frames are warped with flow fields to produce temporally aligned bursts and uncertainty maps. Uncertainty is concentrated near moving objects since those pixels are sampled less often and have a higher likelihood of flow error, which guides the final denoiser.
Figure 5: Real color SPAD samples. Binary bursts captured at 6000 fps with a megapixel passive color SPAD prototype. Notice Burst-gQIR hallucinates semantic details. In column 2, the dice change color and are missing dots. In column 3, a marble not present is shown, and a dice is removed. In column 5, the sticker letters change. UnSPAD does not have similar failures.
Method
ITF ↑
Ewarp∗↓
MANIQA ↑
1 -bit
7.24
228.4
0.233
Average of 77
32.03
18.3
0.169
Burst-gQIR [ 13 ]
30.84
15.0
0.341
UnSPAD (Ours)
39.56
3.5
0.302
Table 1: Temporal consistency and reference-less image quality on real SPAD data. UnSPAD remains competitive in perceptual image quality, while significantly improving temporal consistency.
Figure 6: Color preservation. We project pixel values onto a 2D chromaticity plot to mitigate the effect of lightness. Specifically, we plot the a∗ and b∗ components of the CIELAB colorspace as its meant to preserve perceptual distance. We select 3 patches to highlight the difference in color. Observe UnSPAD’s KDE shape matches more closely with a long average ground truth ( 300 ms), whereas gQIR over saturates as seen by the larger KDE spread.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
E ∗ ↓
QBP [ 27 ]
13.380
0.448
0.496
1.0
QUIVER [ 5 ]
22.429
0.814
0.429
1.7
Burst-gQIR [ 13 ]
29.832
0.856
0.330
2.1
UnSPAD ( 7 : 1 )
33.007
0.914
0.193
0.5
UnSPAD ( 1 : 1 )
31.618
0.902
0.251
0.7
Table 2: Reconstruction on the Burst-gQIR dataset from 77 1 -bit simulated SPAD frames. UnSPAD outperforms existing methods across the board. We also demonstrate robustness to the setting where each ground truth frame is sampled once ( 1 : 1 ) rather than 7× . This is a more accurate model but not adopted in prior work.
Method
PSNR ↑
SSIM ↑
Student-Teacher [ 7 ]
18.720
0.401
RVRT [ 24 ]
19.412
0.354
MemDeblur [ 18 ]
19.488
0.387
EMVD [ 30 ]
20.019
0.587
Spk2ImgNet [ 44 ]
20.395
0.564
FloRNN [ 23 ]
21.034
0.679
Table 3: Reconstruction on I2-2k benchmark from 11 simulated 3 -bit QIS frames. UnSPAD outperforms existing methods in SSIM, and is comparable in PSNR. If switched to the binomial simulator proposed in gQIR [ 13 ] , performance improves greatly, but is not directly comparable to the alternative methods.
Figure 7: Temporal consistency. Two real scenes are visualized, one with motion (left) and another without (right). Each frame is 13 ms after the previous. Since UnSPAD is not stochastic, features are better preserved over time in both cases. Although the diffusion model in gQIR [ 13 ] preserves some semantic information, object appearance is subject to change.
Figure 8: Method component ablation. Each component of UnSPAD improves image quality. Alignment with proxy frames resolves a sharp but noisy image. The uncertainty map helps guide the spatial denoiser to restore some regions more than others.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
Full method
34.95
0.9636
0.2062
No uncertainty
31.49
0.9049
0.4175
No spatial denoising
30.91
0.9196
0.4022
No proxy frames
18.04
0.4528
0.7066
No alignment
20.30
0.5879
0.6413
Table 4: Ablation study. Proper motion compensation is extremely important. Without the uncertainty estimate (second row), or without a spatial denoiser (third row), or if optical flow is computed directly on noisy frames (fourth row), or if no motion model is used at all (last row), image quality significantly degrades.
Table 5: Denoiser ablation. PSNR is reported in dB and runtime in milliseconds for N binary frames. DRUNet in the alignment stage slightly outperforms Restormer in terms of quality, while being substantially faster. For the final spatial denoising, Restormer has a larger quality increase, and since the runtime is dominated by the burst size N , this trade off is worthwhile.
Figure 1: Additional comparison. On the Christmas lights scene, when the lights are off, generative methods hallucinate different color bulbs (green box), or change the vase shape (pink box), UnSPAD on the other had preserves these details.
Figure 2: Photography effects. UnSPAD preserves natural photographic effects such as bokeh and defocus blur, recovering images consistent with the camera’s optics and acquisition settings. In contrast, generative methods tend to hallucinate sharp details, producing visually appealing but potentially semantically inaccurate reconstructions. Burst-gQIR recovers blurred dice as marbles.
Figure 3: Larger bursts improve quality , up to field of view change when quality plateaus. At short burst sizes, the spatial denoiser needs to work harder to recover quality, while at longer bursts the aligned and merged image already has high quality.
Figure 4: Additional qualitative method ablation. Without motion compensation, fine details are quickly lost, while inaccurate flow estimation leaves a noisy result. Our full solution, combining robust temporal aggregation with spatial denoising, recovers low-blur and low-noise images. One limitation is high frequency details (such as the floor in row 2) may get oversmoothed.
Figure 5: Robust alignment across extreme motions. Our method can successfully align 100 kHz binary SPAD frames with virtually no motion blur, even when the motion is in excess of 25 k pixels per second (estimated with our optical flow algorithm).
Figure 6: Single-pixel flow accuracy. As the reconstruction gets perturbed by increasingly strong shifts, image quality decreases. The temporal denoised image does see a small bump at <1 pixel shifts due to bilinear interpolation acting as a spatial denoiser.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
QBP [ 27 ]
13.380
0.448
0.496
QUIVER [ 5 ]
22.429
0.814
0.429
Burst-gQIR [ 13 ]
29.832
0.856
0.330
UnSPAD Temporal
27.894
0.595
0.3757
UnSPAD Full
33.007
0.914
0.193
Table 1: UnSPAD’s temporal denoiser is competitive. Evaluated on the Burst-gQIR dataset with 77 simulated SPAD frames per reconstruction.
Figure 7: QBP comparison on experimental SPAD data. Notice QBP fails to preserve small moving objects, preventing spatial denoiser to improve accuracy.
Figure 8: QBP comparison on simulated data. QBP overestimates brightness and has greater misalignment error. Two spatial denosiers, BM3D and Restormer, barely affect QBP.
Figure 9: Bit2bit comparison. UnSPAD’s result contains residual noise on bit2bit’s simulated sequence.
Image reconstruction recovers images, volumes, or scenes from incomplete, noisy, indirect, or compressed measurements using physical models and learned priors.
Single-photon avalanche diode (SPAD) cameras operate fundamentally differently from conventional cameras due to their photon-counting nature. Each frame produces a binary image: pixels report zero if no photons arrived during exposure, and one if one or more photons arrived. Reconstructing a scene or inferring its properties from a single binary frame is difficult because many different images could produce the same measurement; thus, the inverse problem is fundamentally one-to-many. As we gather more binary measurements, the inherent uncertainty associated with the inverse problem and any associated inference diminishes. With sufficient photon counts, photon noise becomes negligible relative to the signal mean, enabling near-deterministic scene recovery and inference. This work characterizes the transition from stochastic to near-deterministic scene understanding as photon budget increases, analyzing how the stochasticity in photon arrival affects downstream inference tasks. Technically, we develop a conditional generative framework based on a Hypergeometric frame-thinning process for accumulated binary SPAD measurements. Generative models capture the one-to-many nature of photon-starved inverse problems, enabling empirical characterization of how this ambiguity diminishes with increasing measurements and its impact on downstream tasks like character recognition, QR code decoding, and facial analysis.
Capturing high-quality images from only a few detected photons is a fundamental challenge in computational imaging. Single-photon avalanche diode (SPAD) sensors promise high-quality imaging in regimes where conventional cameras fail, but raw quanta frames contain only sparse, noisy, binary photon detections. Recovering a coherent image from a burst of such frames requires handling alignment, denoising, and demosaicing (for color) under noise statistics far outside those assumed by standard restoration pipelines or modern generative models. We present an approach that adapts large text-to-image latent diffusion models to the photon-limited domain of quanta burst imaging. Our method leverages the structural and semantic priors of internet-scale diffusion models while introducing mechanisms to handle Bernoulli photon statistics. By integrating latent-space restoration with burst-level spatio-temporal reasoning, our approach produces reconstructions that are both photometrically faithful and perceptually pleasing, even under high-speed motion. We evaluate the method on synthetic benchmarks and new real-world datasets, including the first color SPAD burst dataset and a challenging Deforming (XD) video benchmark. Across all settings, the approach substantially improves perceptual quality over classical and modern learning-based baselines, demonstrating the promise of adapting large generative priors to extreme photon-limited sensing. Code at https://github.com/Aryan-Garg/gQIR.
Optical flow remains challenging in high-speed and low-light scenes, where the limited frame rate and sensitivity of conventional cameras lead to motion blur and underexposure. Single-photon avalanche diode (SPAD) cameras offer single-photon sensitivity and extremely fine temporal sampling. However, individual slices in these high FPS binary photon streams are too sparse for dense correspondence. Temporal aggregation can provide the spatial cues required by optical flow, but accumulating photons at fixed coordinates blurs moving structures. Motion-aware aggregation can reduce this blur, yet it depends on the flow being estimated. To address this dependency, we propose QuantaFlow, the first method for dense optical flow directly from SPAD streams. Instead of constructing a fixed input representation, QuantaFlow embeds SPAD representation construction into iterative flow refinement. At each iteration, the current flow coarsely aligns the slices within the source and target sub-streams. A photon-flux transformation then constructs multi-scale representations containing intensity and structural cues, while adaptive multi-scale fusion balances photon noise and residual motion blur at each pixel. The fused representations drive a feature-warping flow update, and the refined flow guides representation construction in the next iteration. We further construct a synthetic dataset for SPAD optical-flow training and evaluation. Experiments on the synthetic dataset and real-world SPAD data demonstrate the effectiveness and generalization of QuantaFlow.