WorksEngineering
D2ToF: One-Step Diffusion for Depth Super-Resolution from Lightweight Single-Photon LiDAR Waveforms
Abstract
Single-photon LiDAR (SPL) sensors offer exceptional temporal resolution and single-photon sensitivity for depth imaging. However, cost-effective SPL arrays typically suffer from extremely low spatial resolutions (e.g. pixels), limiting their real-world utility. In this paper, we propose a novel, one-step super-resolution framework, D2ToF, Distilled Diffusion for Time-of-Flight, that bridges the gap between high-fidelity generative modeling and real-time execution. Our approach uses an amortized variational distillation strategy that allows for a measurement-based data consistency term to be incorporated directly into the one-step training objective. This encourages that the synthesized high-resolution depth maps are consistent with the captured low-resolution Time-of-Flight (ToF) histograms. To overcome the severe ill-posedness of upsampling extremely degraded inputs, we introduce a multimodal guidance strategy. By leveraging a foundational monocular depth model to extract structural priors from a high-resolution reflectivity image, our network is guided by a rich triad of the ToF histogram, reflectivity map, and relative depth estimate. Experimental results outperform evaluated baselines, achieving spatial super-resolution (from histograms to depth maps) at an accelerated inference speed of approximately milliseconds.
Introduction
Depth imaging supports a broad range of applications including 3D reconstruction for autonomous navigation [1], [2] and consumer electronics [3]. Among depth-sensing technologies, single-photon LiDAR (SPL) is attractive for its exceptional temporal resolution and single-photon sensitivity [4], [5], which enable accurate ranging even at long range and under low illumination. Cost-effective lightweight SPL sensors however, provide only a small number of pixel waveforms; typically 8×8 or 16×16 [6], [7]. This severe spatial limitation restricts their practical use and motivates computational super-resolution (SR) approaches to recover a high-resolution depth map from such measurements.
To compensate for the low spatial resolution, sensor-fusion methods use a co-registered high-resolution (HR) intensity or reflectivity image to guide the upsampling of low-resolution (LR) LiDAR data [10], [12]–[19]. These methods span statistical reconstruction [20]–[22], learning-based feature fusion [6], [10], [12], [13], and, more recently, to generative models using conditional diffusion [19]. Although these approaches improve reconstruction quality, plug-and-play methods [6] and
iterative diffusion [11], [19] incur substantial inference costs (Figure 1). Recovering detailed depth from a single frame becomes particularly challenging when the LiDAR provides only 8 × 8 waveforms (Figure 2). Earlier approaches exploit multiple frames from a video sequence [6], [10] or operate at a higher base spatial resolution [19], leaving efficient reconstruction from a single frame at such low spatial resolution insufficiently explored.

Figure 1. Reconstruction quality (RMSE in mm) vs. time (sec.) of our proposed method in comparison to DELTAR [8], DepthAnything-V2 [9], DVSR [10], and Marigold-DC [11] evaluated on 12 clean scenes of Middelbury dataset.

Figure 2. Real SPAD measurement and corresponding reconstruction. Right: (top) high-resolution guidance image with the selected field of view, (middle) spatial location of the selected ToF cells within the measured array, (bottom) one-step reconstructed depth. Left: photon-arrival waveforms for the highlighted ToF cells, showing the temporal data available.
We address both obstacles with a framework built on amortized variational distillation. We first learn a multi-step conditional diffusion teacher over HR depth maps and then distil this distribution into a one-step generator. Following the amortized formulation of [23], a data-fidelity term is integrated to enforce consistency of the reconstruction with the observations. We re-render the predicted HR depth back into a LR ToF histogram and penalise its deviation from the measured histogram. This term anchors the reconstruction to the acquired data and, as our ablations show, improves robustness when the guiding image is degraded.
Reconstructing from an 8×8 waveform also requires strong spatial priors, since the histogram itself carries limited layout. An HR reflectivity image supplies fine texture but is inherently ambiguous about 3D geometry. We therefore add a monocular depth guidance: we pass the reflectivity image through a frozen foundational depth model (e.g. DepthAnything V2 [9]) to obtain a relative depth estimate, and condition the network on the resulting triad of ToF histogram, reflectivity map, and monocular depth. The monocular depth estimator adds a small, fixed cost at inference while contributing dense geometric structure, object boundaries and surface layout that the raw waveform lacks.
We evaluate the method on simulated Middlebury scenes at different photon levels and on real captures from a cost-effective STMicroelectronics SPAD sensor, including low-light acquisitions. Across these settings the proposed one-step model outperforms the learning-based [8, 10, 11], and monocular baselines [9] we compare against, while reconstructing a depth map in approximately 83 ms—about 130× faster than the multi-step teacher it is distilled from, while achieving lower average RMSE on the reported benchmark.
In summary, our main contributions are as follows:
We propose a one-step framework for extreme single-photon depth super-resolution that reconstructs a depth map from an histogram in a single forward pass (≈83 ms), roughly 130× faster than the 250-step (DDPM) teacher it is distilled from.
We introduce a differentiable histogram data-consistency term, derived from the single-photon observation model and embedded in the amortized objective. This anchors the reconstruction towards the measured histogram.
We propose a multimodal conditioning strategy that combines the ToF histogram, an HR reflectivity image, and a frozen foundational monocular-depth estimate, enabling the recovery of geometric structure that is absent from the raw waveform.
We validate the method on simulated and real SPAD data, including low-light conditions, showing strong performance and robustness to degraded reflectivity guidance without retraining.
The remainder of this paper is organised as follows. § 2 reviews the single-photon observation model and the background on diffusion models and distillation that our method builds upon. § 3 presents the proposed framework: the conditional diffusion teacher, the multimodal conditioning strategy, the network architecture, and the one-step distillation framework with its histogram data-consistency term. Section IV reports results on simulated Middlebury scenes and on real captures from an SPAD sensor, and § 5 presents the ablation studies. Section VI concludes the paper. Code will be available upon acceptance.
Background
Single-photon LiDAR Observation model
A LiDAR system emits light pulses and records the ToF of reflected photons across pixel locations. For each pixel, a histogram waveform of photon arrivals is constructed with denoting the photon count at pixel and time bin . The waveform follows a Poisson distribution [21, 24]:
where , are the target’s depth and reflectivity parameters, represents the system impulse response function (IRF) assumed Gaussian with variance , and denotes the background counts due to the observation environment or detector’s dark counts.
Given the histogram waveforms , , the goal is to estimate the target’s depth maps to reconstruct the 3D scene. A classical strategy is the log-matched filter algorithm, equivalent to the maximum likelihood estimator (MLE) in absence of background (i.e., ), given by
Observation model for cost-effective SPAD sensors
Recent advances in single-photon technology have enabled the development of SPAD arrays for fast imaging. However, cost-effective sensors often only provide a limited number of pixel waveforms (e.g., ), resulting in low-resolution images where each pixel covers a large surface of the scene. This impacts the observation model and can be expressed as
where denotes a spatial convolution–decimation operator modeling the limited spatial resolution of the sensor. In such cases, the classical log-matched filter algorithm performs poorly due to the presence of multiple signal peaks in each waveform, motivating the need for advanced algorithms that exploit full LiDAR waveform for improved depth estimation.
In the dense-photon regime, Poisson count statistics can be approximated by Gaussian statistics [25, 26]. Following the Gaussian histogram models adopted for SPL in [6, 27]:
where denotes a noise variance. Under this model, the negative log-likelihood reduces to a squared distance, which we adopt as the data-fidelity term in § 3.
The high-resolution reflectivity image that guides the reconstruction is acquired by a conventional camera and is therefore itself photon-limited, particularly in low-light conditions. We model its degradation with the standard Poisson-Gaussian model [28, 29, 30],
where denotes the effective photon-count level and controls the signal-dependent shot noise (smaller yields stronger noise), while denotes the standard deviation of the signal-independent read noise . In Sec. IV, the networks are trained using clean guidance images only, while is introduced exclusively at test time to evaluate robustness to degradation of the guidance modality.
Diffusion Models
Diffusion models [31, 32] are formulated around a forward noising process, denoted , which is governed by a stochastic differential equation (SDE). This process operates over a continuous time interval , starting with a sample from the true data distribution at . As time progresses, the SDE gradually perturbs the data, ultimately transforming it into a sample from a simple prior, , at , where denotes identity matrix. The general form of this forward SDE is:
where is the drift coefficient, is the diffusion coefficient, and represents a standard Wiener process. A corresponding reverse-time SDE exists, which generates data by starting from the prior and evolving backward to . This generative SDE is defined by:
where is a standard Wiener process running in reverse time and signifies an infinitesimal negative time step. This reverse process is critically dependent on the score function, , which is unknown. Therefore, it is approximated using a time-conditioned neural network, , such that . The network is trained by minimizing a denoising score matching objective:
In this objective, is a positive weighting function, and is the Gaussian transition kernel describing the perturbation process from to .
Distribution Matching Distillation of Diffusion Models
Distillation [33] has been widely explored not only for model compression but also for domain adaptation. For instance, Score Distillation Sampling (SDS) in DreamFusion [34] adapts pretrained 2D diffusion models to the 3D generation domain through a distillation process. More recently, Variational Score Distillation (VSD) [35] has demonstrated that a variational framework can provide a more principled and effective formulation for such adaptation tasks, enabling robust knowledge transfer between domains. Distillation has also been employed for accelerating diffusion inference [33, 36]. Diff-Instruct [37] distilled the multi-step diffusion process into a single-step network, dramatically improving inference speed. It introduced Integral KL (IKL) divergence as an effective and numerically stable metric for aligning diffusion model scores, outperforming the conventional KL divergence in cases where the model and target distributions have misaligned supports. By combining the variational framework with the IKL metric, DAVI [23] extended the concept of distillation to posterior estimation in inverse problems. DAVI learns a parameterized distribution that approximates the Bayesian posterior integrating a deep diffusion prior. This formulation jointly accounts for data fidelity and learned priors using the IKL divergence, leading to a stable and efficient inference framework that amortizes the cost of posterior sampling via parameterized variational distribution.
Methodology
We aim to recover a depth map from an SPAD measurement and a reflectivity guidance, which is severely ill-posed: the sensor provides only 64 spatial waveforms, and it has to recover the missing spatial structure. Our key observation is that a faithful depth map reconstruction must satisfy two requirements simultaneously: it must (i) be consistent with the corresponding histogram when passed through the same forward model, and (ii) supply a high-frequency spatial structure that is consistent with the high-resolution reflectivity guidance. In addition, we focus on improving computational efficiency to support practical deployment of low-cost SPAD-based sensing platforms (such as smartphones, or robotic sensors) in the future. We meet these requirements as follows. We address the extreme single-photon depth super-resolution problem with a two-stage framework. In Stage 1 we learn a histogram-conditioned diffusion prior over high-resolution depth maps (III-C) (see Figure 3). In Stage 2, we distill this multi-step diffusion prior into a single-step amortized posterior sampler (III-E) (see Figure 4); crucially, the amortized formulation allows us to incorporate a differentiable data-consistency operator derived from the single-photon forward model, which is what grounds the generator in the measured waveforms. The result is a single forward pass that reconstructs a depth map in approximately 83 ms, while providing much better accuracy than recent monocular depth estimation methods.

Figure 3. Overview of the conditional prior (Stage 1). The three conditioning signals: low-resolution ToF histogram H, high-resolution reflectivity R, and monocular depth prior enter the U-Net asymmetrically: the encoded histogram is injected on the encoder side, while the encoded reflectivity and monocular-depth features are fused into every decoder (synthesis) stage.

Figure 4. Overview of the distillation stage. A one-step generator is trained so that its samples (i) match the learned prior via teacher-student score matching and (ii) enforce estimated histogram to be consistent with the measured histogram .
Problem Formulation
Let denote the target depth map at high resolution (), and a co-registered HR reflectivity image obtained from a passive camera aligned with the SPAD array. A cost-effective SPAD measures waveforms (), each recording photon counts across time bins, giving a histogram tensor . Rather than relying only on the low-resolution histogram, the reconstruction is conditioned on all available observations.
where is a monocular-depth estimator output (see III-B).
Multimodal Conditioning for Extreme Super-Resolution
To complement the limited information available in the low-resolution photon histograms, the reconstruction of a depth map is conditioned on three complementary signals, each providing a different type of information:
ToF histogram : absolute, metric depth via photon-arrival statistics, but limited spatial layout.
Reflectivity map : dense high-resolution spatial structure (edges, textures), but no depth.
Monocular depth estimate : a relative depth field obtained by using (replicated to three channels) as an input to a frozen DepthAnything V2 [9] (ViT-B). This map provides dense geometric and semantic information at full resolution, including object boundaries and surface orientations.
The three signals are complementary. The histogram provides depth but limited spatial layout, and the reflectivity provides spatial detail but no depth. Since a 2D reflectivity image is inherently ambiguous about 3D geometry, we add the monocular depth estimate to inject the geometric structure that the reflectivity alone cannot resolve. DepthAnything provides only relative disparity (scale- and shift-ambiguous). We therefore use as unaligned geometric guidance, allowing the network to combine its spatial structure with the absolute depth information provided by the ToF histogram. We quantify the contribution of each conditioning signal and its robustness to degraded reflectivity in § 5.
Stage 1: Conditional Diffusion Prior
We first learn a conditional prior to generate a high-resolution depth maps while using the observations . We consider a classical diffusion-based framework where the forward process corrupts a clean depth map with Gaussian noise over steps under a cosine schedule [38], , and the theta-parametrized network conditioned on is trained using denoising score matching:
The resulting model is sampled using 250 DDPM steps [31]. Although it produces high-fidelity depth reconstructions, iterative sampling introduces substantial inference latency. In our framework, this model serves as the multi-step teacher whose conditional score given by is used for distribution matching during Stage 2. The following sections introduce the considered network architecture and the distillation process of this teacher into a one-step generator while incorporating explicit histogram data consistency.
Network Architecture
Both teacher and student use a conditional U-Net [39] built on the ADM backbone [40] operating on inputs, with two residual blocks with group normalization per resolution level carrying the sinusoidal timestep embedding. Starting from 128 base channels, the channel width is scaled by across the five resolution levels ( to ), giving 128 to 512 channels as the spatial resolution decreases. Multi-head self-attention is applied at the two coarsest levels, and .
To effectively leverage the complementary nature of our conditioning signals, we employ an asymmetric injection strategy. Following [12], the low-resolution ToF histogram (H) is injected into the encoder to guide the coarse spatial scales, while the high-resolution reflectivity (R) and DepthAnything output () are fed into the decoder stages to synthesize fine structural details. Our UNet consists of 140M parameters including both conditioning encoders.
Histogram injection (encoder side): Before injection, H is processed by a 3D convolutional encoder consisting of two residual blocks with kernels and channel dimensions . To raise the effective photon count per bin in the photon-sparse regime, we first sum every
consecutive time bins, reducing to 25 aggregated bins - trading marginal temporal resolution for higher signal-to-noise while preserving depth discrimination. The temporal dimension is then flattened into channel features and mapped via 2D residual blocks. These resulting features are added into the network at its two coarsest stages: the native resolution and a bicubically upsampled resolution.
Reflectivity and prior fusion (decoder side): Conversely, the HR reflectivity and its corresponding DA output are passed through parallel hierarchical 2D residual encoders to extract multi-scale feature pyramids. At every decoder resolution, the corresponding spatial features are projected via convolutions, concatenated, fused by a residual block, and added to the main decoder stream. Together with the standard U-Net skip connections, this provides high-resolution structural guidance throughout the decoding process.
Stage 2: One-Step Amortized Distillation
The diffusion teacher learned in Stage 1 captures an empirical conditional distribution of high-resolution depth maps given the measurements and visual guidance. The teacher represents the learned conditional distribution . Although the conditional teacher provides a strong reconstruction distribution, conditioning alone does not guarantee that a sample reproduces the observed photon-arrival histogram when passed through the physical forward model. We therefore augment the learned conditional distribution with an explicit sensor-domain consistency term. For a candidate depth map , assume measure the discrepancy between the observed histogram and the histogram rendered from using the differentiable forward model described in § 3-E1. We then define the measurement-regularized target distribution as
We approximate this target with a one-step implicit distribution . The corresponding variational objective is
We expand the KL divergence in (11) to obtain
Substituting (10) into (12) and omitting constant when optimizing the student leads to the objective
The terms in Eq. (13) encode the two requirements established from the outset: consistency with the physical observation model (, § 3-E1) and agreement of the student with the learned teacher conditional distribution (, § 3-E2). This variational formulation provides a natural way to incorporate explicit data consistency for 1-step distillation.
Enforcing data consistency on the estimated depth: Evaluating the data-fidelity term requires propagating the predicted high-resolution depth map back into the low-resolution histogram. Following Eq. (1), we construct the unscaled signal response , where is the predicted depth at pixel in time-bin units, and we set . We assume a normalized Gaussian , with bins to approximate the impulse response of the STMicroelectronics sensor. The operator pools responses over each low-resolution super-pixel footprint , giving . Consistent with (4), the estimated histogram is , where is the low-resolution background level estimated as the median count per bin of the measured histogram. The gain matches the total background-subtracted counts:
Under the Gaussian approximation, we use the mean squared error between the two histograms as the data-consistency loss:
Prior Regularisation via Integral KL: Minimising the KL-divergence term of (13) directly in data space is unstable when and have misaligned supports [23], [37]. We instead minimise the Integral KL (IKL), which compares the two across the diffusion trajectory,
where is a weighting coefficient defined as below,
where is a diffusion noise schedule [31]. The gradient of Eq. (16) is estimated by teacher-student score matching (see [23], [33] for more details regarding the derivations),
where is the frozen teacher score and is an auxiliary score network tracking the evolving one-step generator, trained by denoising score matching,
Training alternates between updating via Eq. (19) and updating the generator parameters through the combined objective , where is an auxiliary supervised reconstruction loss (see [23] for more details).
Perturbed Posterior Bridge: Following [23], we adopt a Perturbed Posterior Bridge (PPB) to make the single-step generator robust across noise levels. By providing an intermediate set of trajectories, the PPB enables more guided training of the generator, . Since and live in different spaces, a coarse per-super-pixel depth is first obtained from the histogram using matched filtering with bins. is then bilinearly upsampled to to obtain . During training the generator input interpolates between this and the clean depth , perturbed by Gaussian noise,
where , , and . We keep the perturbation unless otherwise stated, and is a monotonically increasing diffusion schedule.
Inference
At test time, the generator produces a reconstruction in a single forward pass,
The data-consistency operator is used only in training and adds no inference cost; the inference pipeline is (i) one frozen DepthAnything V2 pass for , (ii) the upsampled coarse depth-map , and (iii) one generator pass.
Experiments
We evaluate the proposed method on both simulated and real single-photon data. We first describe the experimental setup (Sec. IV-A), then report quantitative and qualitative results on simulated Middlebury scenes (Sec. IV-C), validate on real captures acquired with an STMicroelectronics SPAD sensor (Sec. IV-D), and analyze inference runtime (Sec. IV-E).
Experimental Setup
a) Training data and simulation: The teacher and student networks are trained on depth-reflectivity pairs drawn from the Hypersim [41] and MPI Sintel [42] datasets for all the simulated and real data evaluation. From each sequence we extract patches and synthesise the corresponding low-resolution ToF histograms following the observation model in Eq. (1). The histogram is computed on a spatially downsampled grid, yielding an spatial resolution with time bins per pixel. Photon arrivals are drawn as Poisson realizations, with a fixed signal-to-background ratio (SBR) at SBR and an impulse response of bins. For the Hypersim dataset (43140 frames), the photons-per-pixel (PPP) is kept fixed at PPP, whereas the MPI dataset (2128 frames) uses a variable PPP in the range [550, 800]. We apply geometric augmentation (horizontal flips and rotations) jointly to the depth, reflectivity, and histogram tensors. To account for the imperfect co-registration between the SPAD array and the guide camera, we additionally apply a random misalignment to the histogram during training. With probability 0.5, a random translation and scaling is applied to the grid about its center. The horizontal and vertical offsets are drawn independently from image pixels (equivalent to of a cell as one cell spans px on image) and the scale factor from .
b) Implementation details: Our U-Net backbone is trained using the -prediction objective in Eq. (9) with diffusion timesteps. The teacher is trained for 66k iterations using Adam with a constant learning rate of and a batch size of 48. The student is trained for 60k iterations with a batch size of 6. For the student, we use AdamW with a constant learning rate of and a weight decay of 0.01. Training is performed using three NVIDIA RTX 3090 GPUs for 14 hours. In all experiments, we set and . The DepthAnything V2 model (ViT-B backbone) remains frozen throughout both training stages. At inference, the teacher uses 250 DDPM sampling steps, whereas the proposed student produces a reconstruction in a single forward pass. All inference-time experiments are performed on a single NVIDIA RTX 3090 GPU with a batch size of 1.
Evaluation metrics and baseline methods
In simulated case with ground truth depth, we quantitatively evaluate performance using the root mean-squared error
| Method | Art | Baby | Books | Bowling | Cloth | Dolls | Lampshade | Laundry | Moebius | Monopoly | Plastic | Reindeer | Avg. |
| SBR = 2.0, PPP = 450 (Clean Reflectivity) | |||||||||||||
| DA (aligned) | 41.7/26.7 | 10.8/11.3 | 25.3/24.6 | 57.7/18.0 | 13.9/15.1 | 29.3/27.1 | 39.6/19.0 | 54.0/22.2 | 30.9/23.3 | 43.5/17.0 | 22.5/12.1 | 27.1/17.7 | 33.0/19.5 |
| DELTAR | 73.4/64.0 | 28.6/51.6 | 30.7/51.7 | 63.2/50.5 | 20.8/45.5 | 26.0/52.7 | 52.6/48.4 | 53.8/54.4 | 35.0/56.7 | 74.5/58.3 | 24.5/41.6 | 39.6/49.5 | 43.6/52.1 |
| Marigold-DC | 68.9/55.8 | 20.8/19.8 | 30.6/37.5 | 47.3/35.7 | 8.4/31.4 | 20.6/40.4 | 34.1/34.7 | 53.2/55.1 | 30.4/38.4 | 33.8/26.0 | 16.0/23.1 | 31.1/41.1 | 32.9/36.6 |
| DVSR | 74.5/51.4 | 24.9/29.8 | 39.9/39.9 | 52.7/36.4 | 12.5/22.5 | 24.9/44.7 | 63.5/39.0 | 46.3/41.3 | 39.7/38.3 | 46.0/39.2 | 20.7/28.9 | 42.3/34.0 | 40.7/37.1 |
| D2ToF (teacher) | 24.6/23.3 | 8.9/13.1 | 11.2/19.9 | 16.6/16.2 | 5.9/14.6 | 12.9/27.8 | 20.3/14.8 | 17.3/22.4 | 14.8/22.9 | 14.7/12.9 | 6.0/11.1 | 13.1/18.0 | 13.9/18.1 |
| D2ToF | 21.2/25.2 | 7.1/16.0 | 9.7/21.7 | 14.9/18.7 | 4.1/14.6 | 11.0/27.2 | 17.4/18.4 | 13.6/24.2 | 12.5/23.2 | 12.1/15.5 | 4.7/15.4 | 12.0/20.5 | 11.7/20.1 |
| SBR = 1.0, PPP = 300 (Noisy Reflectivity) | |||||||||||||
| DA (aligned) | 63.9/38.9 | 34.8/22.8 | 81.3/31.4 | 78.2/29.6 | 24.3/23.7 | 46.8/34.7 | 57.3/27.5 | 66.5/36.5 | 59.7/32.5 | 146.1/23.7 | 40.9/23.6 | 33.3/22.9 | 61.1/29.0 |
| DELTAR | 70.6/64.2 | 28.5/51.5 | 30.7/51.4 | 60.6/50.2 | 21.5/45.5 | 26.1/52.6 | 55.4/48.1 | 53.9/54.4 | 34.8/56.6 | 74.5/58.5 | 24.8/41.7 | 38.5/49.6 | 43.3/52.0 |
| Marigold-DC | 77.8/70.1 | 40.3/46.4 | 40.3/56.0 | 60.1/46.9 | 16.0/52.8 | 25.8/47.1 | 70.8/69.2 | 51.9/61.8 | 41.4/51.6 | 48.5/45.8 | 26.1/40.5 | 51.7/43.7 | 45.9/52.7 |
| DVSR | 73.6/51.3 | 37.1/45.2 | 38.9/39.1 | 55.5/36.2 | 14.1/24.0 | 27.3/46.3 | 56.6/41.7 | 41.0/43.4 | 37.0/39.6 | 84.4/42.9 | 23.4/31.0 | 40.2/36.2 | 44.1/39.7 |
| D2ToF (teacher) | 48.3/39.5 | 19.4/23.6 | 18.4/26.9 | 31.9/27.7 | 10.2/23.7 | 17.8/37.1 | 42.5/29.4 | 27.6/35.2 | 26.5/33.6 | 34.8/23.4 | 11.5/18.5 | 20.3/21.9 | 25.8/28.4 |
| D2ToF | 40.5/38.3 | 15.4/23.6 | 15.5/26.4 | 28.2/26.4 | 7.9/23.8 | 14.6/33.6 | 39.5/29.5 | 21.7/31.9 | 22.7/33.9 | 28.4/24.0 | 9.4/21.0 | 17.2/23.2 | 21.7/28.0 |
Table 1. Quantitative comparison on Middlebury scenes (8 × 8 → 128 × 128) at two noise levels with a spacing of 7.65 mm/bin. Each cell reports RMSE in mm / mean surface normal error in degrees (both ↓). The clean configuration uses SBR = 2.0, PPP = 450 with clean reflectivity. The noisy configuration uses SBR = 1.0, PPP = 300 with degraded reflectivity (Poisson , Gaussian ). DA: DepthAnything V2-base aligned to the measured histograms, DELTAR [8], Marigold-DC [11], and DVSR [10]. Within each scene, bold marks the best (lowest) value and underline the second best, applied to each metric independently. The shaded rightmost column reports the mean over all twelve scenes.
(RMSE) (in mm) and the mean surface normal error (SNE) (in degrees). For real data, ground truth depth is not available and we use histogram consistency to evaluate performance. We compare our method against a recent diffusion-based depth completion work Marigold-DC [11], a single-photon depth super-resolution methods, DVSR [10] and DELTAR [8]. Finally, we include our multi-step teacher to quantify the fidelity against the one-step student.
DepthAnything (DA). We apply DepthAnything V2 (ViT-B) [9] directly to the high-resolution reflectivity image. Since DA predicts relative disparity, we align it to our measurement histogram. Matched filtering ( bins) of the measured histogram gives a coarse metric depth , and the DA output is average-pooled onto the same grid to give . We regress inverse depth,
solve for by least squares over the super-pixels, and invert the fitted map. Coefficients are re-estimated for every scene and noise level.
DELTAR [8] predicts dense metric depth from a lightweight 8×8 zone ToF sensor paired with an RGB image, matching the input/output shapes of this benchmark. It does not consume the waveform, instead each zone is summarized by a Gaussian having pair.
Marigold-DC [11] performs monocular depth completion using an image together with sparse metric depth measurements. To adapt our 8 × 8 ToF measurements to this input format, we first estimate one metric depth value for each ToF zone using matched filtering. Each estimated zone depth is then placed at the centre of its corresponding 16 × 16 region on a 128 × 128 sparse-depth map. This produces 64 sparse depth anchors, while all unmeasured pixels are set to zero (RMSE) and the mean surface normal error (SNE) indicate missing depth. The resulting sparse-depth map and the high-resolution guidance image are provided to Marigold-DC. We use the released depth checkpoint with trailing timestep spacing and run inference for 50 DDIM steps.
DVSR [10] is a video-based super-resolution network that consumes a per-pixel peak depth rather than a waveform along with an RGB image. We report the dynamic regime it was designed for: the argmax of the measured histogram is bicubic-upsampled to guidance resolution, panned one pixel per frame across a ten-frame clip, finally metrics calculation uses the zero-offset reference frame.
Results on simulated data
a) Testing data: We simulate high-resolution SPAD histograms with a single depth return per pixel and evaluate on 12 scenes from the Middlebury stereo dataset [43] (Art, Bowling, Laundry, Reindeer, Moebius, Lampshade, Dolls, Baby, Books, Monopoly, Cloth, and Plastic), which provide aligned high-resolution reflectivity and depth maps. Missing depth values arising from stereo occlusions are filled using median filtering prior to histogram simulation. The HR histograms are used to build LR versions that have 100 time bins at an 8 × 8 spatial resolution, with 128 × 128 reflectivity maps. We consider two operating points: clean settings (SBR = 2.0, PPP = 450) with clean reflectivity, and a challenging dark condition (SBR = 1.0, PPP = 300) where the guidance reflectivity is degraded by Poisson and Gaussian noise (, ).
b) Quantitative results: Table I reports RMSE (in mm) and SNE (in degrees), averaged across 12 scenes under clean and noisy conditions. For each method, scene, and condition, we generate 50 reconstructions from the same fixed input measurement, compute the metrics of each against ground truth, and average the 50 values. Our proposed one-step model substantially outperforms all baselines in accuracy in both scenarios. Under clean conditions, it achieves an average RMSE of 11.7 mm, roughly one-third the error of the recent Marigold-DC (32.9 mm) and a significant improvement over the histogram-aligned DA baseline (33.0 mm). The one-step student model surpasses its multi-step teacher (13.9 mm) on every scene. SNE shows a slightly different trend from depth RMSE. Under clean conditions, the teacher achieves a lower angular error than the one-step student (18.1° versus 20.1°), despite their similar RMSE. Under noisy conditions, however, the student achieves the lowest value in both RMSE (21.7 mm) and SNE (28.0°). DA also performs well in terms of surface normals under clean conditions (19.5°), despite having a considerably higher RMSE (33.0 mm versus 11.7 mm for D2ToF). This difference arises because SNE measures local surface orientation rather than absolute depth accuracy. Furthermore, D2ToF retains the lowest mean RMSE under the tested degradation. While noise propagates visibly into the baseline estimates and severely degrades the intermediate DA predictions, our approach consistently recovers coherent depth structures. Crucially, we achieve this robustness zero-shot, requiring no retraining or fine-tuning on noisy reflectivity. We note finally that the analysis of the histogram consistency is grouped in § 5-B.
c) Qualitative results: Figure 5 presents qualitative comparisons in both clean and noisy settings. The baseline methods, DELTAR [8] and DVSR [10], fail to recover sharp depth discontinuities from the inputs, yielding over-smoothed reconstructions. Although Marigold-DC [11] performs relatively well in uniform scenes like Bowling, it generates subpar results in scenes containing intricate details, such as Art. The performance degrades further in the noisy setting. The reflectivity-only DA model (aligned to measurement histogram, explained in Sec. IV-B) captures plausible scene structures but fails severely in noisy scenarios. Conversely, our one-step framework successfully reconstructs the global geometry and fine structural details in both environments.

Figure 5. (Top) Comparison with different methods when both histograms (SBR=2.0, PPP=450.0) and reflectivity maps are in clean scenario. Our method achieves superior results compared to other reconstruction methods. (Bottom) Qualitative comparison of different methods under noisy conditions. The inputs are degraded by both reflectivity noise (Poisson =100, Gaussian =0.02) and ToF histogram noise (SBR = 1.0, PPP = 300). Our proposed approach significantly outperforms existing baselines under such noisy scenarios.
Results on real data
To evaluate our approach in practical scenarios, we utilize the VL53L8 time-of-flight (ToF) sensor manufactured by STMicroelectronics. This device actively illuminates the scene with 940-nm light pulses, employing single-photon avalanche diode (SPAD) detectors and dedicated timing electronics to capture the time of arrival of reflected photons, thereby generating histograms of photon counts.
For our real-world validation, the ToF histograms are captured with the sensor configured to an spatial resolution, yielding 144 time bins per pixel. We isolate the first 100 bins for processing, discarding the trailing background data. A Kinect sensor (positioned beside) provides the high-resolution reflectivity maps used for guidance, together with a reference high-resolution depth map for visual comparison. We collect scenes under both well-lit and low-light conditions to characterise behaviour across different photon regimes.

Figure 6. Comparison with different methods for real data collected using an inexpensive STMicroelectronics sensor in 8 × 8 mode. The reflectivity maps (128 × 128) are provided by a Kinect sensor connected adjacent to the SPAD.
Figure 6 shows reconstructions under well-lit conditions (first four rows) and low-light conditions (last two rows). Despite being trained purely on simulated data, the qualitative results demonstrate transfer to real sensor measurements and produces sharp, geometrically consistent depth maps, recovering object boundaries that are absent from the raw 8 × 8 waveforms. Reflectivity-based algorithms, such as DA and Marigold-DC, perform well when clean reflectivity data are available. However, their performance degrades under low-light conditions due to reduced photon counts, which make reconstruction more challenging. In contrast, the proposed algorithm demonstrates greater robustness under these conditions. Figure 2 displays the measured waveforms for nine pixels in the central region of the sensor, illustrating the photon-arrival statistics underlying the reconstruction.
Reconstruction time
A central advantage of the proposed framework is its inference speed. Although the method ingests the full ToF histogram, the amortized single-step formulation enables reconstruction in a single forward pass. As summarized in Figure 1, our one-step method reconstructs a scene in approximately 83 ms on average, of which roughly 39 ms is attributable to the frozen DA. For DVSR, Figure 1 reports the average inference time per frame over the 10-frame clip. The multi-step teacher requires 11 s per sample owing to its 250-step DDPM sampling process. The one-step student is therefore approximately 130× faster than the teacher while
being slightly better in average RMSE. Even without DA, our model performs relatively well at a faster inference speed of 41 ms (see Figure 1). This places our framework in a highly favorable region of the speed-quality trade-off, putting it well within reach of interactive frame rates on a single RTX 3090 GPU. Furthermore, the differentiable data-consistency term is utilized exclusively during training, thereby incurring no additional inference-time overhead.
Ablation Study
The ablations below are evaluated under 4 operating points that degrade the two input modalities (Table II). Quantitative results are averaged over the 12 Middlebury scenes and reported as RMSE (mm). For selected experiments, we also report the mean RMSE and standard deviation (±) computed from 50 stochastic reconstructions of the same input.
| Setting | Histogram SBR | Histogram PPP | Reflectivity | Reflectivity |
| Clean | 2.0 | 450 | ||
| Noise 1 | 2.0 | 450 | 100 | 0.02 |
| Noise 2 | 1.0 | 300 | 100 | 0.02 |
| Noise 3 | 1.0 | 300 | 40 | 0.05 |
Table II. Operating points used throughout the ablation study.
Effect of the visual conditionings
We ablate the two visual conditioning signals: the high-resolution reflectivity map R and the DA output to quantify their individual contributions and, critically, the robustness of D2ToF when the reflectivity itself is degraded. We compare the full model against two reduced variants: one without DA (H+R) and one without the reflectivity channel (H+) and, as a reference, the monocular prediction alone. Quantitative results are reported in Table III and visual results are shown in Figure 7.

Figure 7. Effect of reflectivity map R and DA output when either of them is used as a condition on the network decoder of D2ToF.
| Configuration | Clean | Noise 1 | Noise 2 |
| Monocular ( only) | 33.0 | 61.2 | 61.1 |
| with R only (H+R) | 18.19 ± 0.14 | 25.83 ± 0.08 | 26.69 ± 0.05 |
| with only (H+) | 12.59 ± 0.06 | 21.69 ± 0.07 | 23.11 ± 0.07 |
| Full (H+R+) | 11.71 ± 0.06 | 20.16 ± 0.05 | 21.75 ± 0.05 |
Table III. Ablation of the reflectivity (R) and DA-output () conditioning. The highlighted row is the full model.
Two findings emerge. First, the histogram adds robustness to noise. The aligned monocular baseline, which sees no range information, degrades sharply as the reflectivity is corrupted, since its RMSE rising from 33.0 to 61.2 mm, whereas every variant that includes the histogram keeps RMSE below 30.0 mm across all three conditions. Under noisy conditions, the monocular error is nearly 3 times that of the full model.
Second, the two visual conditionings are complementary and provide robustness. On all scenarios, the full H++R variant has the lowest RMSE. Under noise, the full model retains the lowest RMSE (20.16 and 21.75). The results confirm that both signals are needed for the best overall trade-off. Removing either or R produces more error under noisy conditions compared to the full variant.
Effect of
The data-consistency term is applied only during training. To verify that the reconstruction agrees with the measurement at test time, we take the depth map produced by each method, re-render it into a low-resolution histogram, and compare it with the measured histogram at clean case. Tables IV and V show the effect of the consistency loss in our framework. Including reduces mean depth RMSE across all four simulated conditions. When only the reflectivity guidance is degraded (Noise 1), RMSE decreases from 21.26 to 20.16 mm using , although training uses only clean reflectivity images. The loss also reduces the mean squared error () from 8.62 to 6.74 on the 12 simulated Middlebury scenes under the clean setting and from 19.70 to 16.12 on the four well-lit real captures.
| Config. | Clean | Noise 1 | Noise 2 | Noise 3 |
| w/o | 12.16 ± 0.13 | 21.26 ± 0.09 | 22.42 ± 0.08 | 25.98 ± 0.08 |
| with | 11.71 ± 0.06 | 20.16 ± 0.05 | 21.75 ± 0.05 | 25.01 ± 0.06 |
Table IV. Ablation of the consistency loss under noise levels defined in Table II.
| DELTAR | Marigold DC | DVSR | D2ToF (Teacher) | D2ToF (w/o Cons. Loss) | D2ToF | GT |
| 37.04 | 13.04 | 25.28 | 12.91 | 8.62 | 6.74 | 4.75 |
| 53.41 | 18.00 | 30.02 | 21.04 | 19.70 | 16.12 | – |
Table V. Histogram reconstruction error (; lower is better), averaged over 12 clean simulated Middlebury scenes and four real scenes (well-lit scenes from Figure 6). GT is unavailable for the real scenes.
Figure 8 shows this comparison for a representative low-resolution pixel from the Laundry and Bowling scenes. Both our teacher and our student place the reconstructed peak at the measured depth and reproduce its shape, whereas the baselines deviate from the measured waveform.

Figure 8. Histogram consistency for one low-resolution pixel from the Bowling (top) and Laundry (bottom) scenes. The gray bars show the measured ToF histogram, while the colored curves show the histograms re-rendered from the depth estimates produced by each method. We generate 50 samples for each method to generate the mean (solid line) and the standard deviation (shaded area). The proposed methods (D2ToF-Teacher and D2ToF) reproduce the measured temporal structure more closely than the compared methods.
Effect of DepthAnything Model Size
We also evaluate the smaller DA variant to examine the accuracy-runtime trade-off of the monocular guidance module. Replacing DA-base with DA-small reduces the reconstruction time from 83 ms to 69 ms, while increasing the average RMSE from 11.7 to 13.1 in the clean setting and from 20.2 to 22.2 under Noise 1. This provides a lower-latency alternative when a modest reduction in reconstruction accuracy is acceptable.
Uncertainty Estimation
Figure 9 shows the per-pixel standard deviation maps of the teacher and the student in the clean setting, computed over 50 samples drawn for the same input. The multi-step teacher exhibits a larger sample spread than the student with (in [0,1] space). For the student, increasing the PPB perturbation scale from to (all separately trained) correspondingly increases the variability of the reconstructed samples.

Figure 9. We show the uncertainty estimation (standard deviation) of our teacher and student model for the Art scene from Middlebury in a clean setting. D2ToF (T) denotes our teacher model.
Conclusion
In this paper, we presented D2ToF, a one-step framework for extreme single-photon LiDAR depth super-resolution from ToF histograms to depth maps. Our approach distils a multi-step conditional diffusion teacher into a single-step generator using an amortized variational distillation framework. A differentiable measurement-constrained data-consistency term further encourages the reconstructed depth to reproduce the measured low-resolution ToF histogram under the forward model. To address the severe ill-posedness of the reconstruction problem, we combine the ToF waveform with a high-resolution reflectivity image and an unaligned monocular relative-depth estimate, providing complementary metric and spatial guidance. Our experiments demonstrate that the proposed approach outperforms the evaluated baselines across both clean and degraded measurement conditions, while the ablation studies demonstrate the benefits of the multimodal conditioning and histogram data-consistency term. Qualitative experiments on real SPAD measurements further demonstrate transfer from simulated training data to real sensor captures under both well-lit and low-light conditions. The one-step model reconstructs a depth map in approximately 83 ms on an NVIDIA RTX 3090. Overall, D2ToF provides an efficient approach for high-resolution depth reconstruction from extremely low-resolution single-photon ToF measurements.
References
- [1]J. Rapp, J. Tachella, Y. Altmann, S. McLaughlin, and V. K. Goyal, “Advances in single-photon lidar for autonomous vehicles: Working principles, challenges, and recent advances,” IEEE Signal Processing Magazine, vol. 37, no. 4, pp. 62–71, 2020.DOI
- [2]A. M. Wallace, A. Halimi, and G. S. Buller, “Full waveform lidar for adverse weather conditions,” IEEE Transactions on Vehicular Technology, vol. 69, no. 7, pp. 7064–7077, 2020.DOI
- [3]C. Callender, Z. Shi, F. Heide, and M. Hullin, “Low-cost spad sensing for non-line-of-sight tracking, material classification and depth imaging,” ACM Transactions on Graphics, vol. 40, no. 4, Jul. 2021.
- [4]A. M. Pawlikowska, A. Halimi, R. A. Lamb, and G. S. Buller, “Single-photon three-dimensional imaging at up to 10 kilometers range,” Opt. Express, vol. 25, no. 10, pp. 11919–11931, May 2017.DOI
- [5]Z.-P. Li, J.-T. Ye, X. Huang, P.-Y. Jiang, Y. Cao, Y. Hong, C. Yu, J. Zhang, Q. Zhang, C.-Z. Peng, F. Xu, and J.-W. Pan, “Single-photon imaging over 200km,” Optica, vol. 8, no. 3, pp. 344–349, Mar 2021.
- [6]A. Ruget, L. Wilson, J. Leach, R. Tobin, A. McCarthy, G. S. Buller, S. McLaughlin, and A. Halimi, “Plug-and-play algorithm for 3D guided video super-resolution of single-photon lidar data,” Opt. Express, vol. 33, no. 5, pp. 11775–11792, Mar 2025.
- [7]A. Ruget, M. Tyler, G. M. Martín, S. Scholes, F. Zhu, I. Gyongy, B. Hearn, S. McLaughlin, A. Halimi, and J. Leach, “Pixels2pose: Super-resolution time-of-flight imaging for 3D pose estimation,” Science Advances, vol. 8, no. 48, p. eade0123, 2022.
- [8]Y. Li, X. Liu, W. Dong, H. Zhou, H. Bao, G. Zhang, Y. Zhang, and Z. Cui, “Deltar: Depth estimation from a light-weight tof sensor and rgb image,” in European conference on computer vision. Springer, 2022, pp. 619–636.arxiv.org/abs/2209.13362
- [9]L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,” Advances in Neural Information Processing Systems, vol. 37, pp. 21 875–21 911, 2024.arxiv.org/abs/2406.09414
- [10]Z. Sun, W. Ye, J. Xiong, G. Choe, J. Wang, S. Su, and R. Ranjan, “Consistent direct time-of-flight video depth super-resolution,” in Proceedings of the ieee/cvf conference on computer vision and pattern recognition, 2023, pp. 5075–5085.arxiv.org/abs/2211.08658
- [11]M. Viola, K. Qu, N. Metzger, B. Ke, A. Becker, K. Schindler, and A. Obukhov, “Marigold: Zero-shot monocular depth completion with guided diffusion,” in 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025, pp. 5359–5370.arxiv.org/abs/2412.13389
- [12]D. B. Lindell, M. O’Toole, and G. Wetzstein, “Single-Photon 3D Imaging with Deep Sensor Fusion,” ACM Trans. Graph. (SIGGRAPH), no. 4, 2018.DOI
- [13]A. Ruget, S. McLaughlin, R. K. Henderson, I. Gyongy, A. Halimi, and J. Leach, “Robust super-resolution depth imaging via a multi-feature fusion deep network,” Opt. Express, vol. 29, no. 8, pp. 9117–11937, Apr 2021.DOI
- [19]P. Deka, S. McLaughlin, and A. Halimi, “Conditional diffusion for single-photon lidar depth super-resolution,” in 2025 IEEE Statistical Signal Processing Workshop (SSP), 2025, pp. 71–75.
- [20]A. Halimi, A. Maccarone, R. Lamb, G. S. Buller, and S. McLaughlin, “Robust and Guided Bayesian Reconstruction of Single-Photon 3D Lidar Data: Application to Multispectral and Underwater Imaging,” IEEE Trans. on Comput. Imaging, vol. 7, pp. 961–974, 2021.arxiv.org/abs/2103.10122
- [21]J. Rapp and V. K. Goyal, “A few photons among many: Unmixing signal and noise for photon-efficient active imaging,” IEEE Trans. Comput. Imaging, vol. 3, no. 3, pp. 445–459, Sept. 2017.DOI
- [22]K. Choi, J. Koo, S. McLaughlin, and A. Halimi, “Graph attention-driven bayesian deep unrolling for dual-peak single-photon lidar imaging,” IEEE Transactions on Computational Imaging, 2025.arxiv.org/abs/2504.02480
- [23]S. Lee, D. Park, I. Kong, and H. J. Kim, “Diffusion prior-based amortized variational inference for noisy inverse problems,” in European Conference on Computer Vision. Springer, 2024, pp. 288–304.arxiv.org/abs/2407.16125
- [24]A. Halimi, R. Tobin, A. McCarthy, J. Bioucas-Dias, S. McLaughlin, and G. S. Buller, “Robust restoration of sparse multidimensional single-photon lidar images,” IEEE Trans. Comput. Imaging, vol. 6, pp. 138–152, 2020.DOI
- [25]F. J. ANSCOMBE, “The transformation of poisson, binomial and negative-binomial data,” Biometrika, vol. 35, no. 3-4, pp. 246–254, 12 1948. [Online]. Available: https://doi.org/10.1093/biomet/35.3-4.246
- [26]J. Li, Z. Shen, R. Yin, and X. Zhang, “A reweighted l² method for image restoration with poisson and mixed poisson-gaussian noise,” pp. 875–894, 2015.
- [27]A. Halimi and S. McLaughlin, “Plug-and-play for joint deblurring and guided super-resolution of single-photon 3d lidar data,” in 2023 31st European Signal Processing Conference (EUSIPCO), 2023, pp. 1718–1722.DOI
- [28]A. Foi, M. Trimeche, V. Katkovnik, and K. Egiazarian, “Practical poisson-gaussian noise modeling and fitting for single-image raw-data,” IEEE Transactions on Image Processing, vol. 17, no. 10, pp. 1737–1754, 2008.DOI
- [29]T. Brooks, B. Mildenhall, T. Xue, J. Chen, D. Sharlet, and J. T. Barron, “Unprocessing images for learned raw denoising,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2019, pp. 11 028–11 037.arxiv.org/abs/1811.11127
- [30]C.-X. Li, X. Qu, A. Gnanasambandam, O. A. Elgendy, J.-J. Ma, and S. H. Chan, “Photon-limited object detection using non-local feature matching and knowledge distillation,” 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pp. 3959–3970, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID: 244532182
- [31]J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020.arxiv.org/abs/2006.11239
- [32]Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” arXiv preprint arXiv:2011.13456, 2020.arxiv.org/abs/2011.13456
- [33]T. Yin, M. Gharbi, R. Zhang, E. Shechtman, F. Durand, W. T. Freeman, and T. Park, “One-step diffusion with distribution matching distillation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 6613–6623.arxiv.org/abs/2311.18828
- [34]B. Poole, A. Jain, T. Barron, and B. Mildenhall, “Dreamfusion: Text-to-3d using 2d diffusion,” arXiv preprint arXiv:2209.14988, 2022.arxiv.org/abs/2209.14988
- [35]Z. Wang, C. Lu, Y. Wang, F. Bao, C. Li, H. Su, and J. Zhu, “Prolific-dreamer: High-fidelity and diverse text-to-3d generation with variational score distillation,” Advances in neural information processing systems, vol. 36, pp. 8406–8441, 2023.
- [36]T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,” arXiv preprint arXiv:2202.00512, 2022.arxiv.org/abs/2202.00512
- [37]W. Luo, T. Hu, S. Zhang, J. Sun, Z. Li, and Z. Zhang, “Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models,” Advances in Neural Information Processing Systems, vol. 36, pp. 76525–76546, 2023.arxiv.org/abs/2305.18455
- [38]A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilistic models,” in International conference on machine learning. PmLR, 2021, pp. 8162–8171.arxiv.org/abs/2102.09672
- [39]O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer, 2015, pp. 234–241.arxiv.org/abs/1505.04597
- [40]P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021.
- [41]M. Roberts, J. Ramapuram, A. Ranjan, A. Kumar, M. A. Bautista, N. Paczan, R. Webb, and J. M. Susskind, “Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding,” in International Conference on Computer Vision (ICCV) 2021, 2021.arxiv.org/abs/2011.02523
- [42]D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black, “A naturalistic open source movie for optical flow evaluation,” in European Conf. on Computer Vision (ECCV), ser. Part IV, LNCS 7577, A. Fitzgibbon et al. (Eds.), Ed. Springer-Verlag, Oct. 2012, pp. 611–625.DOI
- [43]H. Hirschmuller and D. Scharstein, “Evaluation of cost functions for stereo matching,” in 2007 IEEE Conference on Computer Vision and Pattern Recognition, 2007, pp. 1–8.DOI