Robust Event-guided Deblurring (RED)

• The paper demonstrates that robust deblurring can be achieved by modeling incomplete and noisy event streams using a Robustness-Oriented Perturbation Strategy and modality-specific disentangled representation.
• It integrates event restoration, cross-modal OmniAttention, and interactive modules (MSEM and ESEM) to fuse sparse motion cues with image semantics for enhanced performance.
• Experiments show that RED achieves state-of-the-art PSNR and SSIM on multiple benchmarks, remaining resilient to event degradation and exposure mismatches in real-world conditions.

Robust Event-guided Deblurring (RED) designates a line of motion deblurring research that uses asynchronous event streams to compensate for the temporal ambiguity of exposure-averaged RGB images while explicitly addressing the fragility of event guidance under incomplete, noisy, misaligned, or distribution-shifted measurements. In current usage, the term also refers to the specific RED network, which frames robustness as the central design objective through a Robustness-Oriented Perturbation Strategy, a modality-specific disentangled representation, and cross-modal interactive refinement modules (Leng et al., 6 Sep 2025). The topic sits within a broader progression from early event-image fusion and physics-inspired formulations toward real-event self-supervision, exposure-aware selection, scale-aware generalization, and degraded-event restoration (Sun et al., 2021, Xu et al., 2021, Kim et al., 2021, Zhang et al., 2023, Shen et al., 2024).

1. Physical model and problem formulation

Event-guided deblurring begins from the mismatch between how frame cameras and event cameras observe motion. A blurry frame is an exposure average, whereas a Dynamic Vision Sensor triggers events when logarithmic intensity changes cross a threshold. The RED paper states the triggering rule as

$$p = \begin{cases} +1 & \text{if } \log \frac{I_t(x, y)}{I_{t-\Delta t}(x, y)} > \theta \ -1 & \text{if } \log \frac{I_t(x, y)}{I_{t-\Delta t}(x, y)} < -\theta \end{cases}$$

which makes the event stream a sparse encoding of motion-induced brightness changes rather than a dense image measurement (Leng et al., 6 Sep 2025).

Several formulations in the literature make this coupling explicit. EVDI models blur formation as

$B = \frac{1}{T} \int_{t \in \mathcal{T}} L(t)\, dt$

and expresses the latent image at reference time $f$ as

$L(f) = \frac{B}{E(f, \mathcal{T})},$

where the double-integral term $E(f,\mathcal{T})$ is computed from the event stream over the exposure interval (Zhang et al., 2022). The unknown-exposure formulation writes the blurred frame as

$$B^{\Delta t_X}(x, y) \simeq \frac{1}{N} \sum_{i=1}^{N} L_{\tau_i}(x, y),$$

and links intensity evolution to events through

$$I^{t_2}(x, y) \simeq I^{t_1}(x, y) \cdot \exp\left( \sum_{t_1}^{t_2} \beta \cdot E^{t}(x, y) \right),$$

making clear that only events occurring during the true exposure phase are physically correlated with the blur (Kim et al., 2021).

These formulations establish the central promise of event-guided deblurring: events provide temporally dense motion cues that a single blurred RGB frame does not contain. They also expose the core difficulty that motivates RED. Because event generation depends on thresholding, contrast, and sensor nonidealities, event streams are incomplete and noisy rather than direct ground-truth motion trajectories (Leng et al., 6 Sep 2025).

2. Why robustness became the central issue

Early event-guided deblurring methods already recognized that simple fusion of blurred frames and raw events is inadequate under real operating conditions. EFNet argued that previous approaches were hindered by ineffective event representations and weak event-image fusion, especially in extreme blur settings, and introduced Symmetric Cumulative Event Representation (SCER), Event-Image Cross-modal Attention (EICA), and Event Mask Gated Connection (EMGC) to improve robustness (Sun et al., 2021). Event-guided Multi-patch Network extended DMPHN by adding event guidance, a stacked pipeline, and self-supervision against random transformations and Gaussian noise, explicitly targeting limited robustness to spatial transformations and noise (Zhang et al., 2023).

The robustness problem widened as the literature moved from synthetic events toward real sensors. “Motion Deblurring with Real Events” identified a sim-to-real gap caused by data inconsistency between synthetic and real blur-event pairs and proposed self-supervision with blur-consistency and photometric-consistency losses, together with a piece-wise linear motion model to handle motion non-linearities (Xu et al., 2021). “Event-guided Deblurring of Unknown Exposure Time Videos” showed that many methods assume the exposure time equals the reciprocal of the frame rate, an assumption described as untrue in real situations where exposure time may be unknown and dynamically varies (Kim et al., 2021). “Generalizing Event-Based Motion Deblurring in Real-World Scenarios” further argued that fixed spatial-resolution assumptions and specific blurriness distributions limit practical usage, motivating scale-aware learning over different spatial and temporal scales (Zhang et al., 2023).

A second robustness axis concerns the event stream itself. RDNet is built on the premise that real-world events are generally degraded, and explicitly models threshold bias, limited bandwidth, and circuit noise before restoring events in a first stage and using those restored events for second-stage deblurring (Shen et al., 2024). The 2025 RED network sharpens this point: existing methods focus on cross-modal interaction while overlooking the inherent incompleteness of event streams arising from the sensitivity-noise trade-off introduced by DVS thresholding (Leng et al., 6 Sep 2025).

A common misconception is that the availability of events by itself resolves motion blur. The literature instead treats event guidance as conditionally useful: it is highly informative when the stream is temporally aligned and sufficiently complete, but brittle when event corruption, exposure mismatch, scale mismatch, or low-contrast sparsity are ignored (Kim et al., 2021, Shen et al., 2024, Leng et al., 6 Sep 2025).

3. The RED network with modality-specific disentangled representation

The method explicitly titled “RED: Robust Event-guided Motion Deblurring with Modality-Specific Disentangled Representation” formulates robustness as the primary objective rather than a side effect of better fusion (Leng et al., 6 Sep 2025). Its architecture has three principal components: the Robustness-Oriented Perturbation Strategy (RPS), disentangled OmniAttention, and two interactive modules, the Motion Saliency Enhancer Module (MSEM) and the Event Semantic Engraver Module (ESEM).

RPS exposes the network to incomplete events during training by random masking. The perturbed event tensor is

$\widetilde{E} = E \odot M,$

with two masking schemes. Temporal-invariant masking applies a fixed 2D binary mask across all time steps, whereas temporal-variant masking samples different masks for different time slices. The dropout ratio $\alpha$ is sampled from a uniform distribution between $\alpha_{\min}$ and $B = \frac{1}{T} \int_{t \in \mathcal{T}} L(t)\, dt$0 for each training iteration (Leng et al., 6 Sep 2025). This is not merely data augmentation in the generic sense. The paper presents it as a robustness-oriented exposure to incomplete patterns so that the network becomes resilient to unknown event disturbance conditions.

The disentangled OmniAttention module separates modality roles instead of applying a modality-agnostic self-attention block. Mid-level image and event features are processed through $B = \frac{1}{T} \int_{t \in \mathcal{T}} L(t)\, dt$1 and depth-wise $B = \frac{1}{T} \int_{t \in \mathcal{T}} L(t)\, dt$2 convolutions to form queries, keys, and values. RED then computes separate attention maps for semantic reasoning in the image branch and temporal or motion dependency in the event branch: $B = \frac{1}{T} \int_{t \in \mathcal{T}} L(t)\, dt$3 The corresponding updates are

$B = \frac{1}{T} \int_{t \in \mathcal{T}} L(t)\, dt$4

Cross-modality attention is then performed in both directions, allowing image semantics to guide event interpretation and event motion cues to guide image restoration (Leng et al., 6 Sep 2025).

The interactive modules implement this complementarity explicitly. MSEM enhances motion-sensitive areas in blurry image features by extracting a high-frequency motion activation map from perturbed events and injecting it into the image branch. ESEM performs the inverse operation: it engraves semantic context from the image branch into incomplete event features through dual-branch attention. Applied iteratively across network levels, these modules produce motion-enhanced image features and semantically enhanced event features rather than a single undifferentiated fusion tensor (Leng et al., 6 Sep 2025).

Empirically, the RED paper reports the highest PSNR and SSIM across all tested event corruption ratios on GoPro, with 37.27 dB at TIR $B = \frac{1}{T} \int_{t \in \mathcal{T}} L(t)\, dt$5, and 36.94/36.80 at TIR/TVR $B = \frac{1}{T} \int_{t \in \mathcal{T}} L(t)\, dt$6. It also reports 29.56 dB on HighREV and 27.01 dB on REVD, leading the compared methods on both real-world datasets (Leng et al., 6 Sep 2025). The paper further states that under moderate or severe event loss, many event-guided methods underperform even the image-only DSTN baseline, whereas RED remains above that level.

4. Robustness mechanisms in the surrounding literature

Robust event-guided deblurring has evolved through several distinct technical strategies. The table summarizes representative mechanisms already present in the literature.

Method	Robustness target	Representative mechanism
EFNet (Sun et al., 2021)	Noisy event-image fusion	SCER, EICA, EMGC
RED (Leng et al., 6 Sep 2025)	Incomplete or corrupted events	RPS, disentangled OmniAttention, MSEM, ESEM
RDNet (Shen et al., 2024)	Degraded real-world events	Event restoration before deblurring
Unknown-exposure deblurring (Kim et al., 2021)	Variable exposure and readout	ETES and feature fusion
SAN (Zhang et al., 2023)	Spatial and temporal scale mismatch	EGER, MSFF, two-stage self-supervision
E-MPN (Zhang et al., 2023)	Transformations and Gaussian noise	Self-supervised consistency under augmentation

EFNet is a foundational case of robustness through representation design. Its SCER encodes cumulative event polarity symmetrically around the frame center, motivated by the Event-based Double Integral model, while EICA performs channel-wise cross-modal attention using image features as queries and event features as keys and values (Sun et al., 2021). EMGC then routes information between stages according to an event-derived mask, preserving detail in sharp regions and focusing correction on blur-dominant regions. EFNet reports 35.46 PSNR and 0.972 SSIM on GoPro, and 38.12 PSNR and 0.975 SSIM on REBlur (Sun et al., 2021).

Unknown-exposure video deblurring attacks a different robustness failure mode: the event subset relevant to blur is not known a priori. Its Exposure Time-based Event Selection (ETES) module computes a cross-modal temporal activation map from frame-event correlations, then weights event features so that only the bins most correlated with the true exposure phase are emphasized (Kim et al., 2021). The scale-aware network SAN addresses the fact that events are often lower resolution than frames and that blur severity changes across time. It introduces Exposure-Guided Event Representation (EGER), Multi-Scale Feature Fusion (MSFF), and a two-stage self-supervised scheme based on the relativity of blurriness, using brightness consistency, structure consistency, temporal generalization, and spatial generalization losses (Zhang et al., 2023).

RDNet treats event degradation as a first-class restoration problem. It models threshold bias, limited bandwidth, and circuit noise to synthesize degraded and undegraded event pairs $B = \frac{1}{T} \int_{t \in \mathcal{T}} L(t)\, dt$7, trains a first-stage restoration network with

$B = \frac{1}{T} \int_{t \in \mathcal{T}} L(t)\, dt$8

and then feeds the restored events into a second-stage deblurring network (Shen et al., 2024). On GOPRO, it reports 37.33 PSNR and 0.980 SSIM; on REBlur, 36.31 PSNR and 0.974 SSIM (Shen et al., 2024). This suggests that robustness can be improved either by learning invariance to degraded events, as in RED, or by restoring the event modality before fusion, as in RDNet.

5. Datasets, benchmarks, and empirical behavior

The empirical study of RED is inseparable from the datasets used to expose real failure modes. REBlur was introduced as a real-world event-blur dataset with blurry and sharp images plus events, captured under controlled illumination and motion conditions, and used to evaluate severe real blur (Sun et al., 2021). RGBlur+E later provided high-resolution RGB blur images and synchronized real event data through a beamsplitter rig, with a high-speed subset for quantitative evaluation and a long-exposure subset for qualitative assessment (Vitoria et al., 2022). MS-RBD was proposed for cross-resolution event-guided deblurring, combining high-resolution RGB frames with low-resolution events in indoor and outdoor sequences (Zhang et al., 2023). EVRB was introduced for event-guided video deblurring with synchronized blurry video, sharp video, and event data (Kim et al., 2024). RELED extended the setting to joint low-light enhancement and deblurring with real low-light blur, real events, and normal-light sharp ground truth (Kim et al., 2024). DavisMCR was presented specifically to test deblurring under diverse levels of event degradation induced by changing environmental brightness and object contrast (Shen et al., 2024).

Across these benchmarks, several recurrent empirical patterns appear. First, event guidance improves deblurring most visibly under severe or non-uniform motion blur when frame-only temporal correspondence becomes unreliable. Dynamic Motion Awareness reports modulated deformable convolutions whose offsets and masks are estimated purely from event features, with improvements of up to 1.57 dB on synthetic data and 1.08 dB on real event data (Vitoria et al., 2022). CMTA extends this into the video domain with recurrent intra-frame enhancement and event-guided inter-frame temporal alignment, reporting 36.78 dB and 0.977 SSIM on GoPro for CMTA-7 and 31.38 dB and 0.927 SSIM on EVRB (Kim et al., 2024).

Second, real-time or near-real-time performance is possible when the architecture is localized and computationally constrained. Event-guided Multi-patch Network reports 30 ms per 1280$B = \frac{1}{T} \int_{t \in \mathcal{T}} L(t)\, dt$9720 image, describes itself as the first real-time deep motion deblurring model for 720p images at 30fps, and achieves 33.83 dB and 0.941 SSIM for E-StackMPN at 5 frames on GoPro (Zhang et al., 2023). This suggests that robustness and deployment efficiency need not be mutually exclusive, although the literature reaches that trade-off through very different design choices.

Third, self-supervision is a major route to robustness on real data. EVDI uses mutual constraints among blurry frames, latent images, and event streams through $f$0, $f$1, and $f$2, reporting 30.40 PSNR and 0.91 SSIM for deblurring on GoPro while emphasizing adaptation to real blurry videos and events (Zhang et al., 2022). Motion Deblurring with Real Events uses blur-consistency and photometric-consistency on real blurry images and real events to bridge the synthetic-to-real gap (Xu et al., 2021). These methods do not use the RED acronym, but they address the same robustness objective from the training-signal side rather than the feature-disentanglement side.

6. Scope, misconceptions, and research directions

A narrow reading of RED would confine it to the 2025 modality-specific disentangled network (Leng et al., 6 Sep 2025). A broader and technically more useful reading treats RED as the robustness-oriented branch of event-guided deblurring research: methods that recognize event incompleteness, corruption, exposure mismatch, scale mismatch, or real-world domain shift as primary obstacles rather than secondary nuisances. On that reading, EFNet, SAN, unknown-exposure deblurring, RDNet, and real-event self-supervised methods are all part of the conceptual formation that the 2025 RED network makes explicit (Sun et al., 2021, Zhang et al., 2023, Kim et al., 2021, Shen et al., 2024, Xu et al., 2021).

Another misconception is to equate robustness exclusively with denoising. The literature shows a more heterogeneous picture. Robustness may mean exposure-aware event selection, as in ETES; spatial-scale flexibility, as in SAN; restoration of degraded events, as in RDNet; invariance to geometric transformations and Gaussian noise, as in E-MPN; or explicit resilience to missing events, as in RPS (Kim et al., 2021, Zhang et al., 2023, Shen et al., 2024, Zhang et al., 2023, Leng et al., 6 Sep 2025). These are not interchangeable. They target different failure modes and often rely on different assumptions about sensor synchronization, event quality, and supervision.

Recent extensions suggest that the robustness agenda is broadening. Low-light event-guided enhancement and deblurring jointly exploit event temporal fidelity and frame semantics under severe noise and illumination loss (Kim et al., 2024). Neuromorphic imaging work on joint image deblurring and event denoising uses iterative coarse-to-fine mutual refinement, where event-regularized image priors improve deblurring and image gradients regulate event denoising (Zhang et al., 2023). A plausible implication is that future RED systems will continue to move away from one-shot fusion toward pipelines that explicitly repair, select, or re-parameterize auxiliary sensing signals before or during cross-modal restoration.

Within that trajectory, the specific RED network is notable because it crystallizes a field-wide conclusion: event guidance is most effective when the model treats the event stream not as a clean motion oracle, but as a sparse, incomplete, and modality-specific signal that must be robustly interpreted before it can guide deblurring (Leng et al., 6 Sep 2025).