FMA-Net++: Motion- and Exposure-Aware Joint Video Super-Resolution and Deblurring¶

Conference: ECCV 2026
arXiv: 2512.04390
Code: Not open-sourced
Area: Image Restoration
Keywords: Video Super-Resolution, Video Deblurring, Exposure-Awareness, Temporal Modeling, Hierarchical Refinement

TL;DR¶

FMA-Net++ proposes a non-recurrent sequence-level framework based on Hierarchical Bidirectional Aggregation (HRBA) blocks. By introducing an Exposure Time-aware Modulation (ETM) module to inject frame-level exposure information into features, and combining it with an exposure-aware Flow-Guided Dynamic Filtering (FGDF) to jointly model motion- and exposure-varying degradations, the method achieves SOTA accuracy and inference speed on the VSRDB task.

Background & Motivation¶

Joint video super-resolution and deblurring (VSRDB) aims to restore sharp, high-resolution videos from blurry, low-resolution videos. In practical filming, blurry low-resolution videos are ubiquitous: camera shake and object motion combined with limited exposure time cause motion blur, while low resolution leads to the loss of high-frequency details. Single-task methods operate in isolation—super-resolution cannot remove blur, and deblurring cannot recover high-frequency details—making joint modeling an inevitable choice. A deeper issue is that blur formation is determined by two coupled factors: the motion field determines the spatial pattern of the blur, while the exposure time controls the temporal range and intensity of the blur. Even more challenging, the auto-exposure mechanism of modern cameras dynamically adjusts exposure duration frame-by-frame, causing the blur severity to vary drastically across different frames within the same video—a complex spatio-temporally varying degradation that existing fixed-exposure assumption models fail to capture.

Existing methods also have limitations in temporal modeling. Sliding-window architectures (e.g., FMA-Net) process only local neighborhoods per frame with a limited temporal receptive field, failing to utilize distant information to reconstruct heavily blurred regions. Although recurrent architectures can propagate information across frames, they face sequential processing bottlenecks—preventing parallel acceleration and making them prone to gradient vanishing on long sequences. Recent Transformer-based methods (e.g., VRT) support parallel processing but suffer from massive computational and memory footprints. More fundamentally, almost all existing methods default to a fixed exposure time for all frames; once exposure changes, the estimation of degradation patterns completely fails. FMA-Net addresses motion-related degradation via flow-guided dynamic filtering but is limited by its sliding-window design and fixed-exposure assumption.

This paper breaks through on two dimensions simultaneously: for temporal modeling, it proposes HRBA blocks to replace sliding-window and recurrent structures, enabling sequence-level parallel processing and hierarchically expanding the temporal receptive field; for degradation modeling, it introduces an Exposure Time-aware Modulation (ETM) layer to explicitly inject frame-level exposure information into features, allowing the subsequent degradation estimation network to output motion- and exposure-aware degradation kernels. Core Idea: Decompose VSRDB into explicit degradation learning (NetD, predicting joint motion-exposure degradation kernels) and degradation-guided restoration (NetR, utilizing degradation priors for high-quality sharp super-resolution). HRBA provides efficient long-range temporal modeling, while ETM and ETE inject frame-level exposure conditions to form a non-recurrent, parallelizable pipeline.

Method¶

Overall Architecture¶

The complete framework of FMA-Net++ consists of two major sub-networks: the degradation learning network (NetD) and the restoration network (NetR), both built by stacking multiple HRBA blocks. The entire pipeline is based on a key observation (formulated in Eq. 1-3 of the paper): the degradation kernel \(\mathcal{K}_i\) of a blurry LR frame at spatial location \(\boldsymbol{p}\) is jointly determined by the frame's exposure time \(\Delta t_{e,i}\) and the continuous motion field \(\boldsymbol{M}\) within its neighborhood. Directly inverting this continuous physical process is infeasible; hence, FMA-Net++ approximates it with a learnable discrete framework: NetD predicts frame-wise degradation kernels \(\mathcal{K}_i^D\) and optical flow priors from the input blurry LR sequence, while NetR receives these degradation priors and works with Degradation-Aware (DA) attention in HRBA to restore sharp HR frames frame-by-frame.

Specifically, the input blurry LR sequence \(\boldsymbol{X}\) is simultaneously fed into a pre-trained Exposure Time-aware feature Extractor (ETE) and NetD. ETE is a lightweight ResNet-18-based network pre-trained via supervised contrastive learning and frozen during joint training, outputting frame-level exposure embeddings \(\boldsymbol{u}_i\). NetD consists of \(M\) stacked HRBA blocks, each containing an ETM layer—which modulates the output features of HRBA with \(\boldsymbol{u}_i\) via SFT, ensuring features at all scales incorporate exposure information. After hierarchical refinement through multiple HRBA blocks, NetD predicts two sets of crucial priors: image flow-occlusion mask pairs (describing motion between sharp HR frames) and exposure-aware degradation kernels \(\mathcal{K}_i^D\). NetR likewise consists of \(M\) HRBA blocks. Its initial features fuse the input LR frames and the context features from NetD, while the flow-mask pairs are inherited from NetD. In each HRBA block of NetR, the multi-attention module includes self-attention and degradation-aware cross-attention (DA attention), where queries come from the predicted degradation kernels \(\mathcal{K}_i^D\), prompting the restoration process to actively focus on degraded regions. Finally, NetR outputs high-frequency residuals \(\hat{\boldsymbol{Y}}_i^{\text{res}}\), which are added to the bilinearly upsampled LR frames to obtain the final sharp HR output.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input blurry LR sequence X"] --> B["ETE<br/>Exposure Feature Extractor<br/>(Frozen)"]
    B --> C["Frame-level exposure embedding u_i"]
    A --> D["NetD Degradation Learning Network"]
    C --> D
    D --> D1["HRBA × M<br/>Each block contains ETM<br/>(SFT-modulated features)"]
    D1 --> E["Predict Degradation Priors<br/>Flow-mask pairs + Degradation kernel K_i^D"]
    E --> F["FGDF: Perform exposure-aware<br/>dynamic filtering along motion trajectories"]
    F --> G["Reconstruction of blurry LR loss<br/>L_D Supervision"]
    A --> H["NetR Restoration Network"]
    E --> H
    G --> H
    C --> H
    H --> H1["HRBA × M<br/>Multi-Attention (Self-Attention<br/>+ DA Cross-Attention)"]
    H1 --> J["Upsampling + Residual Prediction<br/>Y_hat_res"]
    J --> K["Final Sharp HR Sequence<br/>Y_hat = Y_hat_res + X↑"]

Key Designs¶

1. HRBA: Hierarchical Bidirectional Aggregation and Collaborative Refinement of Multi-Flow-Mask Pairs

Existing temporal modeling faces a dilemma: sliding-windows are constrained to fixed neighborhoods, while recurrent structures cannot run in parallel and suffer from gradient vanishing. The core concept of the HRBA block is to decompose temporal modeling into "hierarchical refinement"—after passing through each HRBA block, the features of the current frame aggregate information from increasingly distant past and future frames, thereby expanding the temporal receptive field block-by-block while processing all frames in parallel at the sequence level without sequentially constrained dependencies.

Specifically, the HRBA block maintains two core states: the feature map \(\boldsymbol{F}_i^j\) of the current frame \(i\) and a set of multi-flow-mask pairs \(\mathbf{f}_i^j\) (consisting of \(n\) pairs of optical flows \(\boldsymbol{f}\) and occlusion masks \(\boldsymbol{o}\) pointing to neighboring frames \(i\pm1\)). Multiple flow-mask pairs (\(n>1\)) are required because single optical flow estimation is highly unreliable under severe blur; multiple flows provide a "one-to-many" set of candidate correspondences, allowing the model to choose the most rational alignment among multiple motion hypotheses. In each refinement step, the current flow-mask pairs are used to warp and aggregate neighboring frame features via occlusion-aware warping to obtain \(\tilde{\boldsymbol{F}}_i^j\), and then the residual is predicted based on \(\tilde{\boldsymbol{F}}_i^j\) and current \(\mathbf{f}_i^j\) to update flow-mask pairs (\(\mathbf{f}_i^{j+1} = \mathbf{f}_i^j + \Delta\mathbf{f}_i^j\)). \(\tilde{\boldsymbol{F}}_i^j\) is passed through the multi-attention module (self-attention capturing spatial dependencies, plus degradation-aware cross-attention in NetR) and the ETM layer (injecting exposure information) to output the refined features \(\boldsymbol{F}_i^{j+1}\). As the number of HRBA blocks \(M\) increases (the paper uses \(M=4\)), features become cleaner and structures better aligned (visualized in Fig. S3 of the paper), and the temporal receptive field expands. Compared to sliding windows limited to a 3-frame neighborhood (like FMA-Net), the 4-layer HRBA can aggregate information from more distant frames to reconstruct heavily blurred areas.

2. ETE + ETM: Exposure Time-Aware Feature Extraction and Modulation Pipeline

Inter-frame exposure variation is a crucial factor long neglected in VSRDB. Within the same video, frames with shorter exposure exhibit milder blur and retain relatively more detail, whereas frames with longer exposure suffer from severe blur and spatial information loss. If the network is blind to the exposure conditions of each frame, it cannot distinguish whether "this frame is blurry because of a long exposure" or "this frame originally lacked details," leading to chaotic degradation estimation.

ETE is a ResNet-18 pre-trained via supervised contrastive learning across 5 discrete exposure levels (using duties of 5:1 to 5:5 synthesized in REDS-ME as pseudo-labels). After training, ETE is frozen to prevent the exposure embedding space from drifting during joint training. t-SNE visualization (Fig. 7a in the paper) shows that the 5 categories of exposure embeddings output by ETE form clear clusters in the feature space—with the shortest exposure (5:1) achieving 100% separation, while the longest exposures (5:4 and 5:5) exhibit slight overlap, which naturally reflects the visual ambiguity near the extreme blur boundaries.

The ETM layer is a lightweight SFT (Spatial Feature Transform) module embedded inside each HRBA block. It takes the frame-level exposure embedding \(\boldsymbol{u}_i\) and predicts affine parameters \((\boldsymbol{\alpha}, \boldsymbol{\beta})\) via a shallow network \(\mathcal{M}\), and then modulates the attention output \(\hat{\boldsymbol{F}}_i^j\) channel-wise: \(\boldsymbol{F}_i^{j+1} = (1 + \boldsymbol{\alpha}) \odot \hat{\boldsymbol{F}}_i^j + \boldsymbol{\beta}\). Consequently, all refinement steps in every HRBA block are explicitly aware of the current frame's exposure conditions—enabling degradation learning to predict degradation kernels at the correct temporal scale. Control experiments are highly illustrative (Table 6 in the paper): when the frame-wise exposure guidance \(\boldsymbol{u}_i\) is intentionally shuffled or replaced with fixed assumptions on REDS-RE, PSNR and tOF consistently deteriorate, particularly on transition frames where exposure changes; random incorrect exposure performs even worse than having no guidance at all—proving that \(\boldsymbol{u}_i\) is actively utilized rather than ignored.

3. Exposure-aware FGDF: Performing Exposure-Aware Dynamic Filtering along Motion Trajectories and Architecture-Level Decoupling

Physically, a blurry LR frame \(\boldsymbol{X}_i\) is the result of downsampling continuous clean signals integrated along motion trajectories during the exposure duration. Classic convolution modeling assumes spatially invariant kernels, which fails to express "different regions having distinct blur patterns due to varying motion": a fast-moving object and a static background will have dramatically different blur under the exact same exposure time. FGDF addresses this by performing filtering operations along optical flow trajectories—rather than sampling on a fixed \(k_d \times k_d\) grid, it samples pixel values along corresponding motion trajectory points computed from the estimated continuous inter-frame motion offsets. This essentially learns an implicit approximation of temporal integration.

The key upgrade in FMA-Net++ lies in exposure coupling: the degradation kernel \(\mathcal{K}_i^D\) used by FGDF is predicted by NetD from features already modulated by ETM, thus naturally becoming an exposure-aware FGDF. \(\mathcal{K}_i^D\) has a shape of \(\mathbb{R}^{3 \times H \times W \times k_d^2}\), predicting a set of dynamic filtering weights for each pixel position across three consecutive frames (\(t = i-1, i, i+1\)). The supervision for NetD stems from a self-reconstruction constraint (Eq. 5-6 in the paper): applying the predicted \(\mathcal{K}_i^D\) to ground-truth sharp HR frames via flow-guided dynamic filtering should reconstruct the original blurry LR frames \(\boldsymbol{X}_i\). This ingenious self-supervised design—where the "correctness" of the degradation kernel is validated by its ability to synthesize the blurry input—eliminates the need for ground-truth degradation kernel annotations.

At the architectural level, decoupling into NetD + NetR offers two advantages: first, NetR can focus solely on restoration, resulting in a more compact parameter footprint; second, the DA attention in NetR uses \(\mathcal{K}_i^D\) as queries for cross-attention, allowing the restoration process to selectively focus on spatio-temporal regions correlated with degradation patterns. Ablation studies (Table 8 in the paper) show that replacing DA attention with ordinary SFT significantly drops performance, confirming that degradation-guided cross-attention is indispensable for refined restoration.

A Complete Example: Walking Through the REDS-RE Dynamic Exposure Scenario¶

Consider a 10-frame test sequence in REDS-RE where frame-level exposure labels are generated via a random walk: frames 1-5 are 5:1 (shortest exposure, very mild blur), frames 6-7 jump to 5:3 (moderate blur), and frames 8-10 jump to 5:5 (longest exposure, severe blur). Such sequences are common in reality—when a camera automatically extends exposure in dark conditions, inter-frame blur suddenly worsens.

Step 1: ETE extracts exposure embeddings for each frame: frames 1-5 output embeddings clustered in the short-exposure region, frames 6-7 fall in the middle, and frames 8-10 cluster in the long-exposure region (features of non-adjacent frames are far apart in the t-SNE space). NetD receives these embeddings along with the input LR sequence. In the first HRBA block, local neighborhood alignment is performed: the HRBA block for frame 7 (5:3) receives the \((\boldsymbol{\alpha}, \boldsymbol{\beta})\) corresponding to the 5:3 exposure embedding, and SFT modulation scales its features to match the blur level of 5:3; meanwhile, the multi-flow-mask pairs \(n=9\) provide 9 motion candidates. On the severely blurred frames 8-10 (5:5), different candidates in the multi-flow set capture various possible correspondences of rapid motion. After passing through 4 HRBA layers to progressively expand the temporal receptive field, the features of frame 7 successfully aggregate information from frames 1-5 (clear references) and frames 8-10 (blurry references).

NetD then predicts the exposure-aware degradation kernel \(\mathcal{K}_i^D\) for each frame: on frames 1-5, the kernels tend to be localized (mild blur, kernels concentrated in the center), whereas on frames 8-10, the kernels are visibly diffused, and their shapes are guided by the motion direction estimated by the optical flow (larger weights along the motion vector). NetR uses these kernels for DA cross-attention: the queries for frame 8 come from its diffuse degradation kernels, performing cross-attention with the keys/values from self-attention, forcing the restoration process to focus on the severely blurred regions of that frame. Meanwhile, ETM continues to inject the 5:5 exposure condition. After refinement in the final HRBA layer, upsampling and residual additions are performed, enabling the transition region between frames 7-8 (where the degradation abruptly changes due to exposure transition) to still recover sharp edges and textures.

Loss & Training¶

FMA-Net++ adopts a three-stage training strategy. In the first stage, ETE is pre-trained with a supervised contrastive loss on synthetic data with 5 exposure levels and then frozen. In the second stage, NetD is trained independently with the loss \(\mathcal{L}_D = \sum_i \ell_1(\hat{\boldsymbol{X}}_i, \boldsymbol{X}_i) + \lambda_1 \sum_i \ell_1(\boldsymbol{Y}_{i\pm1 \to i}, \boldsymbol{Y}_i) + \lambda_2 \ell_1(\boldsymbol{f}^{\boldsymbol{Y}}, \boldsymbol{f}^{\boldsymbol{Y}}_{\text{RAFT}})\), where the first term is the blurry LR reconstruction loss (using the predicted kernels on ground-truth HR to reconstruct blurry LR), the second term is the warp loss (aligning optical flow warped HR frames with original frames), and the third term is the RAFT pseudo-supervised flow loss (RAFT is only used during training and not during inference). In the third stage, NetD and NetR are jointly fine-tuned with the total loss \(\mathcal{L}_{\text{total}} = \ell_1(\hat{\boldsymbol{Y}}, \boldsymbol{Y}) + \lambda_3 \mathcal{L}_D\). Hypereparameters are set to \(\lambda_1 = \lambda_2 = 10^{-4}, \lambda_3 = 0.1\). The input sequence length is 10 frames, patch size is \(64 \times 64\), upsampling factor is \(s=4\), the number of HRBA blocks is \(M=4\), the number of multi-flow-mask pairs is \(n=9\), and the degradation kernel size is \(k_d=20\).

Key Experimental Results¶

Main Results¶

Dataset	Metric	Ours	Prev. SOTA	Gain
REDS4-ME-5:4	PSNR / SSIM / tOF	29.66 / 0.8546 / 1.688	29.04 / 0.8275 / 1.891 (FMA-Net*)	+0.62dB / +0.027 / -0.203
REDS4-ME-5:5	PSNR / SSIM / tOF	29.24 / 0.8453 / 1.956	28.51 / 0.8136 / 2.269 (FMA-Net*)	+0.73dB / +0.032 / -0.313
REDS-RE (Dynamic Exposure)	PSNR / SSIM / tOF	30.13 / 0.8643 / 1.360	29.33 / 0.8427 / 1.602 (BSSTNet*)	+0.80dB / +0.022 / -0.242
GoPro (Cross-domain Generalization)	PSNR / SSIM / tOF	30.49 / 0.9018 / 2.091	28.83 / 0.8655 / 2.727 (FMA-Net*)	+1.66dB / +0.036 / -0.636

Note: * denotes baselines retrained on the REDS-ME training set. The advantage of FMA-Net++ on REDS-RE (dynamic exposure) is more pronounced than on REDS-ME (fixed exposure), showing that ETM exposure-aware modeling contributes most in exposure-changing scenarios. The substantial improvement of +1.66dB on GoPro validates strong generalization capability.

Ablation Study¶

Configuration	Key Metric (REDS4-ME-5:5)	Description
Full FMA-Net++	29.24 PSNR / 1.956 tOF	M=4 layer HRBA
Sliding-window variant	28.57 PSNR / 2.231 tOF	Fixed 3-frame neighborhood, constrained receptive field
Recurrent propagation variant	29.11 PSNR / 1.989 tOF	Sequential propagation, non-parallelizable, weak gradient propagation
w/o ETE (No exposure embedding)	29.12 PSNR / 2.054 tOF	After compensating with larger capacity (13.1M vs 12.8M), in-distribution PSNR is close, but REDS-RE PSNR is still 0.25dB worse and GoPro is 0.64dB worse
Self-attention + SFT (No DA attention)	28.86 PSNR / 2.132 tOF	Significantly drops after replacing DA cross-attention with SFT
Contrastive learning ETE vs Frame difference features	REDS-RE: 30.13 vs 29.75 PSNR	Contrastive learning ETE significantly outperforms alternatives like simple frame difference or classifier ETE

Key Findings¶

HRBA's core advantage stems from hierarchical long-range temporal aggregation: The sliding-window variant is the most constrained (PSNR lower by 0.67dB), and the recurrent variant is better but still insufficient (lower by 0.13dB). The block-wise progressive receptive field expansion of 4-layer HRBA avoids both the neighborhood limits of windows and the gradient vanishing of recurrent models.
ETM's contribution is most significant in dynamic exposure scenarios: On REDS-RE (inter-frame exposure changes), ETE exposure-aware modeling contributes a 0.25dB PSNR gain (compared to the capacity-matched baseline without ETE), whereas its impact is limited on the fixed-exposure REDS-ME—perfectly validating that the design is highly targeted rather than just blindly adding capacity.
Exposure guidance \(\boldsymbol{u}_i\) is actively utilized rather than ignored: Shuffling or replacing \(\boldsymbol{u}_i\) (Table 6) consistently degrades performance, especially on exposure-transition frames (PSNR drops from 29.98 to 29.05); random incorrect exposure is worse than using no guidance, proving that the network "believes" this guide.
Strong cross-domain generalization: Trained only on REDS synthetic data, the model achieves 30.49 PSNR without seeing any GoPro data, far exceeding the runner-up FMA-Net* (28.83) which was fully trained. This shows that the decoupled degradation learning + exposure-aware degradation estimation captures a more fundamental degradation mechanism.
Outstanding efficiency: With only 12.8M parameters, it infers in just 0.074 seconds per frame (at 180×320 resolution), which is 5.2x faster than the similarly-sized SOTA method RVRT, 9.2x faster than VRT, and consumes only 6.2GB VRAM (VRT requires 20.5GB).

Highlights & Insights¶

Architectural insight of decoupling degradation learning and restoration: Instead of directly learning the end-to-end LR\(\to\)HR mapping, the method first trains a degradation synthesizer (NetD predicting degradation kernels) and then allows the restoration network to operate under the guidance of these kernels. This "teaching the network what degradation is before asking it to repair it" paradigm is more robust than blind restoration. Moreover, NetD's task can be supervised using existing physical models (verifying kernel correctness by synthesizing blurry LR from ground-truth HR), resulting in higher data efficiency.
Exposure embedding via contrastive learning: ETE does not directly regress continuous exposure durations; instead, it uses contrastive learning in the feature space to separate frames with different exposures. This naturally preserves the ordinal relationship among exposure states (the feature space from 5:1 to 5:5 is continuous), which is clearly illustrated by the t-SNE visualization—and is key to the model's generalization under unseen continuous exposure conditions (as new exposures can be interpolated between anchors).
Practical design of multi-flow-mask pairs: Under severe blur, single optical flow estimation is highly unreliable. Maintaining 9 motion hypotheses simultaneously—allowing the model to select the most rational correspondence among candidates to warp features—proves far more stable than forcing a single flow estimation. The computational overhead of adding extra flows is negligible (adding \(n=1\) to \(n=9\) takes only 0.001s additional time at 12M parameters), yet the performance gain is distinct (28.52 \(\to\) 29.24 PSNR).
Lightweight exposure injection via SFT modulation: The ETM layer is just a lightweight SFT module (learning affine parameters to scale and shift output features). Embedding it in each HRBA block adds almost no parameters or inference time, yet infuses exposure-aware capability throughout every layer's features. This "gentle intervention, deep penetration" design philosophy is highly translatable to other conditional tasks.

Limitations & Future Work¶

Limitation to exposure duration variance without modeling a complete degradation space: The authors explicitly note that REDS-ME/RE only simulates blur variations caused by temporal integration, ignoring other co-occurring degradation factors in real cameras such as ISO gain changes, white balance shifts, dynamic range compression, and sensor noise. Expanding to jointly model multiple degradations is an important future direction.
Dependence on discrete exposure anchors: The contrastive learning pre-training of ETE requires discrete exposure labels (5 levels). Although the exposure embedding space can interpolate intermediate values, the assumption that "exposure is known per frame (obtainable in synthetic setups)" does not hold in real-world scenarios—inference extracts exposure embeddings directly from input frames without ground-truth labels, but training still requires synthetic pseudo-labels. If real-world camera auto-exposure strategies deviate from the synthetic assumptions, generalization may suffer.
Synthesis-to-real generalization gap: Despite superior performance on GoPro and real smartphone videos, the evaluation datasets (REDS) still contain blur synthesized based on temporal integration. Real camera blur may include complex factors not modeled here, such as rolling shutter effects and non-linear sensor responses.
Fixed sequence length: Training uses a 10-frame input. Longer videos may require window-based processing, which introduces window boundary effects. Theoretically, HRBA can handle arbitrary lengths, but it is constrained by GPU memory.

vs FMA-Net (Prior work from this team, CVPR 2024): FMA-Net first proposed FGDF to handle motion-dependent degradation, but was restricted by a sliding window (T=3) and a fixed-exposure assumption. FMA-Net++ replaces sliding windows with HRBA (sequence-level parallelization, hierarchical receptive fields) and adds ETM exposure-aware modeling—yielding a 0.73dB PSNR improvement on 5:5 severe blur while being 4.3x faster (0.318s \(\to\) 0.074s).
vs BasicVSR++ (Recurrent VSR): BasicVSR++ is a representative of recurrent propagation, which partially alleviates information scarcity through temporal propagation, but sequential processing restricts training parallelization and suffers from gradient decay in long sequences. The recurrent variant of HRBA in FMA-Net++ outperforms BasicVSR++ (due to the use of ETM + DA attention), but the full HRBA version significantly outperforms the recurrent variant—indicating that structural-level parallel hierarchical modeling is superior to sequential propagation.
vs VRT / RVRT (Transformer Video Restoration): While VRT also employs sequence-level parallelization (shifted window), its Transformer architecture consumes massive computations and memory on long sequences (20.5GB vs 6.2GB). RVRT reduces complexity with a recurrent design (12.9M parameters), but the inherent recurrent bottleneck remains. FMA-Net++'s HRBA holds a distinct efficiency advantage (0.074s vs 0.385s on RVRT). This suggests that CNN-style hierarchical refinement combined with lightweight attention may be a more suitable architecture choice than full-stack Transformers for video restoration.
vs Ev-DeblurVSR (Event-assisted VSRDB): Ev-DeblurVSR utilizes event cameras to assist in deblurring, obtaining good temporal information without requiring extra data for the VSRDB task. However, it requires non-standard event data (unavailable in standard RGB videos) and does not model exposure changes. FMA-Net++ significantly outperforms it on all metrics using only standard RGB inputs, demonstrating higher practicality and generalization.

Rating¶

Novelty: ⭐⭐⭐⭐ The physical motivation of introducing exposure awareness into VSRDB is solid. Replacing the two dominant structures (sliding-window and recurrent) with HRBA is supported by clear theoretical analysis and experimental validation. The combination of exposure-aware FGDF + ETE contrastive pre-training is highly original.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Establishes 2 benchmarks, conducts horizontal comparison against 10+ methods, and performs 5 sets of ablation studies (temporal strategy/ETE/DA attention/multi-flow number/loss functions). Detailed t-SNE visualization and guide substitution analyses cover all design aspects. The REDS-RE dynamic exposure design is an especially great addition, directly exposing the weakness of the fixed-exposure assumption.
Writing Quality: ⭐⭐⭐⭐ The mathematical derivation of the physical model (Eq. 1-3) aligns clearly with the network design, and the method section flows logically. However, the text is somewhat lengthy, and moving some redundant details to the appendix would improve the reading tempo.
Value: ⭐⭐⭐⭐⭐ Exposure-aware joint VSRDB addresses a real-world pain point in practical applications (smartphones/surveillance cameras). FMA-Net++ is fast, effective, generalizes well, and holds high practical value. As a general efficient temporal modeling module, HRBA also possesses migration potential. The fact that the code is not open-sourced slightly reduces reproducibility.