arXiv CSopen access

DART: A Degradation-Aware Recurrent Transformer for Archival Film Restoration

DART: A Degradation-Aware Recurrent Transformer for Archival Film Restoration Mikołaj Jastrzębski, Wojciech Kozłowski, and Kamil Adamczewski Wrocław University of Science and Technology 0.32

DART (Ours)

arXiv:2607.21219v1 [cs.CV] 23 Jul 2026

MANIQA

0.30 0.28

MambaOFR (CVPR25)

BasicVSR (CVPR21) DeepRemaster (SGA19)

0.26

#Params

0.24 BasicVSR++ (CVPR22)

0.34

(a) Top: original video frames. Bottom: the restored frames.

ShiftNet (CVPR23)

RTN (CVPR22)

0.36

0.38

1M 5M

0.40

10M

0.42

CLIP-IQA+ (b) Performance gain

0.44

15M

0.46

Fig. 1: (a) Real-world old films restored by our method (b) DART beats prior stateof-the-art while remaining efficient.

Abstract. Archival film restoration is a challenging problem because historical footage contains compound degradations such as scratches, dust, blur, noise, flicker, and photometric aging, while clean reference videos are unavailable. Existing video restoration methods largely treat these degradations implicitly, reconstructing frames without explicit knowledge of where damage occurs or how severe it is. We propose DART, a degradation-aware recurrent transformer for archival film restoration. DART predicts and propagates a soft defect mask through time, using it to guide temporal fusion and condition the restoration network on both damage location and severity. This makes the restoration process explicitly aware of film artifacts rather than relying only on reconstruction losses. Experiments on real archival benchmarks show that DART improves no-reference perceptual quality over prior restoration architectures while remaining compact and efficient, producing cleaner and more temporally consistent restorations of structured film damage. Keywords: Old film restoration · Video restoration · Degradation-aware · Recurrent transformer · Mask supervision

1

Introduction

Archival film is an irreplaceable visual record of cultural, historical, and artistic memory. However, much of this footage has survived only in degraded form.

2

Jastrzębski et al.

Old films commonly contain a mixture of structural defects, such as scratches, dust, cracks, and blotches, together with photometric degradation, blur, flicker, exposure instability, and artifacts introduced during digitization. Restoring such material is therefore not a standard denoising or enhancement problem. A model must distinguish genuine scene content from damage, preserve temporal coherence, and operate in a setting where clean reference videos are unavailable for real archival footage. Recent video restoration methods have made substantial progress by exploiting temporal information across frames through alignment, recurrent propagation, and transformer-based reconstruction [1, 2, 13, 15]. These ideas have also been adapted to old-film restoration, where models must handle stronger and more diverse artifacts than in standard video enhancement [8, 16, 19, 27]. However, most existing restorers still treat degradation only implicitly, reconstructing clean frames without being told where damage occurs or how severe it is, so defect localization is learned only indirectly through the reconstruction loss rather than as a supervised signal. Recent supervised detection approaches address part of this issue, but are limited to structural damage and are usually separated from the restoration model itself [18]. This limits the model’s ability to distinguish true film damage from valid scene structures and motivates a restoration framework that is explicitly degradation-aware. This lack of explicit damage awareness is especially problematic for archival footage. A scratch, a fence post, a branch, and a thin object boundary can all produce similar high-contrast patterns yet require different restoration behavior. Conversely, severe localized degradation may require the model to rely more heavily on temporal evidence from neighboring frames. A restoration model should therefore reason explicitly about where the footage is damaged and how strongly to intervene, not just reconstruct plausible images. We propose DART, a Degradation-Aware Recurrent Transformer for archival film restoration. It predicts a continuous soft defect mask, propagates this mask through time, and uses it to guide recurrent fusion between the current observation and the aligned temporal state. Crucially, unlike prior restorers that also rely on defect masks but learn them only implicitly through the reconstruction loss, DART supervises this mask directly against ground-truth defect locations, so damage localization is optimized explicitly during training rather than emerging as a by-product of reconstruction. In addition, DART summarizes the predicted mask and residual evidence into a global degradation condition, allowing the restoration transformer to adapt its behavior to the severity of the current frame. In summary, our main contributions are: – We introduce DART, a degradation-aware recurrent restoration framework for compound archival film degradation, shifting old-film restoration from passive reconstruction to explicit damage-aware processing. – We propose a multi-scale Dilation Pyramid MaskNet trained with direct continuous-mask supervision, enabling the model to localize both the position and severity of film artifacts.

DART: Degradation-Aware Recurrent Transformer

3

– We condition the restoration backbone on degradation through AdaLN-Zero modulation driven by the predicted mask and residual evidence, allowing the network to adapt its restoration behavior to the severity of each frame. – We achieve state-of-the-art results on real archival-film benchmarks while maintaining a compact model.

2

Related Work

Temporal modeling architectures. A central challenge in video restoration is generating details while maintaining consistency across frames. Early approaches incorporated temporal information by aggregating neighboring frames. FastDVDnet [24] and UDVD [23] concatenate adjacent frames to enhance the target frame, while Lei et al. [12] introduces temporal consistency constraints to reduce frame-to-frame variations. Later methods modeled frame correspondences through alignment and propagation mechanisms. EDVR [30] and TDAN [26] introduced deformable alignment modules to handle complex motion without explicit optical flow [6]. Recurrent architectures improved temporal information exchange by propagating features across the sequence. BasicVSR and BasicVSR++ [1,2] demonstrated bidirectional feature propagation, while later works explored propagation in latent representations [16, 28]. RVRT [15] combined recurrent propagation with guided deformable attention to capture long-range dependencies and spatial correspondences. Closer to our setting, RTN [27] brought the recurrent-transformer paradigm to old-film restoration, propagating a hidden state to carry clean content across damaged frames. To reduce the latency of recurrent approaches, recent methods investigate more efficient temporal fusion. DeepRemaster [8] introduced reference-based attention in the bottleneck of an autoencoder, enabling joint processing of multiple input frames. More recently, Mao et al. [19] adopted Mamba-based state-space modeling [4] as an efficient alternative to self-attention, applied along the spatial dimension. Despite their strong temporal modeling, these methods remain degradation-unaware: they reconstruct clean frames without ever being told where a frame is damaged or how severely, and none conditions its restoration on the degradation itself. Making restoration explicitly damage-aware first requires localizing the damage. Degradation segmentation. Explicit mask prediction is a critical strategy in restoration used to isolate corrupted regions and prevent the blurring of clean textures. Historically, masking has been very common in dehazing [17, 21] to estimate spatial density maps that point out exactly where haze is concentrated. Similarly, it is intuitively used to segment out distinct artifacts like rain streaks [5], or to mark extremely dark, information-sparse spots in low-light image enhancement [7]. While these spatial priors are effective for static scenes, masking becomes similarly useful for old video enhancement [16, 19]. Historical films suffer from unstructured, temporally random defects like scratches and dust. To maintain temporal consistency and preserve structural fidelity, modern architectures use degradation masks to isolate these artifacts. Methods such

4

Jastrzębski et al.

as [16, 19, 27] rely on feature map differences to estimate these masks. However, because natural scene dynamics also alter these features, the network often misidentifies camera or object motion as true film degradations. A recent detector [18] instead supervises damage segmentation directly, avoiding this confusion, but it targets only binary, structural defects and runs as a separate detection stage decoupled from the restoration network. DART instead supervises a continuous defect mask end-to-end within the restoration network, capturing not only where damage occurs but how severe it is, across the full range of compound film degradation rather than binary structural damage alone. By anchoring mask generation to ground-truth defects, our approach targets only actual artifacts, disentangling them from scene motion.

3

Method

3.1

Overview

DART restores a T -frame clip {x1 , . . . , xT } with a bidirectional recurrent network. Each clip is processed once forward, x1 → xT , and once backward, xT → x1 ; − − the resulting hidden states → s t and ← s t are merged by a final Fusion convolution into the restored frame x̃t . Figure 2 depicts a single forward step; the backward pass repeats the same computation in reverse temporal order. A shared encoder E maps the current frame to a feature E(xt ), and a frozen RAFT [25] module estimates the optical flow ft−1→t between xt−1 and xt . A Warp operator W (·, ·) resamples its first argument along a given flow field; applied with ft−1→t to the previous hidden state st−1 , the previous soft defect mask Mt−1 , and the previous frame xt−1 , it brings all three into alignment with the align current coordinate system, yielding salign t−1 , Mt−1 , and a warped previous frame. Comparing the warped previous frame against xt gives a residual indicator Rt that estimates where the frame has changed or is occluded, \mathbf {R}_t = \big | W(\mathbf {x}_{t-1}, f_{t-1 \to t}) - \mathbf {x}_t \big | . \label {eq:residual}

(1)

The soft mask Mt , produced by the MaskNet introduced next, then gates a temporal fusion of the encoded current frame with the aligned hidden state into the representation Ft that is passed on to the restoration backbone, \mathbf {F}_t = E(\mathbf {x}_t)\odot (1-\mathbf {M}_t) + \mathbf {s}^{\text {align}}_{t-1}\odot \mathbf {M}_t ; \label {eq:fusion}

(2)

wherever Mt is close to 1, the network prefers the historical, presumably undamaged state over the current, possibly corrupted observation. Three stages then turn this signal into the restored frame: a MaskNet localizes the damage (Sec. 3.2), a Condition Encoder distills the mask and residual into a degradationseverity signal (Sec. 3.4), and a Swin backbone reconstructs the frame under that signal’s guidance (Sec. 3.4), all trained end-to-end with an explicit masksupervision objective (Sec. 3.5).

DART: Degradation-Aware Recurrent Transformer Encoded Frame 𝑬(𝒙𝑡 )

Global Average Pool 𝐶ℎ × 𝐻 × 𝑊 → mean(dim=(2, 3)) → 𝐶ℎ

𝒔𝑡

+

Conv

Conv

+

Unembed

Conv

Embed

Conv

+

Gate

𝒛𝒍+𝟏

𝛼2

Modulate

+

LayerNorm

Modulate

(S)W-MSA

Gate

𝛼1

𝛾2 , 𝛽2

SiLU

SiLU

Linear

Linear

𝒛𝒍 𝑨(∗)

𝛾1 , 𝛽1

LayerNorm

(c) AdaLN Swin Transformer Block

𝑨(∗) 𝑎𝑙𝑖𝑔𝑛

𝐬𝑡−1 , 𝑬(𝒙𝑡 )

AdaLN Swin Transformer Block

𝑎𝑙𝑖𝑔𝑛

𝒔𝑡−1 Glob Avg Pool

SiLU

SiLU

Conv

SiLU

Conv

Conv

Concat input

Sigmoid

Conv

Conv

𝑴𝑡

𝑨(𝐌𝑡 , 𝑹𝑡 )

Swin Restoration Transformer

(b) AdaLN Condition Encoder 𝑨 ∗ = 𝑨(𝑴𝑡 , 𝑹𝑡 )

LReLU

Dil 1

LReLU

Dil 2

LReLU

Dil 3

LReLU

Conv

LReLU

Concat input

(a) MaskNet Dilation Pyramid

C

AdaLN Condition Encoder

Warp 𝑴𝑡−1

Flow

𝐾× 𝑁×

𝒔𝑡−1

𝑎𝑙𝑖𝑔𝑛

C

𝑴𝑡−1

𝒇𝑡−1 →𝑡

Optical Flow 𝒇𝑡−1 →𝑡

𝒔𝑡 backward state forward 𝒔𝑡 state

Fusion

𝐅t

𝑹𝑡

Residual

Mask Gate (Eq. 2)

𝑎𝑙𝑖𝑔𝑛

𝐬𝑡−1 , 𝑬(𝒙𝑡 )

𝑴𝑡

Residual 𝑹𝑡 (𝒙𝑡 , 𝒙𝑡−1 )

Input Frames: 𝒙𝑡 , 𝒙𝑡−1

𝒙𝑡 , 𝒙𝑡−1

MaskNet Dilation Pyramid

C

MLP

𝑬(𝒙𝑡 )

E

𝑴𝑡

Linear

𝒙𝑡

෥𝑡 Restored Frame: 𝒙

𝑴𝑡

Input Frame: 𝒙𝑡

5

Fig. 2: Overview of the proposed DART architecture in one recurrent forward step. C denotes the concatenation operation. The upper portion of the figure illustrates the general flow of the framework, feature alignment and fusion.

3.2

Dilation Pyramid MaskNet

MaskNet is the component that decides, at every timestep, which regions of the current frame should be trusted and which should instead be filled in from the aligned hidden state salign t−1 . A shallow, directly-connected predictor, the design used by RTN [27] and left architecturally unchanged in MambaOFR [19], has a receptive field too small to tell genuine scratches apart from thin scene structure such as fence posts, branches, or drifting smoke: both produce a similarly narrow, high-contrast response at the pixel level, so a predictor that only ever looks at a small local window cannot distinguish them. DART replaces this shallow predictor with a multi-scale MaskNet Dilation Pyramid (bottom of Fig. 2 (a)). A shared convolutional stem branches into three parallel paths with dilation rates d ∈ {1, 2, 3}; their outputs are concatenated and merged by a fused head. Because the three branches observe the same input at three different effective receptive fields, the network can confirm a candidate defect against a wider spatial context before committing to it, rather than reacting to a single narrow window in isolation. Section 3.3 gives the pyramid’s full input and the resulting mask Mt ; combined with the direct supervision introduced in Sec. 3.5, this wider receptive field is what lets DART’s mask track genuine film damage rather than scene geometry, as Sec. 4.3 shows on real archival footage. 3.3

Recurrent Soft-Mask Propagation

Predicting Mt from scratch at every timestep, as RTN does, means that a persistent defect such as a vertical scratch must be independently re-discovered

6

Jastrzębski et al.

in every single frame, a form of temporal amnesia that wastes model capacity and produces flickering, temporally-inconsistent mask predictions. Following the propagation mechanism introduced in MambaOFR [19], DART instead flowwarps the previous soft mask into the current coordinate frame and feeds it back into the Dilation Pyramid alongside the encoded current frame, the aligned hidden state, and the residual indicator: \mathbf {M}_t = \mathcal {M}\!\left (E(\mathbf {x}_t)\,\Vert \,W(\mathbf {s}_{t-1},f_{t-1\to t})\,\Vert \,\mathbf {R}_t\,\Vert \,W(\mathbf {M}_{t-1},f_{t-1\to t})\right ), \label {eq:mask}

(3)

where ∥ denotes channel-wise concatenation and M the Dilation Pyramid of Sec. 3.2. This temporal feedback allows DART to track persistent defects across the clip. As a result, the mask supervision (Sec. 3.5) acts on a temporally coherent sequence rather than relying on independent per-frame estimates. 3.4

AdaLN-Zero Degradation Conditioning

The mask-gated fusion of Eq. (2) operates spatially: pixel by pixel, it decides whether to trust the current frame or the historical state. Nothing in that operation tells the restoration backbone how severely, overall, the current frame is damaged, so a lightly speckled frame and a heavily cracked one would otherwise be processed by identical transformer weights. DART closes this gap with global degradation conditioning. A Condition Encoder first concatenates the newly predicted mask Mt with the residual indicator Rt and passes the result through three alternating Conv/SiLU layers, a Global Average Pool, and a two-layer MLP, collapsing the spatial degradation evidence into a single condition vector A(*) = A(\mathbf {M}_t, \mathbf {R}_t). \label {eq:condition}

(4)

This vector then modulates every block of the Swin restoration backbone. Each AdaLN Swin Transformer Block maps an input feature zl to zl+1 by projecting A(∗), through a shared SiLU-Linear layer, into per-block modulation parameters (γ1 , β1 , γ2 , β2 ) and gate scalars (α1 , α2 ). A normalized feature x is scaled and shifted by \mathrm {Modulate}(\mathbf {x},\gamma ,\beta ) = \mathbf {x}\odot (1+\gamma ) + \beta , \label {eq:modulate} (5) and the block’s two branches, windowed self-attention and an MLP, are each modulated, applied, and re-scaled by their own gate before being added back to the residual stream: \begin {aligned} \mathbf {z}^{l}_{in} &= \mathbf {z}^l + \alpha _1 \odot \text {(S)W-MSA}\big (\text {Modulate}(\text {LayerNorm}(\mathbf {z}^l), \gamma _1, \beta _1)\big ), \\ \mathbf {z}^{l+1} &= \mathbf {z}^{l}_{in} + \alpha _2 \odot \text {MLP}\big (\text {Modulate}(\text {LayerNorm}(\mathbf {z}^{l}_{in}), \gamma _2, \beta _2)\big ). \end {aligned} \label {eq:adaln-block}

(6)

Here (S)W-MSA denotes (shifted) window multi-head self-attention, the standard Swin building block [14] whose window partitioning shifts between consecutive blocks so that information is exchanged across window boundaries, and MLP a two-layer feed-forward network. Following the AdaLN-Zero initialization

DART: Degradation-Aware Recurrent Transformer

7

strategy of [22], the final linear layer that produces (γ, β, α) is initialized to zero, so every block begins training as an identity mapping and only gradually learns to modulate its features as A(∗) becomes informative, which we find stabilizes early training considerably. In effect, the same restoration weights are reused across the whole clip, but the network is free to focus its attention in proportion to how damaged the current frame actually is. This mechanism makes DART’s restoration core, and not merely its mask, degradation-aware. 3.5

Training Objective and Mask Supervision

DART is optimized end-to-end with a combined objective that balances faithful reconstruction against direct supervision of the defect mask. Reconstruction and adversarial terms. A pixel-wise L1 term penalizes deviation from the clean ground-truth frame yt , averaged over the temporal window, and a VGG19-based perceptual term Lperc [10] sharpens texture; together they form Lrec = L1 +λperc Lperc . A Temporal-PatchGAN discriminator [3] built from 3D convolutions is trained with a hinge loss to distinguish restored clips from clean ones, and DART is trained to fool it via the adversarial term Ladv . Mask supervision. Reconstruction alone gives MaskNet only an indirect, weak training signal, since a mask can be substantially wrong while the reconstructed frame still looks acceptable. DART instead supervises the mask directly against a continuous ground-truth mask Mgt , exported alongside the synthetic training pairs (Sec. 4.1), applied to the raw pre-sigmoid MaskNet logits Z in both the forward and backward propagation passes; head frames, for which no temporal gate yet exists, are excluded. Because defect pixels are sparse, the binary cross-entropy term is class-balanced with a positive weight wpos = clamp(Nneg /Npos , wmin , wmax ), where Npos and Nneg count positive and negative ground-truth pixels: \mathcal {L}_{BCE}^{(p)} = \mathrm {BCEWithLogits}\!\left (\mathbf {Z}^{(p)},\, \mathbf {M}^{gt,(p)},\, w_{pos}\right ). \label {eq:bce}

(7)

A differentiable Dice term further rewards structural overlap between the predicted soft mask M = σ(Z(p) ) and the ground truth: \mathcal {L}_{Dice}^{(p)} = 1 - \frac {2\sum (\mathbf {M}\odot \mathbf {M}^{gt,(p)}) + \epsilon }{\sum \mathbf {M} + \sum \mathbf {M}^{gt,(p)} + \epsilon }. \label {eq:dice}

(8)

The two terms are combined and averaged over both propagation directions p ∈ {←, →}: \mathcal {L}_{mask} = \frac {1}{2}\sum _{p\in \{\leftarrow ,\rightarrow \}}\left (\mathcal {L}_{BCE}^{(p)} + \lambda _{dice}\,\mathcal {L}_{Dice}^{(p)}\right ). \label {eq:mask-loss}

(9)

Unlike prior methods that infer defect locations indirectly through reconstruction, this explicit formulation forces MaskNet to learn the exact position and severity of degradations directly from the ground truth. Full objective. The three terms are combined as \mathcal {L}_{total} = \lambda _{rec}\,\mathcal {L}_{rec} + \lambda _{adv}\,\mathcal {L}_{adv} + \lambda _{mask}\,\mathcal {L}_{mask}, \label {eq:total}

(10)

8

Jastrzębski et al. Table 1: Training settings and loss weights of DART. Setting

Value

Optimizer Learning rate Epochs Window / crop / batch Precision λrec , λadv , λmask , λdice

Adam (β1 =0.9, β2 =0.99) 2 × 10−4 (linear decay) 20 7 frames / 256×256 / 1 clip FP16 (mixed) 1.0, 0.01, 0.5, 0.5

with the loss weights λrec , λadv , λmask , λdice reported alongside the remaining training settings in Table 1.

4

Experiments

4.1

Datasets, Metrics, and Implementation

Training data. DART is trained on REDS [20] train_sharp (720p, 24 fps). During training, a synthetic degradation pipeline, adapted from AbsoluteDegradation [9], generates training pairs online, yielding for each 256 × 256 crop a ground-truth sharp frame, a synthetically degraded version of it, and a continuous artifact mask: the mask outlines exactly where the synthetic defects were applied and undergoes the same degradation as the degraded frame itself. This mask provides the supervision target for Lmask (Eq. (9)). The official 30 REDS val_sharp clips are reserved as a synthetic, full-reference proxy test set, used only for the ablation study in Sec. 4.4. Real-world evaluation. No clean references exist for genuine archival footage, so evaluation on real data is necessarily no-reference. AbsoluteDegradation [9], a 13,252-frame, lossless-PNG corpus of public-domain archival footage (1896– 1918), is the primary target domain; SRWOV [19], 216 lower-resolution, JPEGcompressed real-world clips, serves as a secondary benchmark. Metrics. Real-footage quality is scored with three no-reference metrics: CLIPIQA+ [29] for general naturalness and perceptual alignment, and MUSIQ [11] and MANIQA [32], which correlate most closely with human quality judgments and show the widest, most consistent margins throughout this section. On the synthetic proxy set, standard full-reference PSNR, SSIM [31], and LPIPS [33] are reported for ablation studies. Implementation. DART is implemented in PyTorch and trained on two NVIDIA A100 GPUs with FP16 mixed precision, using a frozen RAFT [25] flow estimator. Table 1 lists the remaining training settings and loss weights. Every competing architecture in Sec. 4.2 is retrained from scratch on identical data using its own published hyperparameters, so all comparisons in this section share a common training source and evaluation protocol.

DART: Degradation-Aware Recurrent Transformer

9

Table 2: Quantitative comparison of restoration architectures, all trained on the AbsoluteDegradation pipeline and evaluated on the AbsoluteDegradation and SRWOV [19] datasets. Bold marks the best score, and underline marks the second best, in each metric column. The input frames row denotes the unrestored degraded footage and serves as a reference floor. AbsoluteDegradation Dataset Model

SRWOV Dataset

CLIPIQA+ ↑MUSIQ ↑MANIQA ↑CLIPIQA+ ↑MUSIQ ↑MANIQA ↑

Input frames

0.2483

26.8418

0.1619

0.3890

42.8591

0.2307

BasicVSR [1] BasicVSR++ [2] ShiftNet [13] DeepRemaster [8] RTN [27] MambaOFR [19] DART (Ours)

0.4049 0.3685 0.4309 0.3574 0.4220 0.4190 0.4334

49.5797 42.9144 46.7724 44.7540 46.8737 47.2436 51.2304

0.2844 0.2375 0.2764 0.2542 0.2661 0.2777 0.3011

0.4086 0.4428 0.4578 0.3722 0.4441 0.4402 0.4488

57.6442 58.0442 57.9334 53.5473 58.4585 57.3551 59.4263

0.3104 0.3231 0.3336 0.2637 0.3321 0.3210 0.3427

4.2

Comparison with State of the Art

Quantitative results. Table 2 compares DART against BasicVSR [1], BasicVSR++ [2], ShiftNet [13], DeepRemaster [8], RTN [27], and MambaOFR [19], all retrained on identical data. DART outperforms all baselines on every metric on AbsoluteDegradation. On SRWOV it remains the best on MUSIQ and MANIQA, the two metrics that correlate most closely with human perceptual judgment, and is edged out only on CLIPIQA+, where ShiftNet scores narrowly higher (0.458 vs. 0.449). DART outperforms RTN, the architecture whose Swin restoration backbone it shares, on every metric across both datasets; therefore, this consistent margin isolates what the degradation-aware additions themselves contribute, rather than reflecting a larger or differently-trained network. Qualitative results. Figure 3 tracks three consecutive frames from an archival newsreel clip. BasicVSR++ and MambaOFR leave the emulsion out of focus and lets speckle persist across all three frames; RTN removes some noise but leaves faint vertical scratch tracks; DART alone produces a stable, sharp, and temporally consistent reconstruction. Figure 4 extends the comparison to three further clips: across all of them, DART suppresses structured defects, scratches, and dust while preserving genuine scene texture more faithfully than the generic video-restoration baselines, which either under-react to structured damage or over-sharpen it into the scene. Efficiency. Table 3 shows this quality gain is not bought with a larger network. DART is the second-smallest model evaluated, at 6.6M parameters (only 1.3M trainable beyond the shared, frozen RAFT flow estimator), and among the most memory-efficient at 0.35 GB. It dominates MambaOFR [19] on every axis and improves on ShiftNet [13] in parameters, memory, and speed. As archival restoration runs offline, its moderate per-frame latency is not a practical limitation.

10

Jastrzębski et al.

Frame 1

Frame 2

DART

MambaOFR

RTN

ShiftNet

BasicVSR++

Input

Frame 0

Fig. 3: Qualitative comparison on three consecutive frames of an archival clip. Competing methods leave vertical scratch tracks, over-sharpen emulsion noise, or leave the frame out of focus; DART produces a cleaner, sharper, and more temporally consistent reconstruction across the sequence.

11

DART

MambaOFR

RTN

ShiftNet

BasicVSR++

Input

DART: Degradation-Aware Recurrent Transformer

Fig. 4: Qualitative comparison on archival clips. DART suppresses structured defects and scratches while preserving scene texture more faithfully than the generic videorestoration baselines, which either under-react to structured damage or over-sharpen it into the scene.

12

Jastrzębski et al.

Table 3: Computational cost and restoration quality, measured per frame on a 1×7×3×256×256 clip (NVIDIA A100 GPU, fp32, 30 timed iterations after 5 warmup). MUSIQ is from Tab. 2. Params include the shared, frozen RAFT (5.3M); DART adds 1.3M trainable. Bold/underline: best/second-best. Method

Params (M) ↓

FLOPs (G) ↓

Mem (GB) ↓

FPS ↑

MUSIQ ↑

9.9 13.6 13.0 9.9 6.2 8.7 6.6

910 381 316 131 349 442 372

0.34 0.36 1.06 0.39 0.35 0.64 0.35

27.6 30.6 21.4 529.3 33.8 14.5 29.5

49.6 42.9 46.8 44.8 46.9 47.2 51.2

BasicVSR [1] BasicVSR++ [2] ShiftNet [13] DeepRemaster [8] RTN [27] MambaOFR [19] DART (Ours)

4.3

Mask Behavior on Real Footage

Because the defect masks in RTN and MambaOFR are never directly supervised, their accuracy on real archival footage, where no ground-truth mask exists to check them against, cannot be inferred from training loss alone. Figure 5 and Figure 6 illustrate two recurring failure modes of this unsupervised design, both absent from DART. Figure 5 isolates a diagonal deck railing on a ship, genuine scene structure with exactly the narrow, high-contrast profile of an analog scratch. RTN and MambaOFR both partially erase it, mistaking the railing for mechanical damage and blending it away with the aligned historical state, while DART’s widerreceptive-field mask recognizes the railing as scene content and restores it intact. Figure 6 shows the complementary failure. The input frame contains a severe, unambiguous chemical crack running through the emulsion. The RTN and MambaOFR mask activations (top row) are almost uniformly flat: their unsupervised masks do not activate on this defect at all, so the restoration backbone receives no localization signal, and the crack survives essentially unchanged in both restored outputs (bottom row). DART’s mask activates precisely along the crack, and the corresponding output attenuates it substantially. Taken together, the two figures suggest that an unsupervised mask is not simply less precise than a supervised one. Without a ground-truth target to anchor it, an unsupervised mask can fail in either direction: activating falsly on undamaged structure, or failing to activate on real, severe damage. Either failure leaves the restoration backbone with no reliable notion of where the frame is degraded, or how badly. Direct supervision is what keeps DART’s localization anchored to the defect itself, rather than to whatever correlate of damage the reconstruction loss happened to pick up. 4.4

Ablation Study

Table 4 isolates the contribution of each core component by removing it from the full model (rows marked −). Every component earns its place: removing the

DART: Degradation-Aware Recurrent Transformer Input

RTN

MambaOFR

13

DART (Ours)

Fig. 5: RTN and MambaOFR mistake a ship’s deck railing, genuine scene structure with a scratch-like profile, for film damage and partially erase it during temporal fusion; DART’s mask recognizes the structure and restores it intact. Input

RTN

MambaOFR

DART (Ours)

Fig. 6: Mask activations (top) and restored crops (bottom) for a severe chemical crack in the emulsion. RTN and MambaOFR’s unsupervised masks stay essentially flat and fail to localize the defect at all, so it survives largely unrestored; DART’s supervised mask activates precisely on the crack, and the corresponding output attenuates it substantially.

Dilation Pyramid MaskNet, the recurrent mask pass, or the AdaLN conditioning each lowers every perceptual metric on real footage, with AdaLN conditioning the single most influential component, its removal alone drops MANIQA from 0.301 to 0.263. This pattern follows from how the components depend on one another: a sharper, supervised mask feeds a more informative signal into the Condition Encoder, so the severity vector A(∗) it produces genuinely reflects the damage present, and the restoration transformer is in turn modulated by a trustworthy signal rather than noise. Weaken the mask, and the conditioning that depends on it degrades with it. Several further design alternatives explored during development, deformable alignment, a bidirectional residual indicator, SPADE-style spatial modulation, and direct mask concatenation, as well as variations in temporal window length and optical-flow backbone, did not improve on this configuration consistently across metrics. The one row that looks like a counterexample, − mask supervision, is in fact the clearest illustration of this point: it produces the best PSNR, SSIM, and LPIPS on the synthetic proxy set, yet loses on every perceptual metric on real

14

Jastrzębski et al.

Table 4: Ablation study on the AbsoluteDegradation (real footage, no-reference) and Synthetic (full-reference) datasets. − indicates removing the corresponding mechanism from the full DART model. Best scores are in bold and second-best scores are underlined. AbsoluteDegradation Dataset Ablation

Synthetic Dataset

CLIPIQA+ ↑MUSIQ ↑ MANIQA ↑ PSNR ↑ LPIPS ↓ SSIM ↑

Input frames

0.2483

26.842

0.1619

19.540

0.5421

0.5882

DART - previous mask - AdaLN injection - MaskNet - mask supervision

0.4334 0.4332 0.4167 0.4321 0.4173

51.230 50.772 48.685 49.815 49.451

0.3011 0.2927 0.2629 0.2905 0.2912

23.665 23.815 23.622 23.733 23.975

0.1835 0.1860 0.1984 0.1852 0.1787

0.7294 0.7326 0.7295 0.7309 0.7380

footage, because full-reference distortion metrics reward the safe, slightly oversmoothed reconstructions that an unsupervised mask tends to produce, while no-reference perceptual metrics correctly penalize the resulting loss of authentic detail. Since the goal of this work is faithful restoration of real archival film rather than distortion scores on a synthetic proxy, a strong synthetic score is not evidence of good real-world restoration, and the supervised configuration remains the principled choice.

5

Conclusion and Limitations

We presented DART, a degradation-aware recurrent transformer that shifts archival film restoration from passive reconstruction to explicit damage-aware processing. DART predicts a directly supervised, continuous soft defect mask and propagates it temporally to track artifacts coherently. This mask serves a dual purpose: it guides recurrent temporal fusion and, via AdaLN-Zero conditioning, modulates the restoration backbone based on local degradation severity. Consequently, DART achieves state-of-the-art perceptual quality on two realworld benchmarks while adding just 1.3M trainable parameters. Crucially, our direct mask supervision prevents common unsupervised failure modes, ensuring genuine scratch-like scene structures are preserved and severe damages are correctly localized. However, DART currently targets monochrome footage and, like all flowbased propagators, relies heavily on optical flow accuracy, making extreme occlusions or highly persistent defects challenging. Furthermore, while we rely on established no-reference metrics, they do not fully substitute for human evaluation. Extending this degradation-aware framework to color restoration and validating results through formal perceptual studies remain promising directions for future work.

DART: Degradation-Aware Recurrent Transformer

15

References 1. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: BasicVSR: The search for essential components in video super-resolution and beyond. In: CVPR (2021) 2. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In: CVPR (2022) 3. Chang, Y.L., Liu, Z.Y., Lee, K.Y., Hsu, W.: Free-form video inpainting with 3d gated convolution and temporal patchgan. In: ICCV (2019) 4. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023) 5. Guan, C., Yoshie, O.: Clip-driven rain perception: Adaptive deraining with patternaware network routing and mask-guided cross-attention. Pattern Recognition p. 112886 (2025) 6. Horn, B.K., Schunck, B.G.: Determining optical flow. Artificial intelligence 17(1-3), 185–203 (1981) 7. Hou, J., Zhu, Z., Hou, J., Liu, H., Zeng, H., Yuan, H.: Global structure-aware diffusion process for low-light image enhancement. Advances in Neural Information Processing Systems 36, 79734–79747 (2023) 8. Iizuka, S., Simo-Serra, E.: DeepRemaster: Temporal source-reference attention networks for comprehensive video enhancement. In: ACM TOG (2019) 9. Jastrzębski, M., Glinkowski, D., Zieliński, D., Borkowski, D., Kozłowski, W.M., Adamczewski, K.: AbsoluteDegradation: A physics-inspired synthetic filmdegradation pipeline and archival film restoration benchmark (2026) 10. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV (2016) 11. Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: MUSIQ: Multi-scale image quality transformer. In: ICCV (2021) 12. Lei, C., Chen, Q.: Fully automatic video colorization with self-regularization and diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3753–3761 (2019) 13. Li, D., Shi, X., Zhang, Y., Cheung, K.C., See, S., Wang, X., Qin, H., Li, H.: A simple baseline for video restoration with grouped spatial-temporal shift. In: CVPR (2023) 14. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: ICCV (2021) 15. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Van Gool, L.: Recurrent video restoration transformer with guided deformable attention. NeurIPS (2022) 16. Lin, S., Simo-Serra, E.: Restoring degraded old films with recursive recurrent transformer networks. In: WACV (2024) 17. Liu, Q., Song, T., Qin, A., Liu, Y., Yang, F., Gao, C.: Hdsa-net: Haze density and semantic awareness network for hyperspectral image dehazing. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 18, 3989– 4003 (2025) 18. Liu, Q., Liu, Y., Wang, L., Yan, F., Zhang, Q., Ju, H.: Restoration of archival film with large areas of structural damage. npj Heritage Science (2026) 19. Mao, Y., Luo, H., Zhong, Z., Chen, P., Zhang, Z., Wang, S.: Making old film great again: Degradation-aware state space model for old film restoration. In: CVPR (2025)

16

Jastrzębski et al.

20. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Lee, K.M.: NTIRE 2019 challenge on video deblurring and super-resolution: Dataset and study. In: CVPRW (2019) 21. Ngo, D., Lee, G.D., Kang, B.: Haziness degree evaluator: A knowledge-driven approach for haze density estimation. Sensors 21(11), 3896 (2021) 22. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023) 23. Sheth, D.Y., Mohan, S., Vincent, J.L., Manzorro, R., Crozier, P.A., Khapra, M.M., Simoncelli, E.P., Fernandez-Granda, C.: Unsupervised deep video denoising. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1759–1768 (2021) 24. Tassano, M., Delon, J., Veit, T.: Fastdvdnet: Towards real-time deep video denoising without flow estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1354–1363 (2020) 25. Teed, Z., Deng, J.: RAFT: Recurrent all-pairs field transforms for optical flow. In: ECCV (2020) 26. Tian, Y., Zhang, Y., Fu, Y., Xu, C.: Tdan: Temporally-deformable alignment network for video super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3360–3369 (2020) 27. Wan, Z., Zhang, B., Chen, D., Liao, J.: Bringing old films back to life. In: CVPR (2022) 28. Wan, Z., Zhang, B., Chen, D., Zhang, P., Chen, D., Liao, J., Wen, F.: Bringing old photos back to life. In: CVPR (2020) 29. Wang, J., Chan, K.C., Loy, C.C.: Exploring CLIP for assessing the look and feel of images. In: AAAI (2023) 30. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0–0 (2019) 31. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600–612 (2004) 32. Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y., Cao, M., Wang, J., Yang, Y.: MANIQA: Multi-dimension attention network for no-reference image quality assessment. In: CVPRW (2022) 33. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)