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

arXiv:2607.21219 · cs.CV, cs.LG · Submitted 2026-07-23 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "DART: A Degradation-Aware Recurrent Transformer for Archival Film Restoration".

Jane: Archival film restoration is addressed by DART, a degradation-aware recurrent transformer designed to move from passive reconstruction to explicit damage-aware processing.

Tom: First, who's behind it and why it matters.

Paper summary: Jane: So, looking back at "DART: A Degradation-Aware Recurrent Transformer for Archival Film Restoration," the authors have really put together a system that goes beyond standard restoration by predicting and propagating a soft defect mask through time to guide temporal fusion. They claim this approach yields cleaner and more temporally consistent results while remaining compact.

Tom: It’s fascinating how they moved from treating degradation only implicitly to explicitly processing it through this degradation-aware recurrent transformer structure. The whole mechanism of predicting that soft mask and then using it to condition the network via AdaLN-Zero modulation is what makes this work so interesting when you think about complex film artifacts.

Lu: I see the implication here for future research; because they’ve shown how to decouple damage localization from pure reconstruction loss through direct mask supervision, it gives researchers a clearer target for designing next-generation restoration frameworks.

Meng: From an engineering viewpoint, if this framework can consistently deliver high fidelity on real archival footage without requiring massive, slow models, that suggests we could deploy more sophisticated restoration tools in environments where resources are constrained but quality is paramount.

Lalam: This paper has the potential to improve how we understand and preserve cultural history; having a tool that can accurately model and restore damaged historical media means we can access and interpret that footage with much greater fidelity than before.

Tom: Exactly, Jane; the title itself, "DART: A Degradation-Aware Recurrent Transformer for Archival Film Restoration," tells you precisely what it is—a degradation-aware recurrent transformer specifically for archival film. It’s a very descriptive name for such a complex system.

Jane: And the implications are that we are moving toward restoration methods that understand the *context* of damage across time, not just isolated pixel errors in a single frame. This contextual understanding could be applied to any sequence where temporal coherence is vital, not just film.

Lu: The ability to condition the restoration backbone on severity via that condition vector A suggests a level of fine-grained control over the restoration process that was previously inaccessible with standard recurrent architectures <ref:2607.21219#pg0>.

Meng: I’m still focused on practical application; if this model can handle the diverse issues of old film—scratches, dust, photometric aging—with high fidelity and reasonable speed, it could significantly lower the barrier for digital preservation in museums and archives globally.

Lalam: I think the biggest impact is how this advances AI's capability to handle historical artifacts; it means we can preserve cultural memory with a level of accuracy that was previously unattainable because clean reference videos simply don't exist for most old footage.

Tom: So, to wrap up, DART shows that by explicitly modeling damage through a propagating mask and using that signal to condition the network adaptively, we can achieve strong restoration results on archival film while keeping the model structure relatively efficient. This is a solid piece of work addressing a real bottleneck in digital humanities and media preservation.

Conclusion: Tom: So, we've seen DART move beyond simple reconstruction by predicting that soft defect mask through time to guide the restoration process, and now we're getting to the conclusion of this paper.

Jane: It sounds like the core idea is that DART isn't just fixing pixels; it’s learning *how* damage evolves across a whole sequence of frames. That level of temporal awareness is quite something for archival film work.

Lu: From my perspective, it’s the way the authors shifted from passive reconstruction to active damage-aware processing that really opens up new creative avenues in how we think about media artifacts. It suggests that degradation isn't just noise; it’s a structured signal we can learn from.

Meng: I'm curious about how this translates into a real pipeline; does the recurrent structure keep the computational load manageable for actual production environments? We need to know if this stays compact enough to run reliably on standard hardware.

Lalam: The implication here is huge for cultural preservation; because DART can explicitly model and correct temporal inconsistencies in film, we could potentially restore historical footage with a level of fidelity that current methods just can't reach.

Tom: Exactly, Lalam, that fidelity is what matters most when dealing with fragile historical media like archival film. It’s about making sure the visual narrative remains intact across the entire duration of the clip.

Jane: And when we look at the authors and their work, they’ve done a really neat job of tying together several complex ideas—the mask prediction, the temporal fusion, and that conditioning mechanism—into one cohesive framework.

Lu: The way they structured the Dilation Pyramid MaskNet training with direct continuous-mask supervision is particularly clever; it forces the AI to localize both where something is wrong and how bad it is simultaneously. That’s a sophisticated training strategy.

Meng: I'm still focused on the practical side, though; if we can’t deploy this efficiently, all that theoretical modeling doesn't help in the real world of restoration workflows. We need concrete details on the model size and inference speed for a proper evaluation.

Lalam: Considering everything we've discussed about temporal coherence and explicit damage modeling, I see this as a significant step forward in how AI can engage with cultural heritage; it’s not just about making pictures look better, it's about preserving the context of those pictures.

Tom: And that’s where we end for now; DART really shows us that by treating degradation as something to be explicitly mapped and conditioned upon, we can build restoration tools that are much more robust than what we had before. Next up, we'll look at some of the specific results they achieved on those archival benchmarks.

Mikołaj Jastrzębski, Wojciech Kozłowski, Kamil Adamczewski

Wrocław University of Science and Technology

cs.CV, cs.LG

Submitted: 2026-07-23

Updated: 2026-10-04

Comments: Accepted to ACCV 2026. 14 pages + references, 6 figures, 4 tables. Project page: https://www.mikjas.com/dart Code: https://github.com/TytanMikJas/DART-Degradation-Aware-Recurrent-Transformer

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: Archival film restoration is addressed by DART, a degradation-aware recurrent transformer designed to move from passive reconstruction to explicit damage-aware processing.

Key concepts

Soft Defect Mask (Mt)
This continuous mask predicts where defects are located across frames. It's not just a binary map but a smooth value indicating the severity of degradation at each point in time. DART propagates this mask forward and backward through the film sequence, allowing it to track persistent damage throughout the entire clip.
Temporal Feedback Mechanism
Instead of generating masks independently for every frame, DART uses a feedback loop where the previous soft mask is warped into the current frame's coordinate system. This allows DART to build a temporally coherent understanding of defects by feeding historical damage information back into the current prediction.
AdaLN-Zero Modulation
This technique adapts how the restoration network processes features based on damage severity. The predicted mask and residual evidence are used to modulate feature blocks via AdaLN-Zero. This enables the model to change its restoration strategy dynamically—adapting its behavior to restore severely damaged areas differently than less affected ones.
Mask Supervision Loss (L_mask)
This is a direct training signal where the model is explicitly trained against ground-truth defect locations. Unlike standard reconstruction losses, this loss forces the model to learn precise damage localization during training, ensuring that identifying and correcting film artifacts becomes an explicit optimization goal.

Terminology

Summary

Archival film restoration is addressed by DART, a degradation-aware recurrent transformer designed to move from passive reconstruction to explicit damage-aware processing. The core finding is that 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, leading to cleaner and more temporally consistent restorations while remaining compact.

The gist

DART 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.

Key Contributions

  1. 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.

  2. 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.

  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.

  4. We achieve state-of-the-art results on real archival-film benchmarks while maintaining a compact model.

Method Overview

DART restores a T-frame clip by processing it once forward (x1→xT) and once backward (xT→x1), merging the resulting hidden states via a final Fusion convolution into the restored frame x˜t. The process involves several key components:

** A shared encoder E maps the current frame to a feature E(xt). **

The optical flow module estimates the flow between adjacent frames, and a Warp operator resamples previous elements using this flow field to align them with the current coordinate system. This yields a residual indicator Rt, defined as:

"The residual indicator Rt that estimates where the frame has changed or is occluded, is defined as:

(1) Rt = W(xt-1, ft-1 to t) - xt.

Mask Generation and Propagation

Instead of predicting the mask from scratch at every timestep, DART employs a temporal feedback mechanism. The soft mask Mt is propagated by flowwarping the previous soft mask Mt-1 into the current coordinate frame and feeding it back into the Dilation Pyramid alongside other relevant inputs:

"Mt = M(E(xt) W(st-1, ft-1 to t) Rt W(Mt-1, ft-1 to t>>))" (3)

This propagation allows DART to track persistent defects across the clip, ensuring that the mask supervision acts on a temporally coherent sequence rather than relying on independent per-frame estimates.

Degradation Conditioning

The fusion of the encoded current frame with the aligned hidden state is governed by a mask gate, where areas where Mt is close to 1 prefer the historical state over the current observation. To address how severely damaged a frame is, DART uses a Condition Encoder to distill this evidence:

"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(Mt, Rt>>)." (4)

This resulting vector A(∗) then modulates every block of the Swin restoration backbone, allowing the network to adapt its behavior based on damage severity. The modulation is achieved via AdaLN-Zero:

A normalized feature x is scaled and shifted by Modulate(x, γ, β) = x⊙ (1+γ) + β. (5)

Training Objective

DART is optimized end-to-end using a combined objective that balances reconstruction fidelity with explicit mask supervision. The total loss is:

Ltotal = λrec Lrec + λadv Ladv + λmask Lmask. (10)

The reconstruction term (L rec) combines a pixel-wise L1 term and a VGG19-based perceptual term. The adversarial loss (L adv) uses a Temporal-PatchGAN discriminator to train DART to fool it. Crucially, the mask supervision loss (L mask) directly anchors the model:

DART instead supervises the mask directly against ground-truth defect locations, so damage localization is optimized explicitly during training rather than emerging as a by-product of reconstruction. (3.

Improvements for AI systems

As a fastidious researcher, I have analyzed the DART (Degradation-Aware Recurrent Transformer) paper. The core innovation lies in shifting archival film restoration from implicit reconstruction to explicit, degradation-aware processing via a recurrent transformer architecture guided by a predicted soft defect mask and global severity conditioning.

Here are the specific improvements this system enables for AI applications:

  1. Enhanced Restoration Fidelity for Archival/Historical Media:

The system can restore old films with compound degradations (scratches, dust, blur, noise) with superior perceptual quality (as evidenced by state-of-the-art scores on no-reference metrics like MUSIQ and MANIQA). Unlike prior methods that might over-sharpen texture or under-react to structural damage, DART preserves genuine scene textures while accurately suppressing artifacts like scratches and dust.

  1. Explicit Damage Localization:

The system explicitly predicts a continuous soft defect mask, which is supervised against ground truth. This allows the AI to distinguish between genuine film damage (like a scratch) and similar high-contrast scene structures (like a fence post). This prevents the model from mistaking scene geometry for degradation and ensures that restoration efforts are precisely targeted where needed.

  1. Temporal Coherence Across Long Sequences:

By propagating the defect mask through time (Eq. 3), DART maintains temporal consistency across entire video clips. It tracks persistent defects (like a vertical scratch) across frames, preventing the temporal amnesia and flickering observed in unsupervised methods like RTN or MambaOFR.

  1. Adaptive Restoration Behavior Based on Severity:

The system uses an AdaLN-Zero conditioning mechanism to distill the predicted mask and residual evidence into a global degradation severity signal. This allows the restoration backbone (Swin Transformer) to adapt its behavior frame-by-frame—applying stronger interventions where damage is severe and more conservative restoration where the frame is clean.

  1. Efficient, Compact Restoration:

The architecture achieves state-of-the-art performance while remaining highly efficient, utilizing only 6.6M parameters (trainable) and maintaining a small memory footprint of 0.35 GB per frame for inference, making it practical for offline archival processing where latency is not the primary concern.

In summary, this system creates an AI capable of performing high-fidelity, temporally consistent restoration of historical video footage by treating degradation not as a noise nuisance to be filtered implicitly, but as a quantifiable signal that directly guides and conditions the restoration process.

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.

Sources

Related papers