Archival Film Restoration - ML Research

Absolute Cinema

Bringing century-old footage back to life - denoising, sharpening, defect in-painting and de-flickering - with a physically-grounded degradation pipeline and a state-of-the-art restoration model.

Mikołaj Jastrzębski, Dawid Glinkowski, Dawid Zieliński, Daniel Borkowski - advised by Kamil Adamczewski, Wojciech Kozłowski

Wrocław University of Science and Technology

AbsoluteDegradation

Dataset arXiv Code

DART

Code Weights

NeurIPS 2026 - Datasets & Benchmarks - Accepted ACCV 2026 - Main Track - Accepted

Original Restored - DART

Loading comparison

Drag to compare - original archival footage vs. DART restoration. Source footage courtesy of the National Digital Archive.

01 Motivation

A vanishing visual record,
locked behind cost and complexity.

Thousands of hours of archival film survive only as fragile, damaged prints. Restoring them today is a manual, expert-driven craft - so most footage is never seen again.

Costly & slow

Frame-by-frame manual restoration takes specialist studios weeks per reel - and costs that only major institutions can bear.

Low watchability

Noise, flicker and damage make raw archival footage hard for audiences to engage with, limiting its cultural reach.

Heritage decay

Film stock degrades faster than it can be digitised and restored. Every year, more footage is lost permanently.

Out of reach

No accessible tool exists for smaller archives, historians or private collectors - the scale of the problem demands automation.

02 Our approach

Three contributions,
one end-to-end restoration system.

01 - Benchmark

A new test set

81,576 frames of manually curated genuinely archival footage at up to 1440x1080 - watermark-free, heavily degraded, purpose-built to evaluate real-world restoration.

02 - Pipeline

AbsoluteDegradation

A physically-inspired degradation pipeline that models film-stock damage, generating temporally consistent, reference-paired training data at scale.

03 - Model

DART

Our restoration model - state-of-the-art on every perceptual metric we measured, outperforming previous SOTA architectures.

03 Solution - Benchmark

A test set built for
real archival film.

Where the previous standard (SRWOV) mixed film with animation at low resolution and carried watermarks, our set is archival-only, higher-resolution and heavily degraded - a far harder, more honest benchmark. Dataset is sourced from the public domain collections of U.S. Library of Congress.

- More frames

- Pixels per frame

100% Archival domain

0 Watermarks

Sample clips from our benchmark - archival footage from the U.S. Library of Congress

SRWOV (previous standard) vs. our test set
Property	Previous standard (SRWOV)	Our test set
Frame count	29,256	81,576
Resolution	640 x 360	1440 x 1080
Content domain	Mixed (film + animation)	Archival film only
Watermarks	Present	None
Degradation level	Low	High

04 Solution - Pipeline

No clean original exists -
so we synthesise the damage.

A damaged century-old reel has no pristine counterpart to learn from, so supervised restoration is impossible without paired data. The only way to obtain it is to take clean footage and degrade it ourselves - which means a model is only ever as good as how realistically that synthetic damage mirrors real film decay.

No paired ground truth

You can never capture a clean version of a damaged historical film. Without real pairs, training depends entirely on the realism of synthetic degradations.

Prior pipelines fall short

Earlier synthesisers keep frames structurally close to the clean original - high SSIM, low LPIPS - failing to replicate genuine film damage, so models under-restore real footage.

Authentic, physics-inspired damage

AbsoluteDegradation applies six ordered, physically-grounded stages to produce heavier, more realistic and temporally coherent damage - plus a per-frame degradation mask - closing the domain gap and improving every backbone.

The Pipeline - Six Degradation Stages - clean footage through multiple stages of authentic degradation

Clean input frame before any degradation

INPUT - Clean Clip

Source footage before any processing - propagated unchanged as the ground-truth training reference.

STAGE 1 - Grayscale

Colour channels are collapsed to a single luminance channel, reproducing the monochrome look of original archival black-and-white film stock.

STAGE 2 - Texture Overlay

1–3 scanned emulsion textures overlaid per frame - surface blemishes, particle noise and chemical staining.

STAGE 3 - Degradation Series

Invert, apply blur, grain, downscale and gamma in randomised order, then re-invert - mimicking negative decay. Followed by a random-ordered composition of blur, grain, JPEG compression, downscaling, gamma shift and motion blur.

STAGE 4 - Photometric Artifacts

Resize to target resolution, colour jitter, 8-bit quantize - locking the final degraded appearance.

STAGE 5 - Vertical Line Simulation

Clip-level overlays with Brownian-drift trajectories, per-scratch flicker and multi-frame lifespan.

Output - Degraded / Clean Pair

Ready for
Training

The degraded clip and the clean original are saved as a matched pair - the complete supervised training signal for training.

05 Solution - Model

DART - Degradation-Aware
Recurrent Transformer.

Where prior restoration models treat every frame identically, DART first asks where the damage is and how severe - then restores accordingly. The result is a content-adaptive process that outperforms every prior architecture on our archival benchmark.

DART architecture - mask prediction, temporal warping, and severity-conditioned Swin Transformer blocks.

Dilation Pyramid MaskNet

Three parallel branches at dilation rates d∈{1,2,3} localise artifacts across scales - from fine grain to coarse scratches - producing a dense per-frame damage mask.

AdaLN-Zero Severity Conditioning

Per-frame damage severity is distilled into a global signal that modulates every Swin Transformer block via Adaptive Layer Norm - shifting restoration from static to content-adaptive.

Mask-Weighted Restoration Loss

The loss is weighted by the degradation masks produced by our synthetic degradation pipeline, steering the model to repair only the genuine artifacts while leaving authentic footage untouched.

06 Results - Visual

Drag the handle.
See the restoration.

Original Restored - DART

Loading comparison

Original Restored - DART

Loading comparison

Original Restored - DART

Loading comparison

Original Restored - DART

Loading comparison

Original Restored - DART

Loading comparison

Original Restored - DART

Loading comparison

Source footage: public-domain archival film from the U.S. Library of Congress.

07 Results - Quantitative

State-of-the-art on every
perceptual metric.

Swapping the training data alone makes the difference: BasicVSR++, RTN and MambaOFR all score substantially better on every no-reference perceptual metric when trained on AbsoluteDegradation instead of the prior BOF pipeline. Built on the same data, DART outperforms every competing method on our archival benchmark.

#1 All perceptual metrics

5 Baselines beaten

SOTA Archival restoration

Degradation pipeline

BasicVSR++ backbone
Training data	BRISQUE ↓	CLIPIQA+ ↑	MUSIQ ↑	MANIQA ↑
BOF	30.62	0.313	32.55	0.170
AbsoluteDegradation (ours)	20.68	0.369	42.91	0.238

Shift-Net backbone
Training data	BRISQUE ↓	CLIPIQA+ ↑	MUSIQ ↑	MANIQA ↑
BOF	26.74	0.370	36.74	0.213
AbsoluteDegradation (ours)	16.59	0.438	46.77	0.276

MambaOFR backbone
Training data	BRISQUE ↓	CLIPIQA+ ↑	MUSIQ ↑	MANIQA ↑
BOF	32.79	0.347	38.81	0.221
AbsoluteDegradation (ours)	20.28	0.419	47.24	0.278

RTN backbone
Training data	BRISQUE ↓	CLIPIQA+ ↑	MUSIQ ↑	MANIQA ↑
BOF	19.89	0.366	41.46	0.220
AbsoluteDegradation (ours)	14.39	0.422	46.87	0.266

DeepRemaster backbone
Training data	BRISQUE ↓	CLIPIQA+ ↑	MUSIQ ↑	MANIQA ↑
BOF	35.52	0.320	39.48	0.225
AbsoluteDegradation (ours)	23.33	0.357	44.75	0.254

Swapping in AbsoluteDegradation training data improves every backbone - the gains are pipeline-driven, not model-specific.

Model comparison

SOTA model comparison - AbsoluteDegradation benchmark
Method	BRISQUE ↓	CLIPIQA+ ↑	MUSIQ ↑	MANIQA ↑
Input frames	47.86	0.248	26.84	0.162
RVRT	38.55	0.300	35.55	0.200
BasicVSR++	21.94	0.369	40.27	0.237
RTN	17.77	0.401	45.51	0.262
MambaOFR	30.11	0.349	39.80	0.233
DART (ours)	15.87	0.433	51.23	0.301

08 Results - Human study

People prefer it, too.

No-reference metrics can reward hallucinated detail, so we asked people instead. In a blinded two-alternative forced-choice study, raters compared restorations of our archival clips from the same architecture trained on the baseline (Bringing Old Films) pipeline versus AbsoluteDegradation - and picked ours in 64.7% of trials.

64.7% Trials preferring ours

93 Raters

1,395 Blind comparisons

Human preference

Preference for AbsoluteDegradation-trained output · dot = win rate, line = 95% CI

Per architecture

BasicVSR++General-purpose 89.9%213/237 · p <0.001
Shift-NetGeneral-purpose 71.5%163/228 · p <0.001
RVRTGeneral-purpose 59.8%137/229 · p = 0.004
DeepRemasterOld-film-specific 56.8%133/234 · p = 0.043
MambaOFROld-film-specific 55.8%n.s.129/231 · p = 0.087
RTNOld-film-specific 54.2%n.s.128/236 · p = 0.216

Pooled

General-purposeBasicVSR++, Shift-Net, RVRT 73.9%513/694 · p <0.001
Old-film-specificDeepRemaster, MambaOFR, RTN 55.6%390/701 · p = 0.003
All6 architectures 64.7%903/1395 · p <0.001

09 Achievements

From research question
to recognised results.

- Core deliverables -
SOTA model, pipeline & dataset

- Conference presentations
in Poland

- Scientific publications
in progress

10 Publications

Accepted.

'26

AbsoluteDegradation: A Physics-Inspired Synthetic Film Degradation Pipeline and Archival Film Restoration Benchmark

NeurIPS 2026 - Sydney Datasets & Benchmarks Track

Accepted

'26

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

ACCV 2026 - Osaka Main Track

Accepted

Cite this work BibTeX

@inproceedings{jastrzebski2026absolutedegradation,
  title     = {{AbsoluteDegradation}: A Physics-Inspired Synthetic Film-Degradation Pipeline and Archival Film Restoration Benchmark},
  author    = {Jastrz{\k{e}}bski, Miko{\l}aj and Glinkowski, Dawid and Zieli{\'n}ski, Dawid and Borkowski, Daniel and Koz{\l}owski, Wojciech and Adamczewski, Kamil},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026}
}

Absolute Cinema

A vanishing visual record,locked behind cost and complexity.

Costly & slow

Low watchability

Heritage decay

Out of reach

Three contributions,one end-to-end restoration system.

A new test set

AbsoluteDegradation

DART

A test set built forreal archival film.

No clean original exists -so we synthesise the damage.

No paired ground truth

Prior pipelines fall short

Authentic, physics-inspired damage

Ready forTraining

DART - Degradation-AwareRecurrent Transformer.

Dilation Pyramid MaskNet

AdaLN-Zero Severity Conditioning

Mask-Weighted Restoration Loss

Drag the handle.See the restoration.

State-of-the-art on everyperceptual metric.

People prefer it, too.

From research questionto recognised results.

Accepted.

AbsoluteDegradation: A Physics-Inspired Synthetic Film Degradation Pipeline and Archival Film Restoration Benchmark

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

Four researchers, two advisors.

A vanishing visual record,
locked behind cost and complexity.

Three contributions,
one end-to-end restoration system.

A test set built for
real archival film.

No clean original exists -
so we synthesise the damage.

Ready for
Training

DART - Degradation-Aware
Recurrent Transformer.

Drag the handle.
See the restoration.

State-of-the-art on every
perceptual metric.

From research question
to recognised results.