🤖 machine learning

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

This paper introduces AbsoluteDegradation, a physics-inspired synthetic pipeline and a large-scale real-world benchmark designed to overcome the lack of paired training data and standardized evaluation in archival film restoration, demonstrating that models trained with this approach achieve superior generalization to real footage.

Original authors: Mikołaj Jastrzębski, Dawid Glinkowski, Dawid Zieliński, Daniel Borkowski, Wojciech Kozłowski, Kamil Adamczewski

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mikołaj Jastrzębski, Dawid Glinkowski, Dawid Zieliński, Daniel Borkowski, Wojciech Kozłowski, Kamil Adamczewski

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a dusty, scratched, and fading old home movie from the 1920s. You want to restore it so your family can see it clearly again. The problem is that the original, perfect version of that movie is gone forever; it rotted away with time. You can't just compare the bad version to a "perfect" one to teach a computer how to fix it, because the perfect one doesn't exist.

This paper, titled "AbsoluteDegradation," solves this problem by building a virtual time machine and a new testing ground for computers trying to fix these old films.

Here is how they did it, explained simply:

1. The Problem: The "Fake" vs. The "Real"

To teach a computer to fix a broken film, you usually need to show it a "broken" version and the "perfect" version side-by-side. Since the perfect version is lost, scientists have been making fake broken videos to train the computers.

The Old Way: Previous methods were like a child drawing on a photo. They would just randomly add scratches, dust, and blur to a clean picture. But real film damage is more complex. It's not just random; it happens in a specific order (like how film is exposed, developed, and then scanned), and the damage moves in a specific way over time (like a scratch that stays in the same spot for ten seconds). The old fake videos didn't feel "real" enough, so the computers learned the wrong lessons.
The New Way (AbsoluteDegradation): The authors built a physics-based simulator. Instead of just throwing random dirt on the screen, they simulated the actual chemical and mechanical process of how film gets damaged.
- The "Gate Weave": Old projectors and scanners wobble slightly. The authors programmed the computer to mimic this specific, rhythmic shaking (using a mathematical concept called an "Ornstein-Uhlenbeck process," which is just a fancy way of saying "wobbly but returning to the center").
- The "Grain": Real film grain isn't just static noise; it changes based on how bright the image is. They modeled this specific behavior.
- The "Scratches": Real scratches don't just appear and vanish instantly; they last for a while and move smoothly. Their system creates scratches that behave exactly like real physical damage.

The Analogy: Imagine trying to teach someone to fix a car engine.

Old Method: You give them a toy car and tell them to "break it" by throwing sand at it. They learn to fix sand, but not a real engine.
New Method: You give them a real engine, take it apart, and simulate exactly how rust, heat, and friction break it down over 50 years. Now, when they try to fix a real car, they actually know what they are doing.

2. The New "Exam" (The Benchmark)

Even with better training, scientists needed a better way to test if the computers were actually doing a good job.

The Problem: Existing test datasets were small, low-quality, or mixed with cartoons (which look different from real film). Also, the "grades" (metrics) used to score the restoration were tricky. Sometimes, a computer would make the image look sharper but also fake (adding details that weren't there), and the grading system would give it a high score.
The Solution: The authors created a massive new dataset called the AbsoluteDegradation Benchmark.
- They gathered 81,576 high-resolution frames from real, public-domain films from the Library of Congress (dating from 1896–1918).
- These are "in the wild" films with real damage, no watermarks, and no cartoons.
- They proved that current "grading systems" often trick computers into making fake-looking images just to get a high score. Their new benchmark forces the computers to be honest and preserve the original history, even if it looks a bit grainy.

3. The Results

When they tested their new training method (AbsoluteDegradation) against the old methods:

Better Generalization: Computers trained on their "physics-based" fake videos were much better at fixing real old movies than computers trained on the old, simple fake videos.
Avoiding "Hallucinations": The old methods often tried to "fix" things by inventing new details (like sharpening a scratch until it looked like a line in the drawing). The new method learned to be careful, keeping the original texture and history intact without adding fake details.
The "Over-Smoothing" Trap: The paper found that if you just use simple "white noise" (like static on an old TV) to train the computer, the computer learns to smooth everything out, turning a person's face into a blurry blob. By using their complex, realistic noise, the computer learned to keep the fine details (like the folds in a hat or the texture of a face).

Summary

The paper introduces two main things:

A Physics-Inspired Simulator: A way to create fake damaged videos that are so realistic (mimicking real chemical and mechanical processes) that computers trained on them can actually fix real old movies.
A Real-World Test: A massive, high-quality dataset of real old films to prove that the computers are actually doing a good job, rather than just tricking the scoring system.

The ultimate goal isn't just to make the video look "clean," but to restore it authentically, preserving the true history of the footage without inventing fake details or smoothing away the character of the original film.

Technical Summary: AbsoluteDegradation

Problem Statement
The restoration of archival film is hindered by two fundamental challenges. First, the absence of paired training data (pristine ground truth vs. deteriorated footage) forces supervised learning to rely on synthetic approximations. Existing degradation models often treat artifacts as independent image-domain perturbations, failing to capture the complex, temporally coherent, and physics-grounded nature of real film decay (e.g., signal-dependent grain, gate-weave jitter, and exposure fluctuations). Second, there is a lack of high-quality, standardized evaluation benchmarks. Existing datasets are often limited in scale, mix animation with archival footage, suffer from low resolution, or contain non-analog artifacts (like watermarks), making reliable comparison and evaluation difficult.

Methodology: The AbsoluteDegradation Pipeline
The authors propose AbsoluteDegradation, a modular, physics-inspired synthesis pipeline that transforms clean video clips into realistically degraded counterparts. The pipeline models the analog-to-digital process as a structured composition of artifact families, governed by three principles: physics-inspired modeling, temporal coherence, and compositional diversity.

The pipeline operates through six ordered stages:

Frame Crop with Motion Padding: Creates a buffer for subsequent geometric transformations, ensuring spatial alignment with the ground truth.
Dust and Texture Overlays: Simulates emulsion staining and projection marks using a multi-slot persistence model where contaminants have random lifetimes, encouraging networks to utilize temporal information.
Negative-Film Emulation: Models the chemical reality of analog recording by applying degradations (blur, grain, downsampling, gamma) in the inverted domain (simulating the raw negative) before inverting back.
Permuted Degradation Core: Composes seven operators—blur, grain, JPEG, resolution change, gamma, and horizontal/vertical gate-weave—in a random order ( $7! = 5,040$ $7! = 5, 040$ permutations) per clip.
- Signal-Dependent Grain: Replaces standard white Gaussian noise with a Signal-Dependent Noise (SDN) formulation using a blue-noise prior to mimic the structured, high-frequency spectral behavior of silver halide film grain.
- Camera Gate Weave: Models mechanical film transport inaccuracies using Ornstein-Uhlenbeck (OU) processes. Horizontal motion uses a first-order OU process, while vertical motion employs a second-order process to model time-varying mean-reversion strength.
Finalization: Restores spatial resolution, applies mild photometric jitter, and quantizes to 8-bit to match real digitization precision.
Mechanical Scratches: Renders persistent vertical streaks using a sinc²–Gaussian hybrid kernel to approximate Fraunhofer diffraction, with intensity flickering and longitudinal width variation.

Key Contributions

AbsoluteDegradation Synthesis Engine: A framework that captures complex temporal and spatial dependencies of authentic film decay, including signal-dependent grain, parametric scratches, and temporally coherent camera motion. It addresses limitations of prior works by modeling the layered processing structure of analog-to-digital pipelines and using physically grounded stochastic processes.
Large-Scale Real-World Archival Benchmark: A curated dataset of 81,576 high-resolution, lossless frames sourced from 30 public-domain films (1896–1918) from the Library of Congress. Unlike previous benchmarks, it focuses exclusively on real-world analog degradations, avoids mixed content (e.g., animation), and provides rigorous conditions for zero-shot evaluation.

Results
Extensive experiments across four architectures (RTN, MambaOFR, BasicVSR++, RVRT) on both the proposed benchmark and the SRWOV dataset demonstrate:

Improved Generalization: Models trained with AbsoluteDegradation generalize better to real archival footage than those trained on the "Bringing Old Films" baseline.
Benchmark Insights: The new benchmark reveals systematic failure modes in current methods. Standard no-reference metrics (e.g., BRISQUE) were found to be unreliable, sometimes rewarding "historically inauthentic hallucinations" and over-smoothing.
Qualitative Superiority: The proposed pipeline guides networks to produce authentic, artifact-free restorations. In contrast, baselines relying on Gaussian noise tend to over-smooth fine textures and artificially sharpen artifacts (like scratches), creating flickering edges. The signal-dependent blue noise approach preserves structural integrity and fine details (e.g., fabric folds, star emblems) that Gaussian-based models erase.
Ablation Studies: Removing temporal coherence or relying solely on Gaussian noise led to higher metric scores but visibly degraded visual quality, confirming that current metrics can be misaligned with true restoration goals.

Significance
The paper claims to establish a foundation for reproducible and domain-authentic evaluation in archival film restoration. By bridging the gap between synthetic training data and real-world physics, AbsoluteDegradation enables the training of models that prioritize historical integrity over artificial sharpness. The work highlights that misaligned evaluation proxies have partially obscured real progress in the field and argues for the necessity of physics-inspired data synthesis and representative benchmarks to achieve scalable, high-fidelity preservation of audiovisual heritage.

1. The Problem: The "Fake" vs. The "Real"

2. The New "Exam" (The Benchmark)

3. The Results

Summary

Technical Summary: AbsoluteDegradation

More like this