Archivist & Project Degrader

A comprehensive suite for cel animation restoration: specialized AI models and the physics-based degradation simulator used to train them.

---

📚 Table of Contents

• Archivist Models (Download & Usage)
• Project Degrader (Dataset Generation Tool)

---

🎞️ Archivist Models

Archivist is a set of Real-ESRGAN Compact (48nf16nc) models trained to handle specific defects found in old cel animation (1940s-1980s), such as Metrocolor degradation, film tears, chemical stains, and emulsion shifts.

Unlike general-purpose denoisers, these models were trained on a physically-simulated degradation pipeline (see Project Degrader below), allowing them to distinguish between intended line art and physical damage.

⚙️ Technical Specs

• Architecture: Real-ESRGAN Compact (SRVGGNetCompact)
• Config: 48 filters, 16 blocks (48nf16nc)
• Scale: 1x (Restoration/Denoising)

🧩 Model Zoo

Model (Click to Download)	Iterations	Role & Best Use Case	Comparison
AntiLines	457k	The Cleaner. Specifically targets horizontal lines, film tears, and scratches that cut through the frame.	View on ImgSLI
Rough	493k	The Rescuer. For heavily damaged footage. Hallucinates lost details.	View on ImgSLI
Medium	478k	The Workhorse. Balanced removal of grain and dirt while preserving original texture. The best starting point.	View on ImgSLI
Soft	453k	The Artist. Gentle restoration. Keeps film grain aesthetic. ⚠️ Note: In some scenarios, standard DRUNet might yield subjectively better results. Always compare.	View on ImgSLI
RGB	193k	The Specialist. Targets heavy chromatic noise and color channel degradation. Note: overlaps partially with Rough.	View on ImgSLI

Legacy Models: Older versions (BW/RGB Denoise Compact) are available in the Archived_2024 folder.

🛠 Recommended Workflow

For "Hollywood-grade" results, use a Two-Stage Pipeline. Archivist models restore the structure, while a mathematical denoiser stabilizes the result.

1. Stage 1 (Restoration): Process with Archivist to remove physical defects (scratches, lines, stains).
2. Stage 2 (Stabilization): Process the result with DRUNet (low strength). This removes residual mathematical noise and stabilizes the video temporally.

🚀 Usage (REAL-Video-Enhancer)

The easiest way to use these models is via REAL-Video-Enhancer, which supports TensorRT optimization and the DRUNet pipeline.

1. Download the .pth files from this repository.
2. In RVE, click "Add Model" and select the .pth file (it will convert to TensorRT automatically).
3. Select the Archivist model as the main upscaler (1x).
4. Enable Denoise (DRUNet) in the settings for stabilization.

---

🧪 Project Degrader Software

Located in the Degrader/ folder, this is the GUI application written in Python (PyQt6) used to generate the training dataset for Archivist.

Standard noise generation (Gaussian/Poisson) is insufficient for training restoration models for old films. Project Degrader simulates the physics and chemistry of film aging.

🌟 Key Features

• Physics-Based Simulation:
  • Geometry: Simulates film creases, warping, and emulsion shifts (chromatic aberration).
  • Defects: "Smart" scratches (Bezier curves/hairs), debris, and dust.
  • Chemistry: Simulates uneven emulsion degradation, chemical stains, and color fading.
• Digital Artifacts: Simulation of Banding (quantization) and MPEG compression.
• Advanced GUI:
  • Comparison Viewer: Real-time preview with a "Magnifier" tool and split-screen.
  • Profile Manager: Save and load complex degradation presets (JSON).
  • Batch Processing: Multi-threaded generation of LQ/GT pairs with probability distribution for different profiles.

🖥️ Installation & Running

The application is located in the Degrader directory.

Prerequisites: Python 3.10+, FFmpeg (for MPEG simulation).

Linux / macOS

Use the included launcher to automatically set up the virtual environment:

cd Degrader
chmod +x launcher.sh
./launcher.sh run