🔧

Image Restoration - Image Colorization

🛠

Automatic Image Colorization

Transforms grayscale images into naturalistic color. Supports deterministic colorization and diverse multi-variant generation (up to 3 color interpretations per image).

What it does

DISCO is an automatic image colorization framework that converts grayscale images into color RGB images.

Problem it solves

  • Automatic colorization of grayscale photographs
  • Restoration and enhancement of old black and white images
  • Diverse colorization generation (multiple color interpretations for same image)
  • Batch processing of large image collections
  • Creative variations for design/content applications

Input/Output

Input:
Input: Grayscale image (or color image that will be converted to grayscale)
Output:
Output: Colorized RGB image
image
image
icon

Model Version:

DISCO-c0_2
DISCO-rand
Color Saturation: Less aggressive → more muted, subtle colors
Anchor Robustness: Higher robustness to varying anchor site locations
Best for: Conservative colorization where you want natural, less vivid colors
Advantage: More stable performance across different image types and compositions
Calibration: The "c0.2" likely refers to a color saturation coefficient of 0.2
Best for: Varied datasets where image structure differs significantly
Flexibility: Works better when anchor locations (used for color distribution) vary

Technical Details

  • Architecture: Anchor-based disentangled colorization
  • Core Components:
    • Anchor Color Representation Module
      • Predicts global color anchors for the image
      • Outputs: Anchor locations (spatial coordinates) and anchor colors
      • Purpose: Disentangles color multimodality from structural details
    • Anchor-Guided Color Generation Module
      • Synthesizes per-pixel colors using predicted anchors
      • Uses anchors as reference for color generation
      • Purpose: Ensures structural consistency while enabling diverse colorization
  • Framework & Dependencies:
    • Framework: PyTorch (v1.8.0+)
    • GPU requirement: CUDA 10.2+
    • Parameters: Multi-component architecture (exact count not specified in original source)
  • Training:
    • Original training details not specified in model source
    • Intended for generative colorization (not discriminative classification)

Compliance & Provenance

Provider
Open-source
Provider type
Specialized
License
MIT
EU AI Act risk class
Minimal Risk
Art. 50 transparency
Not applicable
Region availability
Available globally
Training data summary
Pending — provider has not yet published per Art. 53(d)

For more on how we classify models and mark outputs, see our AI Policy.

Model Source