DISCO — Colorization AI Model
DISCO colorizes a black-and-white photograph by predicting a distribution of plausible colors rather than one flat guess, which is what keeps old portraits and archive shots from looking painted. Ten to thirty seconds.
Demo
Input — a grayscale image
Output — colorized by DISCO at its default settings
What it does
DISCO is an automatic image colorization framework that converts grayscale images into color RGB images.
It reads the content of a photograph and assigns plausible color to each region — skin, clothing, foliage, masonry — rather than applying a single tint. One run produces one interpretation; changing the seed produces a different one, and a fixed seed reproduces the same result exactly.
When to use it
- Automatic colorization of grayscale photographs
- Restoration and enhancement of old black and white images
- Alternative color interpretations of the same photograph by varying the seed
- Batch processing of large image collections
- Creative variations for design and content applications
Details
- CategoryColorization
- Year2022
- LicenseMIT
Compliance & Provenance
- ProviderOpen-source · Specialized
- EU AI Act RiskMinimal Risk
- Art. 50 TransparencyNot applicable
Inputs & Outputs
- ImageInputimage
Grayscale source image
- ImageOutputimage
Colorized output image
Parameters
- seedintegerdefault -1range -1–2147483647
Random seed (-1 = random)
Random seed for reproducibility. -1 (default) randomizes; fix to reproduce a previous result.
- versionstringdefault randoptions: rand
Model weights
Which DISCO weights to use. 'rand' is the only variant currently offered and is the default; it samples the color distribution per run, so re-running the same input can produce a different but equally plausible coloring. Fix the seed to reproduce one.
Limitations
- Photographs without recognizable subjects — macro detail, heavy bokeh, abstract texture — tend to come back as a single warm tint rather than color.
- Color is plausible, not recovered. The result is what the scene could have looked like, not what it did look like, so it should not be treated as evidence of the original colors.
Tags
Alternatives in Image Restoration
- Image Denoiser
Denoise photos (SwinIR). 5-15s.
- JPEG Quality Restoration
Remove JPEG compression artifacts (SwinIR). 5-15s.
- Motion Blur Removal (MSSNet)
Balanced motion-blur removal (MSSNet). 5-15s.
- Real-World Blur Removal
Diffusion-based deblur (highest quality). 1-3 min.