CODSR — Image Upscaler AI Model
CODSR upscales images 2x or 4x in a single diffusion step. It tags the source with a recognition model first and conditions the diffusion on those tags, and weights how much generative correction each region gets from its own gradient — aiming to add detail where there is structure without inventing it on flat areas.
What it does
CODSR turns a low-resolution image into a sharp 2x or 4x one in a single diffusion step rather than iterating a sampler.
It first runs a recognition model over the input to get content tags, then conditions the diffusion on them, so the detail it restores follows what the picture actually contains.
A per-region weight derived from image gradients decides how much generative correction each area receives, keeping flat regions closer to the source.
When to use it
- A low-resolution image from the web, brought up to a usable size
- A screenshot or video thumbnail, upscaled and clarified
- An old camera or phone photograph with age and degradation on it
- A file compressed to save space, with quality recovered
- Real-world degradation - blur, noise and compression at once
Details
- CategoryImage Upscaler
- Year2026
- LicenseMIT
Compliance & Provenance
- ProviderOpen-source · Specialized
- EU AI Act RiskMinimal Risk
- Art. 50 TransparencyNot applicable
Inputs & Outputs
- ImageInputimage
Low-resolution source image
- ImageOutputimage
Upscaled high-resolution image
Parameters
- scale_factorintegerdefault 2options: 2, 4
Upscale ratio (2/4)
Enlargement factor. 2 (default) — fastest, good for moderate upscaling. 4 — strongest enhancement but much slower and more VRAM-intensive.
Tags
Alternatives in Image Upscaling
- FiDeSR
One-step diffusion image upscaler (2x/4x), detail-preserving.
- PiSA-SR
Highest-quality image upscaler (2x/4x). ~50s.
- SMFANet+
Fast lightweight image upscaler (2x/3x/4x). 0.1-3s.
- Swin2SR
Balanced image upscaler (2x/4x) — Swin Transformer V2. 10-30s.
- SwinIR
Image upscaler with widest scale range (2x/3x/4x/8x). 10-30s.