CLARITY FIELD NOTES
AI upscaling vs traditional resizing
Both methods make a larger grid of pixels. The difference is how they choose the new pixel values.
Ordinary resize interpolates
A standard resize uses existing pixels to calculate nearby colors. It is fast, consistent and useful for layout, exact dimensions and downscaling. It cannot recover a texture that the source never captured. Enlarged lettering or hair often becomes smooth or soft.
AI estimates a plausible larger image
Real-ESRGAN is a trained super-resolution network. It learns relationships between degraded small images and larger training images, then estimates new detail from your source. Its output can sharpen contours or produce more coherent textures. It can also introduce errors, alter a face, smooth grain or distort lettering.
Clarity’s actual processing paths
| Mode / requested scale | Processing |
|---|---|
| AI, 1×–4× | Native 4× AI tile inference, resized to the requested output |
| AI, above 4× to 8× | 4× AI → downsample to 2× → second 4× AI → final dimensions |
| Standard resize | Canvas resampling only; no AI model |
The result panel reports the engine, model, pass count and successful inference tile count. It does not call a normal enlargement “AI”. Both available models run entirely on your device.
Which model should I try?
General x4v3 is a small model for general scenes and photos. AnimeVideo v3 is a compact model trained for animation imagery; it can be useful for flat colors and line art. Neither mode provides specialized face restoration or guarantees that tiny text will become readable. Compare both on a small crop when quality matters.
Check the result at 100%
Look at outlines, repeated patterns, fine lines, eyes and lettering. The comparison labels the original as resized because it must be displayed on the same canvas as the result. Retain the original and prefer standard resize when exact factual fidelity matters.
Model source: Real-ESRGAN project. Browser inference: ONNX Runtime Web.
Open the free image studio · Calculate print size