Scale Factor

Also known as: scale, scale_factor

Scale factor controls the upscaling multiplier for super-resolution models. A scale of 2 doubles the image dimensions, 4 quadruples them. Higher values produce larger images but take more time and can introduce artifacts if pushed too far.

What It Does

Super-resolution models like Real-ESRGAN take a low-resolution input image and upscale it to a higher resolution while adding realistic detail. The scale factor determines the multiplier applied to both width and height. A 512×512 image at 2× becomes 1024×1024; at 4× it becomes 2048×2048.

The model doesn’t simply stretch pixels — it intelligently fills in detail that wasn’t present in the original, using patterns learned from training data. Higher scale factors require more computation and more “hallucinated” detail. For most use cases, 2× or 4× is sufficient. Pushing beyond 4× often introduces visible artifacts or overly smoothed textures.

Adjusting It in CSF

Scale factor belongs to upscaling models, so it shows up in the Advanced Model Settings modal only when the selected model exposes a scale parameter. Edit a production design, open its Image tab, click Advanced Settings, and choose the model to configure.

Depending on how the model defines it, scale factor renders either as a slider between its minimum and maximum or as a row of buttons for a fixed set of multipliers (such as 2, 3, and 4). Save keeps your choice for this design and model when it differs from the default; Reset to Defaults restores it.

Common Scale Factors

2× Scale

Doubles dimensions (4× pixel count). Fast and reliable. Best for images that are slightly undersized.

3× Scale

Triples dimensions (9× pixel count). Good middle ground for moderate upscaling.

4× Scale

Quadruples dimensions (16× pixel count). Maximum commonly supported. Good for significant upscaling from small originals.

Tips

2× is usually sufficient for web display and general use.
4× is ideal when you need print-resolution output from a generated image.
Upscaling an already high-resolution image wastes time and credits with minimal benefit.
For best results, generate your image at the model’s native resolution first, then upscale as a separate step.