How do I calculate the size of an upscaled image?

Multiply both the width and height by the upscale factor. A 512×768 image at 4× becomes 2048×3072. The total pixel count grows by the factor squared, so 4× upscaling produces 16 times as many pixels.

Why does 4× upscaling make the file so much larger?

Upscaling multiplies each dimension, so the pixel count scales by the factor squared. A 4× upscale has 16× the pixels of the original, and file size grows roughly in proportion, which is why a small image can balloon into tens of megabytes.

What is the maximum size Real-ESRGAN can produce?

Real-ESRGAN processes images in tiles, so it can technically go very large, but in practice keeping the long edge under about 8000 pixels avoids tiling seams and VRAM pressure. Topaz Gigapixel handles much larger outputs.

Should I upscale in one pass or several?

For very high factors like 8×, two chained 2× or 4× passes often produce cleaner results and stay within each model's comfortable range, compared to a single aggressive pass that can introduce artifacts.

Is the file-size estimate exact?

No. It is a rough estimate assuming photographic PNG content at about 1.2 bytes per pixel after compression. Flat or simple images compress smaller; highly detailed ones can be larger.

AI Upscaling Size Calculator

Work out your upscaled image size before you run it

AI upscalers like Real-ESRGAN, Topaz Gigapixel and SwinIR multiply every dimension of your image, and the total pixel count grows by the square of the factor. That means a modest 512×768 generation becomes a 2048×3072, 6.3-megapixel file at 4× — and a 50-megapixel monster at 8×. This calculator shows you the exact output dimensions, megapixels and a file-size estimate before you spend GPU time, plus whether your target exceeds the model’s practical ceiling.

How upscaling math works

Upscaling is dimensional, not area-based. If your source is W × H and you upscale by factor f, the output is (W × f) × (H × f). The pixel count — and roughly the file size — scales by f²:

output_width  = source_width  × factor
output_height = source_height × factor
megapixels    = (output_width × output_height) / 1,000,000

So 2× quadruples the pixels, 4× multiplies them by 16, and 8× by 64. Most upscalers also have a practical long-edge limit driven by VRAM and tiling behaviour, beyond which you get seams or out-of-memory errors.

Tips for clean upscales

Chain passes for big jumps. For 8×, run two 4× or 2× passes rather than one aggressive pass — it usually yields cleaner detail and stays inside each model’s comfortable range.
Watch the long edge, not just megapixels. Tiling artifacts appear when the longest dimension is large, so the calculator flags when you cross the model’s recommended ceiling.
Pick the model for the job. Real-ESRGAN and 4x-UltraSharp are great general-purpose open-source options; Topaz Gigapixel handles the largest outputs; SwinIR preserves fine texture but is heavier on VRAM.