Skip to content

Upscaling a small photo for print (and knowing when it's actually big enough)

You have a small photo — say a phone screenshot or a cropped image — and you want to print it on A4, or just have it larger and sharper. Two separate questions come up, and it's worth keeping them apart:

  1. How do you enlarge it? (interpolation vs AI super-resolution)
  2. Is the result big enough for the paper size you want? (the DPI maths)

Two ways to make an image bigger — they are not the same

Interpolation (Lanczos, bicubic) just spreads the existing pixels over a larger grid and smooths between them. The picture gets bigger but no new detail appears — edges go soft and text stays as fuzzy as it was, only larger. It's instant and needs no model. Good enough when you only need a modest size bump.

AI super-resolution (Real-ESRGAN and friends) is different: a neural network reconstructs plausible fine detail — it has learned what sharp edges, text and texture tend to look like, and paints them back in. On a photo this genuinely recovers legibility: a signboard that was a blur becomes readable. The catch is that it is inventing detail it thinks was there, so it's reconstruction, not truth — fine for a photo, not something to trust for forensic detail.

Rule of thumb: interpolation for a small bump, AI when you need the detail back. Doing both and comparing side by side is the quickest way to see whether the AI pass was worth it for your image.

The DPI maths: will it print cleanly at A4?

This is the part people skip and then wonder why the print looks soft. Print sharpness is pixels ÷ physical inches = DPI:

  • A4 is 210 × 297 mm = 8.27 × 11.69 inches.
  • 300 DPI = crisp, magazine quality → needs 2480 × 3508 px.
  • 150 DPI = acceptable for a photo at arm's length → needs 1240 × 1754 px.
  • Below ~120 DPI you start to see softness/pixelation on paper.

So to check any image: divide its longest side by the paper's longest side in inches.

A 321 × 452 px photo printed at A4 → 452 / 11.69 ≈ 39 DPI. Nowhere near. Even doubled (642 × 904) it's only ~77 DPI — fine on screen, still too soft to print. A 4× pass (1284 × 1808) lands at ~155 DPI, which is good enough for a normal A4 photo print.

DPI metadata is not resolution. You can tag a file "300 DPI" and it changes nothing about the actual detail — it only tells the printer what physical size to lay the pixels out at. Real sharpness comes from the pixel count, full stop. Don't chase literal 300 DPI by upscaling 8×+ from a tiny source; past a point the AI is inventing more than it's recovering and it starts to look plastic.

The recipe (Python + uv, no system Python)

This uses spandrel to load the Real-ESRGAN weights — it sidesteps the old basicsr/realesrgan packages, which are a dependency headache on modern Python.

uv venv --python 3.11 .venv
uv pip install --python .venv/bin/python \
    pillow "numpy<2" torch torchvision spandrel

Grab a model checkpoint (x2 for a modest bump, x4 for print):

curl -sL https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth \
    -o RealESRGAN_x4plus.pth

import numpy as np, torch
from PIL import Image
from spandrel import ModelLoader

im = Image.open("input.jpg").convert("RGB")

model = ModelLoader().load_from_file("RealESRGAN_x4plus.pth").eval()
t = torch.from_numpy(np.array(im)).permute(2, 0, 1).float().div(255).unsqueeze(0)
with torch.no_grad():
    out = model(t).clamp(0, 1)
arr = (out.squeeze(0).permute(1, 2, 0).numpy() * 255).round().astype("uint8")

# dpi=(300, 300) only sets the print-layout size, not the detail — see above
Image.fromarray(arr).save("output_4x.png", dpi=(300, 300))

A 0.3 MP image runs in a few seconds on CPU — no GPU needed. For a plain interpolation comparison, im.resize((w*4, h*4), Image.LANCZOS) is a one-liner.

Gotchas worth remembering

  • RuntimeError: Numpy is not available — this is torch 2.2 built against NumPy 1.x while NumPy 2.x is installed. Pin numpy<2 and it goes away. Easy to lose an hour to if you don't recognise it.
  • You can't upscale a picture pasted into a chat. The tool needs the file on disk — a preview embedded in a conversation has no path to read pixels from. Point it at the real file.
  • AI upscaling can't recover what was never captured. It reconstructs plausible detail; it cannot bring back text or faces the source genuinely doesn't contain. Start from the best original you have, not a small re-save.
  • Check the size before printing, not after. Do the DPI division first — it tells you whether you need 2× or 4× before you spend time on the pass.

See also