2× Image Upscaling: AI vs. Lanczos on Two Test Images

We took two reference photos, degraded them on purpose, and upscaled them back with 2× AI and with a plain Lanczos resize. On both, the AI result sat closer to the original: PSNR +0.09 dB and +0.58 dB, SSIM +0.024 and +0.029.

That is a small experiment, not a general claim about upscalers. Below are the exact inputs, the machine, the scoring method, and what these numbers cannot tell you.

Key takeaways

• Two test images, one scale, one control. 2× only, AI versus OpenCV INTER_LANCZOS4 on identical inputs.
• The AI result scored higher on both images. astronaut PSNR 27.331 vs 27.240, SSIM 0.8546 vs 0.8301. coffee PSNR 27.088 vs 26.506, SSIM 0.7622 vs 0.7333.
• This measures fidelity to the original, not beauty. PSNR and SSIM reward similarity; they do not tell you whether the result looks natural.
• The test was run on the desktop component. Not the browser tool, and not a full app session.

What we actually measured

Two reference images were downscaled by half with INTER_AREA, then compressed to JPEG at quality 50. That input was upscaled 2× and compared with the untouched reference.

Test image	Input	Output	Reference
astronaut	256 × 256 JPEG Q50	512 × 512 PNG	512 × 512 original
coffee	300 × 200 JPEG Q50	600 × 400 PNG	600 × 400 original

• Machine: Windows 11 Enterprise 26100, Intel i9-14900HX (24 cores / 32 threads), 31.7 GiB RAM, RTX 4060 Laptop with 8188 MiB.
• Path: the Windows app's native 2× component through DirectML. The adapter in use was not logged, so the numbers are not an RTX 4060 benchmark.
• Method: each run started a fresh process and was timed from process start to exit, so the figures include model load and file I/O. Three runs per image per method; the quality scores come from the first run, whose output hashed identically across all three.
• Scoring: full-frame RGB PSNR (data_range=255) and SSIM (Gaussian window, sigma 1.5, population covariance), scikit-image 0.26.0. No edge cropping.

Why Lanczos is the right control

Most "AI upscaler" demos compare against a stretched image with no interpolation quality argument at all. Lanczos4 is a strong classical resampler, so beating it is a meaningful result — and beating a bilinear stretch is not.

It is also the honest control for the question a reader actually has: if I just resize in my editor, what do I lose?

The results

Test image	Method	PSNR (dB)	SSIM
astronaut	Lanczos 2×	27.240	0.8301
astronaut	AI 2× (Fast)	27.331	0.8546
coffee	Lanczos 2×	26.506	0.7333
coffee	AI 2× (Fast)	27.088	0.7622

The larger jump is on the coffee image, which has soft edges and limited texture — the case where interpolation has the least information to work with.

What these numbers do not tell you

• Two images are not a population. Both are well-lit reference photos with a single obvious subject.
• The input is synthetic. We degraded a good photo ourselves, so the test rewards recovering a known original. A real low-resolution photo has no such original to recover.
• PSNR and SSIM punish invention. A model that adds plausible detail is scored as an error if the original did not have that detail. A lower score is not automatically a worse-looking result.
• JPEG at quality 50 is one degradation, not all of them. Motion blur, sensor noise and heavy compression behave differently.
• This is the native component, not an app session. No GUI import, no export dialog, no queue.

When 2× is the right move — and when it is not

The limits decide this before quality does:

• Input ceiling. The online enhancer accepts up to 8 MP and 8192 px on a side; anything larger is refused rather than downscaled for you.
• Output ceiling. 32 MP caps the result. Because 4× produces sixteen times the pixels, 4× only fits a source of 2 MP or less — even a 1920 × 1080 frame is over that line.
• So 2× is the practical choice for anything from a phone or a 1080p frame, while 4× is for small crops and old low-resolution files.
• Do not upscale twice. Two 2× passes compound both the artefacts and the file size; start from the largest original you still have.

How to check your own result in a minute

1. Keep the original open next to the upscaled file at the same display size.
2. Look at a hard edge, then at a flat area. AI upscalers usually win the edge and can lose the flat area by adding texture that was never there.
3. Zoom to 100% on a face or text. If letters or eyelashes turn into wavy strokes, the model is inventing.
4. Compare file sizes at the same resolution. A large jump means added detail — not necessarily correct detail.

Bottom line

On two deliberately degraded test images, 2× AI upscaling stayed closer to the original than a Lanczos resize by a small but consistent margin, and the gain was larger on the softer image. That supports using it for 2× enlargements of ordinary photos; it does not establish that it beats every resampler, or that it improves a photo that was never downscaled in the first place.

The raw measurements, inputs, outputs and scoring script for this test are kept together so the numbers can be re-run rather than taken on trust. Other studies, and the reporting standard they all follow, are listed on the measurements page.