Resizing looks symmetric in every dialog box — type a number, get pixels. Underneath it is not symmetric at all: downscaling deletes information that exists; upscaling invents information that doesn’t. Everything about quality follows from that one asymmetry.

Downscaling: throwing away data is easy

Shrink 3840×2160 to 1920×1080 and each output pixel summarizes a 2×2 block of input. The resampler has more information than it needs and its whole job is picking what to keep:

• The result is mathematically cleaner than the original — noise averages out, aliasing disappears. This is supersampling, and it’s why 4K footage downscaled to 1080p looks better than native 1080p.
• Every resampling filter (box, bilinear, bicubic, Lanczos) is decent at downscaling; they differ mainly in how much ringing vs softness they leave. Bicubic-sharper and Lanczos are the usual picks for photos; area-average (“box”) is mathematically ideal for exact integer reductions.
• The only real trap is aliasing from nearest-neighbor — skipping pixels instead of averaging them, which produces moiré on fine patterns. Any modern editor defaults to something smarter.

Integer factors are the sweet spot. ×0.5 and ×0.25 downscales map cleanly onto pixel blocks; a ×0.37 downscale is fine too, just marginally less crisp than the integer cases.

Upscaling: the information has to come from somewhere

Grow 640×480 into 1920×1080 and each output pixel must be synthesized — 4.5 new pixels for every real one. No resampler creates detail; they only choose how to hide the invention:

• Nearest neighbor doesn’t hide it at all — each pixel becomes a block. Correct for pixel art, terrible for photos.
• Bilinear/bicubic interpolate smoothly — the honest “soft” look. Fine up to ~125–150%.
• Lanczos interpolates with a wider window — slightly sharper edges, slight ringing halos on high contrast.
• AI/deep-learning upscalers (Real-ESRGAN, waifu2x, the neural modes in editors) hallucinate plausible texture trained on millions of images. Convincing on common subjects; it guesses, and guesses can be wrong.

The quality ceiling is fixed by the source: upscaling never recovers what the original didn’t capture. A blurry 480p frame upscaled to 4K is a blurry 4K frame with extra pixels — larger file, same information.

Why 50% is not the mirror of 200%

The arithmetic makes the asymmetry concrete. Downscale 1920×1080 by 50% → 960×540: each output pixel condenses 4 real inputs. Upscale 960×540 by 200% → 1920×1080: each output pixel is made from ¼ of a real input. Same endpoints, opposite information flow.

This is also why “shrink for upload, enlarge for print” splits into different advice. Shrinking to a social spec is free quality insurance; enlarging to a print size is a budget — the question stops being “does it resize” and becomes “is the density enough,” which is pixel-per-inch math rather than pixel geometry (the sister tool PixelCalc is built exactly for that).

A practical decision list

• Downscale whenever you can. Exact spec sizes (the social presets table) or integer fractions of the source are the cleanest targets.
• Upscale only to a requirement. If nothing demands the bigger size, ship the smaller file — it’s sharper and lighter.
• Prefer integer factors for hard-edged content. Pixel art, game captures, terminal screenshots: 2×/3× nearest-neighbor keeps edges crisp; fractional scaling shimmers.
• Sharpen after resizing, never before. Halos and noise scale with the image and get baked in.
• Don’t re-resize a resized file. Always return to the largest master; compounding resamples compound artifacts.
• Check the “will it fit” verdict. If the source is smaller than the slot on both axes, you’re upscaling whether you noticed or not — the scaler flags it so the trade-off is explicit.

Frequently asked questions

Is a 2× upscale better than 1.5×?

Surprisingly, often yes — for pixel art, game captures and UI screenshots, an exact integer scale (2×, 3×) with nearest-neighbor keeps every edge hard and every pixel the same size. A non-integer upscale has to blend, producing uneven 'shimmer' edges. For photographs the difference matters less; a good bicubic or Lanczos resampler hides 1.5× fine.

Does downscaling to 50% really look sharper?

Yes — it's called supersampling. Going 2×-down averages each 2×2 pixel block into one, which discards noise and aliasing along with resolution. It's the same reason screenshots and thumbnails look crisper than the full frame: the downscale did free anti-aliasing.

Can AI upscalers actually add real detail?

They add plausible detail, not real detail — a learned guess at what texture probably occupied those pixels. On photographs of common subjects it's often convincing; on anything unusual it hallucinates (fabric becomes skin texture, text becomes squiggles). Great for presentation, worthless as evidence.

If I must upscale, how much is 'safe'?

Rules of thumb: up to ~125% is nearly invisible on photos with a decent resampler; ~150% shows softening on close inspection; beyond 200% only AI-assisted tools or very forgiving content stay presentable. If the target is print, the honest ceiling is whatever gives ~240 ppi at the final size — see the PixelCalc density calculator for that side of the math.

Why do my downscaled images sometimes look worse, not better?

Usually one of three causes: a nearest-neighbor resampler (fast but aliases diagonals — pick bicubic/Lanczos), sharpening before instead of after the resize (the sharpen halo gets scaled too), or double compression (downscale the already-compressed JPEG, save as JPEG again). Resize once, sharpen last, and keep the master lossless.