How to Resize an Image Without Losing Quality
Resizing smaller is easy and lossless-looking. Resizing larger is where the trouble starts. Here's the difference between resize, resample, and upscale — and which one you actually need.
"Resize without losing quality" is usually two different questions:
- Make the image smaller while keeping it sharp → easy, well-solved, every tool does it correctly.
- Make the image larger without it looking blurry → hard, fundamentally limited, requires either AI upscaling or a tolerance for soft edges.
This guide breaks down the difference and points you at the right tool for each case.
The fundamental constraint
A digital image is a grid of pixels. When you resize down, you're throwing away pixels — the algorithm just decides which to keep and how to blend the discards into the survivors. Modern resize algorithms (Lanczos, bicubic) do this well; the result is visually indistinguishable from a native-resolution capture at the target size.
When you resize up, you're inventing pixels that weren't in the source. The naive approach (nearest-neighbour) duplicates pixels — the result is pixelated and visibly worse. Smarter approaches (bicubic, Lanczos) interpolate — they smooth the gaps, which avoids pixelation but introduces blur. AI upscalers (ESRGAN, SwinIR, real-time models) hallucinate plausible texture into the gaps — sharper results, but the output is invented, not recovered.
There is no algorithm that recovers information not present in the source. "Lossless upscale" is marketing.
When to use which tool
| Goal | Tool | Method |
|---|---|---|
| Shrink a 4000×3000 photo to 1920×1080 for the web | Resize Image | Lanczos / bicubic downsample |
| Same photo, but compress for smaller file too | Compress Image | Resize + quality slider |
| Enlarge a 500×500 logo to 2000×2000 | Image Upscaler | Real-ESRGAN AI upscale |
| Enlarge a vector / icon (SVG) | Re-export at the higher size from the SVG source | Re-rasterise from the vector |
The third row is the interesting one. AI upscale (Real-ESRGAN with the WebGPU pipeline Dropvert uses) gets you genuinely-sharp output at 2× or 4× the source resolution. The result isn't "the same as if you'd captured it at higher resolution" — but it's much closer than naive bicubic.
Shrinking: the easy case
If your goal is "downsize this photo from a 4000×3000 phone capture to 1080p for Instagram", any modern tool produces visually identical output. Resize Image has a width / height input, a target-resolution picker, and an aspect-ratio lock. Drop the file, set the target, download.
For best results when shrinking by more than 2×, save as JPEG quality 85+ or WebP. PNG is fine for screenshots; for photos it bloats the output unnecessarily.
Enlarging: the AI case
Image Upscaler ships with Real-ESRGAN running in your browser via WebGPU (with a WASM fallback). It's roughly the same engine as paid services like Topaz Gigapixel or Let's Enhance — same model family, no monthly subscription, your image never leaves your device.
Practical scale factors:
- 2× upscale: usually indistinguishable from a sharp native capture for most subjects.
- 4× upscale: visible "AI signature" on detailed textures (faces, fabric, foliage), but dramatically sharper than the source. Good enough for most use cases.
- >4×: hallucinations start to dominate. The model invents plausible textures that may or may not match what was actually in the source. Use sparingly.
A 500×500 source upscaled to 2000×2000 (4×) takes 10-30 seconds on a modern laptop with WebGPU; longer on the WASM fallback path.
Common gotchas
- JPEG-then-resize loses more quality than resize-then-JPEG. If you have the source in a lossless format, do all your resizing first, then save the final result as JPEG.
- Sharpening after a downsample is normal practice. Most photo editors apply a light sharpen automatically after a resize-down; standalone resize tools usually don't. If the output looks slightly soft, that's why.
- Resizing a screenshot of text is different from resizing a photo. For text, nearest-neighbour at integer scale factors (2×, 3×) is sharper than bicubic. Most general-purpose resize tools always use bicubic and the text gets blurry. For screenshots of text specifically, capture at the final size if you can.
FAQ
Why do my upscaled photos look "plastic"? Real-ESRGAN's default model favours sharpness over fidelity on faces and skin. If you find this distracting, scale to 2× rather than 4× — the artefacts are subtler.
Is there any way to truly upscale without artefacts? No, for the reasons above. The only "true" upscale is re-capture or re-render at the higher resolution from the source — which is only possible if you have a vector or 3D source.
Does the resize tool support batch? Yes. Drop multiple images at once on Resize Image; each gets resized with the same settings and the results download as a zip.
Tools mentioned in this guide
Related guides
What Is WebP and Should You Use It?
WebP is a modern image format from Google that produces smaller files than JPEG and PNG with comparable quality. Here's when to use it, when not to, and how to convert.
PNG vs JPEG: Which Image Format Should You Use?
PNG is lossless and supports transparency. JPEG is smaller for photographs. The right choice depends on the image content — here's how to decide.
How to Compress a GIF for Discord, Slack, and Email
GIFs are huge. Here's how to compress one to the size limits Discord (10 MB free / 500 MB Nitro), Slack (1 GB), and email (typically 25 MB) actually allow.