Upscaling is sold as recovery — enhance, sharpen, restore. Nothing recovers information that is not in the file. What upscalers actually do is choose between two honest failures: blur that admits what is missing, or invented detail that does not. Both are legitimate; using the second where you needed the first is how people end up presenting synthesised pixels as evidence.
The distinction that matters
Resampling (bicubic, Lanczos) interpolates between the pixels you have. Larger, softer, adds nothing. Honest.
Learned upscaling (super-resolution, diffusion) generates plausible new detail. Sharp, convincing, invented.
Neither recovers anything. Choose based on whether invented detail is acceptable for your use.
Traditional resampling
Interpolation asks: given the pixels I have, what value belongs at a position between them? Different algorithms answer differently.
| Method | Character | Use for |
|---|---|---|
| Nearest neighbour | Blocky, exact | Pixel art, hard edges |
| Bilinear | Soft, fast | Quick previews |
| Bicubic | Smooth | General enlargement |
| Lanczos | Sharpest, slight ringing | Photographic, downscaling |
Two practical notes. Lanczos is the best default for photographic content in either direction, and its one artefact — faint halos at high-contrast edges — is usually less objectionable than bicubic's softness. And nearest neighbour is not a worse option, it is a different one: for pixel art, anything that interpolates destroys exactly the crisp edges that define the style.
All of these are the right tool for downscaling, which is a genuinely lossless-in-spirit operation — you are discarding information you have, not inventing information you lack.
Learned upscalers
Super-resolution networks are trained on pairs of small and large images, learning what detail typically accompanies a given low-resolution pattern. Given a new small image, they generate the detail that would plausibly belong.
This works impressively well on content resembling the training distribution and fails in characteristic ways outside it. The tells are worth knowing:
- Texture where there should be smoothness. Skin gets pores, walls get stippling.
- Confident nonsense in small text. The upscaler renders sharp letterforms that spell nothing, because it learned what text texture looks like rather than what the words were.
- Uniformity. Every leaf gets the same invented detail, because the same pattern maps to the same output.
- Tile seams. Large images are processed in tiles, and boundaries can show.
🚨 Never upscale evidence
Learned upscaling of a licence plate, a face in CCTV, a document or anything else being used to establish a fact produces a sharp, confident image of something that was never photographed. The output is a statistically plausible reconstruction, not a clearer view.
This has real consequences — enhanced images have been treated as identification when what they show is the upscaler's training distribution. For any forensic, legal, medical or documentary purpose, use honest interpolation and accept the blur, or do not enlarge at all.
Face restoration deserves its own warning
Face restoration models are trained specifically on faces and are extremely effective at making a degraded face look good. They are also the clearest case of the invention problem, because faces are the thing humans are most sensitive to and most likely to over-trust.
The characteristic result: skin smoothed to poreless perfection, eyes made symmetrical and bright, distinguishing features softened toward the training mean. It looks better and it looks less like the person. For portraits of real people this is a meaningful misrepresentation; for a generated character it does not matter at all, which is why the same tool is fine in one context and not the other.
For AI images, regenerating usually wins
This is the section that saves the most time, because upscaling a generated image is frequently the wrong approach entirely.
Detail in AI images is missing because it was never represented in the latent space. A hand with two latent pixels does not have hidden detail an upscaler can bring out — there is nothing there. A separate network invents fingers; the original model, given more latent pixels, produces them properly.
| Approach | Detail source | Best for |
|---|---|---|
| Regenerate at higher resolution | The model itself | Coherent large images |
| Image-to-image at larger size | The model itself | Keeping the composition |
| Inpaint a region at full size | The model itself | Fixing hands and faces |
| Learned upscaler | A separate network | Final polish, 2× |
| Lanczos | Nothing — interpolation | When honesty matters |
The image-to-image route is the standard technique: take the generation, enlarge it with plain interpolation, then run it back through the model at the larger size with moderate denoising strength. The model now works with more latent pixels and adds detail consistent with the image it can see. Denoising strength is the control — too low and nothing improves, too high and the composition changes.
Note the constraint from the other direction: generating far above the model's training resolution causes duplicated subjects. The image-to-image pass avoids this because the existing composition guides it, which is precisely why it works better than generating large from scratch.
💡 A practical ladder
2× — reliable by any method. Learned upscaling here is genuinely useful and rarely obvious.
4× — invented detail becomes visible on inspection. Fine for art, not for anything documentary.
8× and beyond — the output is mostly synthesis. If you need this much, you needed to generate larger in the first place.
Order of operations
Sequence matters more than people expect:
Two rules underpin it. Fix before you enlarge — upscaling a mangled hand gives you a large, sharp, mangled hand, and any later fix has more pixels to work against. And compress last: every lossy save adds artefacts that the next processing step will amplify, so all editing should happen on lossless data with a single conversion at the end. That is the master-and-derivative workflow applied to a processing chain.
One more that catches people out: upscaling then downscaling is not a no-op. It is a slow, lossy way to sharpen, and if the final image is small, generating at or near that size and skipping the round trip gives a better result.
Resizing images for delivery?
Resize and crop images entirely in your browser — nothing is uploaded to a server.
Open Image Resizer →Summary
- No upscaler recovers detail. It interpolates honestly or invents convincingly.
- Lanczos is the best general resampler; nearest neighbour for pixel art.
- Learned upscalers generate plausible detail from training data, not from your image.
- Never enhance evidence. The output is a reconstruction, not a clearer view.
- Face restoration changes identity — it regenerates toward the average face.
- For AI images, regenerating beats upscaling: the detail was never in the latent.
- 2× reliable, 4× visible, 8× is synthesis.
- Fix before enlarging; compress last.
Frequently Asked Questions
Does upscaling actually add detail to an image?
No upscaler recovers information that was never captured. Traditional resampling interpolates between existing pixels, producing a larger but softer image. Learned upscalers generate new detail that is plausible for the content but invented — it looks sharp because it was synthesised, not because it was recovered.
What is the best way to upscale an AI-generated image?
Often not upscaling at all. Regenerating at a higher resolution, or running the image back through the model at higher size with moderate denoising, usually beats post-hoc upscaling because the model produces genuine detail rather than a separate network guessing at it.
Is Lanczos or bicubic better for resizing?
Lanczos generally produces sharper results for photographic content and is the better default for downscaling. Bicubic is smoother and slightly faster. For pixel art or hard-edged graphics, nearest neighbour is the correct choice because both of the others blur the edges you want preserved.
Why do faces change when I use face restoration?
Because the restoration model reconstructs a plausible face rather than recovering the original one. It draws on what faces look like in its training data, so it tends to produce a generic, smoothed version — recognisably a face, and often not recognisably the same person.
How much can I upscale before it stops being useful?
Roughly 2x is reliable, 4x is where invented detail becomes noticeable on close inspection, and beyond that the output is mostly synthesis. If you need a large final image, generating at a higher resolution from the start produces a better result than aggressive upscaling.