FreeVector

Guide · Design workflow

How to Upscale Images Without Losing Quality

Upscaling an image means increasing its pixel dimensions, and no method can recover detail that the original file never captured. What a good method can do is enlarge the image with smooth edges and plausible detail. This guide explains how interpolation and AI upscaling differ, how to calculate the pixels you need for screen or print, and how to check the result so a larger file is also a better one.

Photo editor comparing an enlarged photograph with the original on a desktop monitor

Key takeaways

  • Upscaling adds pixels, not information: classic resizing guesses by averaging neighbors, and AI upscaling guesses by drawing on training examples.
  • Work out the target size first: pixels equal print size in inches times pixels per inch.
  • Use interpolation for simple graphics and AI upscaling for photos, then inspect the result at 100% zoom.
  • AI upscalers can invent details that were never in the photo, so avoid them for evidence, documents and faces that must stay accurate.
  • Start from the best source file you have, because every lossy re-save removes data that cannot be recovered.

What does it mean to upscale an image?

Upscaling is making an image larger by creating new pixels between and around the existing ones. The original file has a fixed number of pixels, and software has to decide what color each added pixel should be.

An image's pixel resolution is its width times its height. Wikipedia's example is a 2048 by 1536 image, which totals 3,145,728 pixels, or 3.1 megapixels, and it says pixel counts serve as upper bounds on image resolution (Wikipedia, Image resolution).

That upper bound explains why upscaling never restores a blurry photo to the sharpness of a high-resolution shot. It changes the size of the file, and the quality depends on how well the new pixels are guessed.

Why does a plain resize look soft or blocky?

Classic resizing fills new pixels by interpolation, and each method trades sharpness against smoothness. The choice of method decides whether an enlarged image looks blocky, soft or crisp.

Wikipedia's article on image scaling describes the common methods. Nearest-neighbor replaces each output pixel with the nearest input pixel, so enlarged areas show repeated blocks of one color, which suits images with few smooth edges, such as pixel art.

Bilinear interpolation creates smooth transitions even where the original had hard boundaries, which suits continuous-tone images but softens sharp edges. Bicubic gives substantially better results than bilinear at higher computational cost, and Lanczos, an approximation of the sinc method, yields better results than bicubic (Wikipedia, Image scaling).

  • Nearest-neighbor: keeps hard edges, shows blocks. Use for pixel art and screenshots of interface elements.
  • Bilinear: smooth but soft. Fast and fine for small enlargements of photos.
  • Bicubic: a solid default in many editors for photos.
  • Lanczos: often the sharpest of the classic options for photos.
  • Vector conversion: not interpolation at all, but the right answer for logos and line art (see below).

How does AI upscaling work, and what can it get wrong?

AI upscalers use a neural network trained on example images to predict what the enlarged version should look like. Instead of averaging neighbors, the model draws on patterns it learned, such as the texture of skin, fabric, foliage or text strokes.

Wikipedia says machine learning methods can add detail and sharpness that is not present in the original, which can introduce errors. It quotes experts saying such tools inevitably have to invent, or at least recreate, details that were or were not there (Wikipedia, Image scaling).

The article on super-resolution imaging is blunter: there is no guarantee that the upscaled features actually exist in the original image. For that reason it calls deep convolutional upscalers unsuitable for ambiguous inputs where the presence or absence of a single feature is critical (Wikipedia, Super-resolution imaging).

  • Good fit: product photos, portraits for web use, scanned prints, old family pictures, game and illustration art.
  • Use with care: faces of real people (features can shift), text in photos (letters can change), and low-resolution crops.
  • Do not use: forensic images, medical images, evidence or any file where invented detail could mislead.
  • Check for artifacts: waxy skin, repeated textures, warped letters, extra fingers, over-sharpened edges.

Which free and open tools exist for AI upscaling?

Open-source models such as Real-ESRGAN show what AI upscaling does under the hood, and commercial apps wrap similar ideas in an easier interface. Real-ESRGAN describes its aim as developing practical algorithms for general image and video restoration, and it is trained with purely synthetic data (GitHub, xinntao/Real-ESRGAN).

Its repository lists models for general photos, anime images and anime video, and an option to set arbitrary output sizes after the model runs. It is released under the BSD-3-Clause license, and the project's portable executables run without CUDA or PyTorch.

For most people a web or desktop app is simpler than running a model by hand. Our ranking of AI image upscalers compares the products on output quality, batch options, file limits and pricing, and our list of AI photo editors covers apps that bundle upscaling with other retouching.

How many pixels do you actually need?

Multiply the print size in inches by the pixels per inch you want, and that gives the pixel dimensions you need. For screens, match the largest display size the image will fill, and for print, work from the print size.

Wikipedia's resolution article shows the math with its 2048 by 1536 example: about 28.5 inches wide at 72 pixels per inch, which it describes as very low quality, but only about 7 inches wide at 300 pixels per inch, which it describes as very good quality. It also lists A-series paper sizes at 300 pixels per inch: A5 is 1748 by 2480, A4 is 2480 by 3508 and A3 is 3508 by 4961.

The article presents 300 pixels per inch through examples and does not state it as a universal standard. Ask your printer for the exact requirement, because large posters viewed from a distance often need less.

  • Step 1: write down the final size (inches or centimeters) or the display size in pixels.
  • Step 2: choose the pixels per inch; use your printer's number, or 300 as a common starting point for close viewing.
  • Step 3: multiply width and height by that number to get the target pixel dimensions.
  • Step 4: compare with your file. If you already have enough pixels, you do not need to upscale at all.
  • Step 5: if you fall short, decide how much of the gap interpolation can cover, and how much needs AI or a new source.

What is a step-by-step workflow for upscaling a photo?

Prepare the source, upscale once, then sharpen lightly and inspect at full size. Doing the steps in this order avoids stacking artifacts on top of each other.

  1. Find the best original: the camera file, the design file or the largest download you can legitimately get.
  2. Crop and rotate first, so you do not pay for pixels you will throw away.
  3. Fix obvious problems such as noise or exposure with a tool from the photo editing software category, before enlarging.
  4. Pick the method: interpolation for graphics and small enlargements, AI upscaling for photos and larger jumps.
  5. Set the target size from your pixel calculation and upscale.
  6. Zoom to 100% and inspect faces, edges, text and textures against the original.
  7. Apply light sharpening only if the result is still soft, and export once.

Keep the original and the upscaled file side by side, with the scale factor in the file name. You can then redo the job with a different tool without starting over.

Why does the source file matter so much?

Every lossy re-save permanently removes data, so start from the least compressed version you have. Wikipedia's article on lossy compression says repeatedly compressing and decompressing a file causes it to progressively lose quality, and that the data already lost cannot be recovered (Wikipedia, Lossy compression).

The same article notes that some operations such as crop, rotate and flip can avoid re-encoding with tools like jpegtran, but other edits, including resizing, generally require decoding and re-encoding, which adds generation loss. Its practical advice is to keep a lossless original and produce JPEG copies from it.

For upscaling this means: avoid taking a screenshot of an image, avoid saving the same JPEG several times between edits, and export the final upscaled file once, in a format that suits its use.

  • Prefer original camera files, PNG or TIFF over a JPEG that has been shared and re-saved.
  • Do not upscale a thumbnail if the full-size file exists somewhere.
  • Keep edits in a layered or project file, and export the final image once.
  • Save a lossless master next to every web export.

When is upscaling the wrong fix?

For logos, icons and line art, converting to vector is better than upscaling. A vector file stores shapes rather than pixels, so it can be scaled to any size without blur. Our guide to vector vs raster images explains the difference, and the guide on converting an image to vector shows how to trace your own artwork.

Upscaling is also the wrong fix when a photo is blurry because of motion or missed focus. AI can make the pixels look sharper, but it is inventing the sharpness, so the result may not match the real scene.

Finally, consider replacing the image. Re-shooting, requesting the original from the owner, or using a stock photo with a larger size may be quicker and safer than repairing a poor file. If you need to isolate the subject first, see our guide on removing a background from an image.

Does upscaling need to be disclosed or labeled?

Usually not for minor enhancement, but it depends on the platform and on what the AI changed. YouTube's disclosure help page lists upscaling among minor edits that do not require an AI label, together with beauty and color filters, background blur, captions and voice repair (YouTube Help, Disclosing use of altered or synthetic content).

The same page asks creators to disclose realistic content that is generated or meaningfully altered, such as footage that alters a real event or place. If an AI upscaler visibly changes what a face or a scene looks like, treat that as an alteration and check the rules for your channel, client or publisher.

How do you check that an upscale worked?

Compare against the original at 100% zoom, then at the final display size. A result that looks sharp in a small preview can still hide artifacts that show up in print or on a large screen.

Use this checklist before you deliver the file.

  • Edges: no halos, jagged steps or double outlines around objects.
  • Faces: eyes, teeth and skin look like the same person as in the original.
  • Text: every letter and number matches the source.
  • Texture: no repeated patterns, plastic skin or painterly smearing.
  • Color: skin tones and brand colors did not shift.
  • Size: the pixel dimensions match your calculation, and the file opens in the target app.
  • File: exported once, with a sensible format, and the original kept safe.

Frequently asked questions

Can you upscale an image without losing quality?

Not in the strict sense. The original has a fixed amount of information, and upscaling adds pixels that are estimated from it. A good method can make the enlarged image look clean and natural, but it cannot recover detail that was never captured, so the aim is a result that holds up at the size you need.

What is the best way to upscale a photo?

For most photos an AI upscaler gives the most convincing detail, while bicubic or Lanczos resizing in an editor works for small enlargements. Start from the best source, upscale once, and inspect at 100% zoom. Try two methods on the same image and keep the one with fewer artifacts.

How many pixels do I need for printing?

Multiply the print size in inches by the pixels per inch. At 300 pixels per inch, an A4 page needs 2480 by 3508 pixels, according to Wikipedia's table. Your printer may accept less for large formats seen from a distance, so ask for their requirement before you spend time upscaling.

Does AI upscaling change what is in the picture?

It can. AI upscalers add detail from patterns they learned in training, and Wikipedia notes there is no guarantee that the added features exist in the original image. That is fine for a web product shot, but it is a problem for evidence, medical images or documents where accuracy matters.

Should I upscale a logo?

Usually you should vectorize it instead. A vector file describes shapes and scales to any size without blur, while an upscaled raster stays an estimate. If you own the logo, trace it in a vectorizing tool or redraw it, and keep the vector file as the master version.

Is DPI the same as pixel dimensions?

No. Pixel dimensions are the width and height of the image in pixels, and they decide how much detail the file holds. Pixels per inch only describes how densely those pixels are laid out when printed. Changing the number alone does not add detail, so calculate from pixel dimensions.

Sources and further reading