# What is image upscaling?

URL: https://dreambench.ai/glossary/image-upscaling

Image upscaling makes a picture larger, with more pixels than it started with. Simple resizing spreads the existing pixels out and looks soft. AI upscaling, also called super-resolution, uses a model trained on pairs of small and large images to predict the extra detail, so edges and textures stay sharp.

Written and reviewed by the DreamBench team. Published 2026-10-05; updated 2026-10-22.

## Image upscaling at a glance

- **Makes:** A larger picture, usually 2x or 4x the width and the height.
- **How:** A super-resolution model predicts detail for the new pixels.
- **Keeps:** The composition, the colors and the shapes of the original.
- **Invents:** Fine texture: skin, foliage, fabric, the crispness of edges.
- **Can’t:** Recover text or faces too small to make out in the original.

**Which upscale do I need?**

- **A social post or a website:** 2x: it doubles the width and the height, which is plenty for a screen, and keeps the result close to the original. Choose 2x on the [AI image upscaler](https://dreambench.ai/tools/ai-image-upscaler).
- **A print, a poster or a 4K screen:** 4x, for the most pixels: choose it on the [AI image upscaler](https://dreambench.ai/tools/ai-image-upscaler), or use [upscale an image to 4K](https://dreambench.ai/tools/upscale-image-to-4k). The result is capped at 4096 pixels on the long edge, so 4x applies when four times your photo’s long edge fits in that; a larger photo gets 2x.
- **A soft or small photo that just needs to look better:** Enhance rather than enlarge for its own sake: the [AI photo enhancer](https://dreambench.ai/tools/ai-photo-enhancer) runs the same model, set up for a soft phone photo or a small picture saved from a message. It sharpens what is there; it doesn’t retouch faces.
- **A blurred photo:** Unblurring first, then upscaling. Sharpening raises contrast at edges without adding pixels, and an upscaler adds pixels but can’t undo focus or motion blur; [unblur an image](https://dreambench.ai/tools/unblur-image) is set up for the blur a model can fix.

## How AI upscaling works

AI upscaling works by looking at small patches of a picture and predicting what a sharper, larger version of each would contain. [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN), the model DreamBench uses, learned this from sharp pictures degraded on purpose to imitate blur and compression, so it knows how they look and how to undo them as it enlarges.

Upscaling adds pixels, not information the photo never held. A face too small to recognize won’t become a particular person, and some upscalers invent detail to look sharper. DreamBench’s upscaler runs one faithful pass with no face retouching, so a face is still the face.

#### From ESRGAN to Real-ESRGAN

Real-ESRGAN builds on [ESRGAN](https://arxiv.org/abs/1809.00219) (2018), which improved an earlier network to give more realistic and natural textures. Real-ESRGAN kept that kind of network and trained it on pictures degraded on purpose, in ways meant to imitate real photos, so it could cope with them.

## Ways to make a picture bigger

**What each method adds**

|  | What it adds | Faithful to the original | Good for |
|---|---|---|---|
| Resizing | More pixels, each an average of its neighbors | Yes, but soft | Small enlargements, previews |
| AI super-resolution | Predicted edges and texture | Mostly: it sharpens what is there | Photos, product shots, prints |
| Generative upscaling | New detail drawn by an image model | Less: it can change faces and text | Illustrations and generated images |
| [Outpainting](https://dreambench.ai/glossary/outpainting) | More scene around the edges, not more detail | The original is kept; the rest is new | A wider or taller frame |

## Where people use upscaling

- Printing a photo larger than its pixels allow.
- Old or small pictures that need to fill a slide, a website or a video frame.
- Product images a marketplace wants at a minimum size.
- Pictures that were generated or saved small.

## Where upscaling falls short

- Faces and text: when they are too small to read, an upscaler can only guess, and a guess is not the original.
- Repeating patterns such as brick, fabric or leaves can come out too regular.
- Heavy compression: blocky JPEG artifacts can be sharpened along with everything else.
- Video: upscaling each frame on its own can make detail shimmer from frame to frame, a failure of [temporal consistency](https://dreambench.ai/glossary/temporal-consistency).

If an upscaled image still looks blurry, the blur was in the original: a photo out of focus or shaken when it was taken has no sharp edge to rebuild, so it comes back larger with the same blur, and a soft copy saved from a chat has already lost the detail the model needs. Start from the largest, least-compressed original you have. A 4x run on a large picture is another cause: when four times the long edge would pass the output cap, the model doubles the picture and the last step resizes it the rest of the way, which adds pixels but no detail, so a picture that size gains nothing from 4x over 2x.

Scratches, tears and fading should be repaired before upscaling, with [photo restoration](https://dreambench.ai/tools/ai-photo-restoration); upscaling first would only make the damage larger and sharper.

## How DreamBench upscales an image

The [AI image upscaler](https://dreambench.ai/tools/ai-image-upscaler) runs Real-ESRGAN at 2x or 4x. A photo larger than 2048 pixels on its long edge is scaled down first, and the result is at most 4096 pixels on its long edge.

An upscale uses 7 credits. Two presets are set up for one job each: the [AI photo enhancer](https://dreambench.ai/tools/ai-photo-enhancer) for a soft or small photo, and [upscale an image to 4K](https://dreambench.ai/tools/upscale-image-to-4k) for the largest result.

**What the upscaler returns, by the picture you start from**

|  | At 2x | At 4x | Why |
|---|---|---|---|
| A 500-pixel picture | 1000 pixels on the long edge | 2000 pixels | Both are the model’s own pass; four times 500 is well inside the cap. |
| A 1000-pixel picture | 2000 pixels | 4000 pixels | The model’s 4x pass, just inside the 4096-pixel cap. |
| A 1920 × 1080 frame (1080p) | 3840 × 2160, which is 4K UHD | 4096 pixels on the long edge | Four times 1920 would pass the cap, so the model doubles it and the last step resizes it to the cap: pixels, not detail. |
| A picture of 2048 pixels or more | 4096 pixels | 4096 pixels | It is scaled to 2048 pixels before the run, so the result is the cap either way. |

A PNG upload comes back as a PNG, its [transparent background](https://dreambench.ai/glossary/transparent-background) included, and a JPEG as a JPEG; the file grows with the pixel count, which is four times at 2x and sixteen at 4x, a PNG most of all. The upscaler takes a photo up to 40 MB, works in a phone browser, and the upload is removed from the tool’s storage after 7 days with the result; neither is used to train a model.

## Questions

- **Is upscaling the same as resizing?** No. Resizing stretches the existing pixels, so a bigger picture looks soft. Upscaling with a super-resolution model predicts new detail for the extra pixels, so edges and textures stay sharper.
- **What is the difference between 2x and 4x upscaling?** 2x doubles the width and the height, so the picture has four times the pixels. 4x makes both four times larger, for sixteen times the pixels. On DreamBench a result is at most 4096 pixels on its long edge.
- **Does AI upscaling actually improve image quality?** AI upscaling improves how large a picture can be shown or printed while staying sharp, by predicting edges and texture for the new pixels. It does not recover information the photo never held: a face or a word too small to read stays that way, and an upscaler that seems to recover them is inventing them. Check faces and text at full size.
- **Why does my upscaled photo look plastic or waxy?** The upscaler smoothed away real texture, usually by denoising a compressed or soft input and filling the gap with learned skin. Start from the least-compressed copy you have, and prefer an upscaler that runs one faithful pass without face retouching, as DreamBench’s does.
- **Is upscaling the same as sharpening or unblurring?** No. Sharpening raises contrast at edges without adding pixels; unblurring tries to undo focus or motion blur at the same size; upscaling adds pixels and predicts detail for them. A soft photo may need unblurring first, then upscaling.
- **Can I upscale a screenshot, a small logo or a video?** A screenshot or logo, yes: flat graphics with clean edges upscale well, though tiny text stays a guess, so retype it where you can, and a vector file is better for a logo when one exists. A video, no: DreamBench’s upscaler works on photos, and upscaling frames one by one makes detail shimmer.
- **Does upscaling improve print quality?** It lets you print larger at the same sharpness, because there are more pixels per inch. For 300 dpi, divide the long edge in pixels by 300 for inches; DreamBench’s upscaler returns at most 4096 pixels on the long edge, and posters viewed from a distance print well at lower densities. The extra detail is predicted, not recovered, so check faces and text at full size before you print.
- **How long does AI upscaling take?** AI upscaling takes well under a minute on DreamBench for most pictures. The super-resolution model makes one pass over the photo, and the result appears beside the original on the page when it finishes. A 4x run on a large picture takes the model’s 2x pass and a resize instead of a 4x pass that would take minutes only to be shrunk back to the 4096-pixel cap.
- **What file do I get back from an upscaler, a PNG or a JPEG?** On DreamBench a PNG upload comes back as a PNG, transparency included, and anything else as a JPEG. Upload graphics, screenshots and logos as PNG so flat color, straight edges and lettering stay lossless; saving a JPEG as a PNG first adds nothing, since the compression damage is already in the pixels. The file grows with the pixel count: four times at 2x, sixteen at 4x.
- **Does upscaling change the aspect ratio or crop the picture?** No. Upscaling multiplies the width and the height by the same factor, so the framing and the aspect ratio stay exactly as they were: nothing is cropped, extended or moved, and a 16:9 picture is still 16:9. To change a picture’s shape, outpainting adds new scene around it and cropping takes some away; DreamBench’s image expander does the first.
- **Does upscaling cost anything?** On DreamBench you create a free account and choose a plan, and each upscale uses 7 credits. The upscaler works in a phone browser, with nothing to install.

## Sources

- [Real-ESRGAN on GitHub](https://github.com/xinntao/Real-ESRGAN), Tencent ARC: read 2026-10-05
- [Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data (arXiv)](https://arxiv.org/abs/2107.10833), arXiv: read 2026-10-05
- [ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks](https://arxiv.org/abs/1809.00219), arXiv: read 2026-10-13

DreamBench is an independent platform and isn’t affiliated with or endorsed by the companies named on this page.