Skip to content
UtilPick.
My toolbox
All tools
Favorites
Saved tools

Loading…

Recently used

Tools you open will appear here.

Get the desktop app
Settings
UtilPick.
All tools
All tools/AI Image Upscaler (Enhance Resolution)

AI Image Upscaler (Enhance Resolution)

Enlarge blurry photos and small drawings 2× or 4× with AI (Real-ESRGAN). Photo and art/anime modes, a before/after slider, and PNG, JPG or WebP saving. Images never leave your browser.

Loading the tool…

Used this tool? Share what worked for you — it helps the next person.

User reviews

No reviews yet.

Write a review

Checking your sign-in status…

    From first step to finished result

    AI Image Upscaler (Enhance Resolution): complete guide

    Walk through the steps and examples, or jump straight to the topic you need.

    How to use
    Examples & choices
    Troubleshooting
    FAQ

    At a glance

    Models
    Two models made from the official weights of the public Real-ESRGAN model (BSD-3-Clause), converted to ONNX. Photo uses realesr-general-x4v3 (2.4MB) and Art / anime uses RealESRGAN_x4plus_anime_6B (9.0MB); each is downloaded once, the first time you use it.
    Accepted files
    PNG, JPG, WebP, GIF, BMP and AVIF, up to 25MB. The original's long side can be up to 2,048px at 2x and 1,024px at 4x (result long side up to 4,096px).
    Result
    Shows a before/after slider with a 100% view, the result size, the original size and the processing time, and saves as PNG, JPG or WebP.
    Speed
    A few seconds in browsers with graphics acceleration (WebGPU); tens of seconds to a few minutes with the standard engine (WebAssembly). It shows the estimated time before you run it, and the tile count and time remaining while it runs.
    Privacy
    The image is processed only inside this browser and is never sent to a server. The internet is used only to download the model.

    How to use AI Image Upscaler (Enhance Resolution)

    1. 01

      Add an image

      Drag a file in or click 'Choose file'. To try it first, click 'Try a sample image' and one UtilPick illustration (240x160) is loaded.

    2. 02

      Choose the image type

      Choose 'Photo' for photos and screenshots, or 'Art / anime' for illustrations, comics, anime and logos with clear lines. The wrong choice can make a photo look painted or leave the lines of a drawing less clean.

    3. 03

      Set the scale and run

      Pick 2x or 4x and the result size appears right below. If the image is too large for that scale, you are told why, and if 2x would work, a 'Switch to 2x' button appears.

    4. 04

      Compare and save

      Move the slider to compare the original (simply enlarged) with the AI result, and switch to '100%' to check at actual pixel size. Choose a file format and click Save; the file is named like 'originalname-upscaled-4x.png'.

    What it’s good for

    How AI upscaling differs from simple enlarging

    Ordinary enlarging (bilinear or bicubic interpolation) fills the new pixels by blending their neighbors, so the bigger you go, the blurrier it gets. Real-ESRGAN is a model trained on many pairs of low-quality and high-quality photos, so it redraws blurry edges as crisp lines and compression blotches as smooth areas. The result looks sharper at the same size, but that detail is the model's estimate, not information that was in your original.

    Photo or Art / anime: which to choose

    For photos and screenshots, choose 'Photo' (realesr-general-x4v3). It reduces noise and blur while keeping texture fairly intact. For illustrations, comics, anime and logos with clear lines, 'Art / anime' (RealESRGAN_x4plus_anime_6B) keeps outlines cleaner. Using Art / anime on a photo can make skin or grass look painted with a brush.

    When not to use it

    Do not use it on images that must stay factual, such as ID cards, ID photos, contracts or accident photos. Text and facial contours generated by AI can differ from the original, and submitting the result as if it were the original can cause problems.

    How it runs and your privacy

    The first time you use it, the tool downloads the official Real-ESRGAN (BSD-3-Clause) weights converted to ONNX (2.4MB for Photo, 9.0MB for Art / anime) once and runs them in this browser. Your image is never uploaded to a server. It is fast in browsers with graphics acceleration (WebGPU); otherwise it falls back to the standard engine, and large images can take a few minutes.

    Worked examples

    Checking whether a scale will work

    Input and settings
    A 1,500x1,000 photo at 4x
    Result
    4x is not possible (original long side up to 1,024px) - 2x gives 3,000x2,000
    The original's limit is the 4,096px result limit divided by the scale. At 4x it would be 6,000px, which is over the limit; at 2x it is 3,000px, which fits.

    Tile count and estimated time

    Input and settings
    A 1,024x1,024 photo at 4x (no graphics acceleration)
    Result
    25 tiles of 231px - about 67 seconds
    It splits the image into 5x5 tiles of up to 256px with 32px of overlap. The measured processing time was 65.1 seconds, and the same image took 2.0 seconds with graphics acceleration (Apple M-series Mac).

    How much sharper did it get

    Input and settings
    Sample illustration 240x160 -> Art / anime 4x (960x640)
    Result
    Edge sharpness (Laplacian variance) bicubic 52.9 -> AI 2,685
    This compares against the same image enlarged by bicubic interpolation. A higher number means crisper edges, and that detail is drawn by the model as its estimate.

    Choosing between Photo and Art / anime

    Swipe the table sideways to see the rest.

    Choosing between Photo and Art / anime
    ModeBest forGood to know
    PhotoPhotos, screenshots, blurry capturesReduces noise and blur while keeping texture. The model is small (2.4MB), so it is relatively fast even with the standard engine.
    Art / animeIllustrations, comics, anime, logosKeeps outlines clean. The model is heavier (9.0MB), so without graphics acceleration it takes about 5 times longer than Photo.
    2x and 4xUsually 2x; 4x for very small drawingsBoth models are 4x models, so 2x takes the same processing time. 2x shrinks the result to exactly half, so the file is smaller and noise is less visible.

    When the result isn’t what you expected

    It says 'Stopped because memory ran out'

    Larger results use more memory. Switch to 2x or shrink the image, close other tabs, then click 'Try again'. A PC is more stable than a phone.

    It takes too long

    Browsers without graphics acceleration (WebGPU) use the standard engine, which is slow. Open the page in the latest Chrome or Edge, or shrink the image to make it proportionally faster. Switching to 2x does not shorten the time. You can click 'Stop' at any time.

    My photo looks smeared like a painting

    You probably processed it as 'Art / anime'. Switch to 'Photo' and run it again.

    Text or faces look different from the original

    This happens because the AI draws new plausible shapes. For text, numbers and faces that must be accurate, go by the original, and do not use the result as proof or evidence.

    Key terms

    Upscaling
    Increasing the number of horizontal and vertical pixels in an image. In AI upscaling, a trained model draws the added pixels.
    Real-ESRGAN
    A public model trained on pairs of low-quality and high-quality images that enlarges while removing blur, noise and compression artifacts (BSD-3-Clause).
    WebGPU
    A feature that lets the browser compute on the graphics card. When available, it is many times faster.
    Bicubic interpolation
    The common enlarging method that blends the surrounding 16 pixels to make each new pixel. The bigger you go, the blurrier it looks.

    AI Image Upscaler (Enhance Resolution) FAQ

    Is it really free? Is there a watermark?

    It is free, with no watermark, no sign-up and no usage limit. Processing happens in your browser, so it costs no server resources. The result is capped at 4,096px on the long side, so the original's long side can be up to 2,048px at 2x and up to 1,024px at 4x.

    Will blurry text become readable?

    Slightly blurry text becomes sharper, but text that was smeared beyond recognition cannot be restored accurately. The model may even draw different letter shapes, so do not rely on it to verify numbers or names.

    How is 2x made?

    Both models are 4x models, so 2x is made by shrinking the 4x result to exactly half (averaging 2x2 pixels). The file is smaller than a 4x result and noise is less visible.

    Does it work with transparent PNGs?

    Yes. Color is enlarged with AI and transparency with bicubic interpolation, so the transparent background is kept. If you save as JPG, transparent areas are filled with white.

    Can I process several images at once?

    Currently one at a time. When one finishes, click 'Start over with another image or setting' and add a new image. The model is not downloaded again.

    The PNG file is too big.

    PNG is saved losslessly, so it is large. Switching the file format to JPG or WebP shrinks it to a fraction of the size. JPG fills transparent areas with white.

    Jump to a topic

    • #photo vs art mode
    • #estimated time
    • #out of memory
    • #Real-ESRGAN

    Reviewed 2026.10.07

    Related tools

    View all
    Compress ImagesImages
    Image Converter (PNG, JPG, WebP, AVIF)Images
    Background RemoverImages
    UtilPick

    Simple tools for a complicated day. A free collection of online utilities.

    Service

    About UtilPick Get the desktop app

    Help

    Request a toolReport a problemContactinfo@utilpick.com

    Policies

    Privacy Policy (Korean)Terms of Use (Korean)
    Business registration no. 365-47-00936 (Republic of Korea)
    © 2026 UtilPick
    Featured on Twelve ToolsUtilPick - Featured on Startup FameFazier badgeFeatured on LaunchIgniter
    HomeSearchFavoritesMore
    Drop an image to upscale

    PNG, JPG, WebP, GIF, BMP or AVIF up to 25.0MB · result up to 4096px on the long side

    Image type

    Tuned for real scenes such as photos and screenshots; reduces noise and blur.

    Scale

    2× takes up to 2048px on the long side, 4× up to 1024px.

    AI invents plausible detail that was not in the original. Text, faces and licence plates can change, so do not use it for ID, certificate or evidence photos.

    Images are processed only in this browser and never sent to a server. The internet is used only to download the AI model the first time.

    Models and how it works

    Photo uses realesr-general-x4v3 (2.4MB); art / anime uses RealESRGAN_x4plus_anime_6B (9.0MB). Both are the official Real-ESRGAN weights converted to ONNX and halved in size (fp16); the SHA-256 fingerprint is checked after download.

    The image is split into overlapping tiles that are processed one by one and blended smoothly where they overlap. 2× is made by halving the 4× result exactly, and transparency is enlarged with bicubic interpolation instead of AI. Results are limited to 4096px on the long side.

    Real-ESRGAN © 2021 Xintao Wang, BSD 3-Clause License. Source and license