Skip to main content
Preparing workspace

Photo Upscales

Upscale photos 2x to 8x with ESRGAN models, or remove noise, deblur and brighten dark shots with MAXIM, run in your browser through TensorFlow.js.

Still loading. If the tool doesn’t appear, reload this page.

About this tool

Photo Upscales

AI image upscaler and photo enhancer that runs in your browser: upscale 2x to 8x, denoise, deblur and brighten low-light photos with ESRGAN and MAXIM models.

Photo Upscales is an in-browser AI image enhancer built on UpscalerJS and TensorFlow.js. It enlarges images 2x, 3x, 4x or 8x with ESRGAN super-resolution models, and uses MAXIM models to remove sensor noise, reduce motion blur, brighten low-light shots or retouch tone and colour. Your image is processed in this browser tab and is not uploaded; the model you choose is downloaded from cdn.jsdelivr.net the first time you run it, using WebGL when your browser supports it. If a model cannot be downloaded, you can run a basic canvas sharpening filter instead, clearly labelled as not AI.

How to use Photo Upscales

  1. 1

    Upload a PNG, JPG, WEBP or BMP image, or pick one of the sample images under the drop zone.

  2. 2

    Choose an operation: Super Resolution (2x to 8x), Denoising, Deblurring or Image Enhancement for low-light photos.

  3. 3

    Pick a model (ESRGAN Slim, ESRGAN Medium, Default Model or one of the MAXIM models) and, for ESRGAN, the tile size and seam padding.

  4. 4

    Click Run: the first run downloads the model from cdn.jsdelivr.net, then the image is processed in your browser.

  5. 5

    Drag the before-and-after divider, zoom to 2x or 4x to inspect detail, then download the result as PNG, JPG or WEBP.

Why use Photo Upscales

  • Processed in your browser

    Inference runs on your own device with TensorFlow.js, on the GPU through WebGL when available. Only the model files are downloaded; your image is never uploaded.

  • Interactive before/after comparison

    A draggable split slider with 1x, 2x and 4x zoom and panning lets you check edges and textures pixel by pixel; the arrow keys move the divider too.

  • Tile processing with overlap padding

    ESRGAN models run on 32, 64 or 128 px tiles with 2 to 8 px of overlap, which keeps GPU memory use down on larger images and hides most tile seams.

  • Seven pretrained models

    Three ESRGAN-family upscalers (Slim, Medium and the 2x default model) and four MAXIM restoration models for denoising, deblurring, low-light enhancement and retouching.

  • Size check before processing

    When an image is larger than the recommended size for the chosen model and scale, the tool warns you and offers a one-click downscale first.

Common use cases

  • Enlarging small web graphics, old digital photos and low-resolution scans before reuse.
  • Cleaning high-ISO sensor noise from phone and camera photos with the SIDD-trained MAXIM denoiser.
  • Reducing camera-shake and motion blur with the GoPro-trained MAXIM deblurring model.
  • Brightening underexposed evening and indoor shots with the LOL-trained low-light model.
  • Trying client-side neural inference with tile-based processing, progress reporting and cancel support.

Frequently asked questions

How does in-browser AI photo upscaling work?

The tool loads ESRGAN or MAXIM neural networks compiled for TensorFlow.js and runs them in your browser tab, on the graphics card through WebGL when available. The network predicts the missing high-frequency detail instead of only blending neighbouring pixels.

Are my photos uploaded to a server?

No. Your image is read and processed inside this browser tab. The only downloads are the UpscalerJS library and the model you choose, which come from cdn.jsdelivr.net the first time you run that model (UpscalerJS fetches the weights from unpkg.com if jsDelivr fails).

What is the recommended maximum image size?

It depends on the model and scale: about 680 to 800 px on the longest side for 2x, 420 to 480 px for 4x, 220 to 256 px for 8x and 512 px for the MAXIM models. Larger images can run out of GPU memory, so the tool offers to downscale them first.

What is the difference between ESRGAN and bicubic resampling?

Bicubic interpolation blends neighbouring pixels, so enlarged edges look soft. ESRGAN-family models are trained neural networks that generate plausible edges and textures, so results look sharper, although invented detail can differ from the real scene.

What happens if a model will not load?

You see an error that says the model could not be loaded, with the reason. You can try again, or run the basic canvas filter (resampling plus sharpening, denoise or brightening) instead; its result is labelled as not AI.

Can I export the enhanced image in different formats?

Yes. Download the result as lossless PNG, JPG at 92% quality or WEBP at 92% quality from the bar under the comparison.