Image Upscaler: 2x and 4x AI super resolution in bulk

Give it a list of image URLs and it enlarges each one by 2x or 4x with Real-ESRGAN, an open super resolution model. The model runs inside the actor, so your images go to no outside API and there is no key to set up. Every enlarged file goes into the run's key value store, and the dataset gives you a direct download link with the size before and after.

What it does

• Upscales by 2x or 4x. A 550 by 368 photo comes out at 1100 by 736 or 2200 by 1472.
• Sharpens while it enlarges. The model rebuilds edges and texture instead of stretching pixels, so the result is crisper than a plain resize.
• Three noise levels. Low keeps the most texture and grain. High smooths away noise. Medium sits between them. On a photo with added noise, high left 29% less fine noise than low.
• Keeps transparency in PNG and WebP output. The alpha channel is enlarged with Lanczos. JPEG output fills transparent areas with white.
• Turns photos upright from their EXIF orientation first, so portrait photos stay portrait.
• Handles many images in one run. Each image gets its own result, and one bad link never stops the run.

Reads JPEG, PNG, WebP, GIF, BMP, TIFF and the other formats Pillow reads. An animated GIF is upscaled from its first frame, and the item says so in note.

Example

A 400 by 301 PNG with a transparent background came out at 1600 by 1204 at 4x, with its transparency kept. A 1024 by 772 WebP photo came out at 4096 by 3088.

Input

• Image URLs (required)
• Scale: 2x (default) or 4x
• Noise removal: low, medium (default) or high
• Output format: PNG (default), WebP or JPEG
• Quality (default 90), for JPEG and WebP
• Largest source image in megapixels (default 1, ceiling 4). A larger source gets an error item and is not charged. 1 megapixel is about 1000 by 1000 pixels.
• Maximum file size in MB (default 25)

Output

Dataset: one item per URL with outputUrl (download link), outputKey (the file name in the key value store), scale, denoise, inputFormat, inputWidth, inputHeight, outputFormat, outputWidth, outputHeight, outputBytes, hasTransparency, processingSeconds, sourceMegapixels and chargedEvent. Files are named by their position in your list, the source file name and the scale, such as 0001-photo-x4.png.

A URL that cannot be upscaled gets ok: false and an error, such as HTTP 404, an unknown host, a file over the size limit, a source over the megapixel limit, or "the file is not an image format this actor can read" for a web page.

Speed

The model runs on the CPU, so time grows with the number of source pixels. At the default 4 GB of memory, a 550 by 368 photo took 6 seconds at 4x and a 400 by 301 PNG with transparency took 4 seconds.

Model and license

Real-ESRGAN, model realesr-general-x4v3, by Xintao Wang and contributors, released under the BSD 3-Clause License. The three noise levels blend its two published weight files the same way the official Real-ESRGAN inference script does. For 2x, the 4x result is scaled down with Lanczos, as that script also does.

Pricing

Pay per event, one event per image upscaled, set by the size of the source image:

• image-upscaled for a source of 0.5 megapixels or less, such as 800 by 600
• large-image-upscaled for a source above 0.5 megapixels, up to the megapixel limit you set

Each item shows its sourceMegapixels and the chargedEvent. URLs that fail are not charged. The run stops when it reaches the maximum charge you set.

Scope and conduct

Only the URLs you supply are downloaded. The actor does not log in anywhere and uses no proxies. Upscale images you have the right to use. An upscaled image adds detail the model predicts, so do not treat it as evidence of what the original showed.

Support

Questions and problems: the Issues tab on this page. Replies are written by an AI assistant on behalf of the developer and reviewed by the developer.