AI Image Upscaler 2x / 4x — Real-ESRGAN on CPU, no GPU, pay per image

Make small, blurry or compressed images bigger and cleaner. Give a list of image URLs, get back 2x or 4x upscaled files (PNG, JPG or WebP) as links. Runs the open-source Real-ESRGAN (general x4v3) model on CPU inside the Actor: no GPU, no subscription, no API key, no third-party service.

Good for

• Product and marketplace photos that are too small for a store's size requirement
• Old photos, scans, screenshots, memes, social-media images that were compressed
• Illustrations, anime/manga art, game assets, logos with soft edges
• Batches: hundreds of images in one run from a list of URLs

How it works

The image is cut into overlapping tiles, each tile is enlarged 4x by the neural network, and the tiles are stitched back together without visible seams. For 2x, the 4x result is reduced with a high-quality filter, which usually looks crisper than a plain 2x enlargement. Transparency (alpha) is kept. EXIF rotation is respected.

Input

Field	Meaning
urls	Public image URLs (up to 500 per run).
images	Optional inline files: [{"name": "a.jpg", "base64": "..."}].
scale	2 or 4 (default 4).
outputFormat	png (default), jpg, webp; quality for JPG/WebP.
maxInputMegapixels	Larger inputs are reduced to this size first (default 4 MP ≈ 2000×2000, maximum 8 MP).

Output (one row per image)

{ "ok": true, "source": "https://example.com/small.jpg", "outputUrl": "https://api.apify.com/v2/key-value-stores/…/records/upscaled-0001.png",
  "inputWidth": 640, "inputHeight": 480, "outputWidth": 2560, "outputHeight": 1920, "format": "png", "seconds": 9.4, "warnings": [] }

Pricing

Pay per event: one result event per successfully upscaled image. Unreachable URLs, unreadable files and blocked addresses return an ok: false row and are free.

What it does and does not do (read before you buy)

• It is a fast, general-purpose upscaler (a 1.2-million-parameter network). It sharpens edges, removes compression blocking and mild noise, and adds plausible fine detail. It does not invent faces or text that are not in the picture, and it will not rescue a 30×30-pixel thumbnail.
• The result is visibly sharper than a plain bicubic/Lanczos enlargement, but it is a perceptual model: pixel-accuracy scores (PSNR) are about the same as or slightly below a plain enlargement, and fine textures such as foliage can look slightly painted. Judge by eye on your own images.
• Heavily damaged photos (severe blur, very strong noise) improve less than with large diffusion-based upscalers that need a GPU.
• No face-specific restoration, no colourisation, no object removal.
• Run time grows with input size: about 5–10 seconds per input megapixel on the default 4 GB Actor (a 160×120 or 640×480 image takes well under a few seconds).
• Only images you have the right to process: you are responsible for the files you submit.

Credits and licence

Model: Real-ESRGAN realesr-general-x4v3 by Xintao Wang et al., BSD-3-Clause (https://github.com/xinntao/Real-ESRGAN), distributed through Qualcomm AI Hub Models as an ONNX export (BSD-3-Clause). The Actor code is MIT.