AI Image Upscaler: Real-ESRGAN 2x, 3x, 4x Photo Enhancer

AI Image Upscaler makes images 2x, 3x or 4x larger and sharper with Real-ESRGAN, the open-source super-resolution model. It removes blur, JPEG blocks and noise while it enlarges: good for photos, product shots, old scans, screenshots, thumbnails and AI-generated art. Paste image links or point it at your own uploads, and get PNG, JPEG or WebP files with a link for each. The model runs inside the actor: no API key, no GPU service, no per-call fees to anyone else.

• ✅ 2x, 3x or 4x upscaling with Real-ESRGAN (realesr-general-x4v3), tile by tile so large images fit in memory.
• ✅ Any common image: JPEG, PNG, WebP, GIF, BMP, TIFF and AVIF. EXIF rotation is applied, transparency and grayscale are kept.
• ✅ PNG, JPEG or WebP output with adjustable quality, saved in the run's key-value store with a public link.
• ✅ Your own files: upload images to an Apify key-value store and list them; no hosting needed.
• ✅ Batch-safe: a 404, a web page, a corrupt file or an over-size image never fails the run. You get an item with the reason, for free.
• ✅ Pay per image: $0.01 per image up to 0.5 MP (e.g. 800x600), $0.02 up to 1 MP, $0.04 up to 2 MP. Same price at 2x, 3x and 4x. No start fee.

Start in one minute

Click Start with the prefilled input:

{
  "urls": [
    "https://upload.wikimedia.org/wikipedia/commons/thumb/9/98/Aldrin_Apollo_11_original.jpg/500px-Aldrin_Apollo_11_original.jpg"
  ],
  "scale": 4
}

You get: One Apollo 11 photo upscaled 4x (500x503 → 2000x2012 PNG), with outputUrl, outputWidth, outputHeight and billedUnits.

Cost: $0.01: 1 image × $0.01 (the input is 0.25 MP). No start fee.

Use with an AI assistant

Connect the Apify MCP server to Claude, ChatGPT, Cursor or any MCP client (https://mcp.apify.com?tools=cprussin/image-upscaler loads only this actor), then paste:

Use the Apify actor cprussin/image-upscaler to upscale https://upload.wikimedia.org/wikipedia/commons/thumb/9/98/Aldrin_Apollo_11_original.jpg/500px-Aldrin_Apollo_11_original.jpg 2x to JPEG, and give me the download link.

Recipes

Product photos 2x as JPEG

Small product shots made sharper, as compact JPEGs.

{
  "urls": [
    "https://example.com/products/shoe-small.jpg",
    "https://example.com/products/bag-small.png"
  ],
  "scale": 2,
  "outputFormat": "jpg",
  "quality": 92
}

Web-ready WebP

The smallest output files, for a website.

{
  "urls": [
    "https://upload.wikimedia.org/wikipedia/commons/thumb/9/98/Aldrin_Apollo_11_original.jpg/500px-Aldrin_Apollo_11_original.jpg"
  ],
  "scale": 2,
  "outputFormat": "webp",
  "quality": 85
}

Your own uploads

Images uploaded to a key-value store in your Apify account.

{
  "keyValueStoreRecords": ["my-user~photos/scan-001.jpg"],
  "scale": 4
}

Cap the price per image

Skip inputs over 0.5 MP (free error item), so no image costs more than $0.01.

{
  "urls": [
    "https://upload.wikimedia.org/wikipedia/commons/thumb/9/98/Aldrin_Apollo_11_original.jpg/500px-Aldrin_Apollo_11_original.jpg"
  ],
  "maxInputMegapixels": "0.5"
}

What can I use an AI image upscaler for?

• E-commerce: enlarge small supplier or marketplace product photos for zoom views and print.
• Old photos and scans: sharpen low-resolution family photos, archive scans and document images.
• AI art and design: upscale Stable Diffusion, Midjourney or DALL-E outputs for prints, wallpapers and thumbnails.
• Web and social content: turn small logos, screenshots and thumbnails into crisp large images.
• Automation and AI agents: add an "upscale this image" step to Make, Zapier, n8n or an MCP-enabled agent through the Apify API.

How does it work?

1. Each image is downloaded (with size and time limits), or read from your key-value store.
2. The format is checked from the file's first bytes; the size is read from its header. Images over Max input size are skipped (free).
3. Real-ESRGAN upscales the image 4x on CPU with ONNX Runtime, in overlapping 512 px tiles so seams don't show. For 2x and 3x the 4x result is resized down with Lanczos filtering (sharper than running a 2x model).
4. The result is encoded as PNG, JPEG or WebP and saved to the run's key-value store. The dataset item has the link, input and output sizes and the processing time.

How do I upscale images?

Field	Description	Default
urls	Direct links to images, one per line	a NASA photo
startUrls	Same, in Apify request-list format (combined with urls)	none
keyValueStoreRecords	Uploaded images, as storeId/recordKey or username~store-name/recordKey	none
scale	2, 3 or 4	4
outputFormat	png (lossless, transparency), jpg (small) or webp (smallest)	png
quality	JPEG/WebP quality, 1-100	90
maxInputMegapixels	0.25, 0.5, 1 or 2: larger images are skipped with a free error	2
maxFileSizeMb	Larger downloads are skipped with a free error (at most 100)	25
timeoutSecs	Max download time per image	60

Upscaling your own files: upload them to a key-value store in your Apify account (Storage → Key-value stores → your store → upload a record) and list them in keyValueStoreRecords, e.g. my-user~photos/scan-001.jpg. Private record URLs (https://api.apify.com/v2/key-value-stores/.../records/...) also work in urls; the actor reads them with your run's token.

Social-network links (Instagram, Facebook, TikTok, X, Pinterest, LinkedIn) are refused with a free error item: upscale images you have the rights to.

What data do you get?

One item per image (shortened):

{
  "url": "https://upload.wikimedia.org/wikipedia/commons/thumb/9/98/Aldrin_Apollo_11_original.jpg/500px-Aldrin_Apollo_11_original.jpg",
  "sourceType": "url",
  "fileName": "Aldrin_Apollo_11_original.jpg",
  "inputFormat": "jpg",
  "inputWidth": 500,
  "inputHeight": 503,
  "inputMegapixels": 0.25,
  "scale": 4,
  "outputWidth": 2000,
  "outputHeight": 2012,
  "outputFormat": "png",
  "outputBytes": 5189662,
  "outputKey": "upscaled-0001-Aldrin_Apollo_11_original.png",
  "outputUrl": "https://api.apify.com/v2/key-value-stores/.../records/upscaled-0001-Aldrin_Apollo_11_original.png",
  "model": "realesr-general-x4v3",
  "billedUnits": 1,
  "processingSeconds": 8.72,
  "warning": null,
  "error": null,
  "upscaledAt": "2026-10-08T07:00:00.000Z"
}

• The upscaled files are in the run's key-value store as upscaled-0001-<name>.<format>, linked from every item.
• Failed images are returned with an error and no output, such as Download failed: HTTP 404., This link is a web page, not an image., The image is 2000x1500 (3 MP), over the 2 MP input limit (maxInputMegapixels). or HEIC/HEIF images (iPhone photos) are not supported.
• warning notes things like an animated GIF (only the first frame is upscaled) or transparency flattened for JPEG.

How much does image upscaling cost?

Pay per event, by the size of the input image. No subscription, no start fee, compute included:

Input size	Example sizes	Price per image
up to 0.5 MP	640x480, 800x600, 700x700	$0.01
0.5 to 1 MP	1024x768, 1280x720	$0.02
1 to 1.5 MP	1280x1024, 1440x900	$0.03
1.5 to 2 MP (the maximum)	1600x1200, 1920x1040	$0.04

That is one image-upscaled event ($0.01) per started 0.5 megapixel of input. The scale and output format don't change the price. Failed and skipped images are free. If you set a maximum cost per run, the actor never upscales an image the remaining budget can't pay for: it skips it with a free error item and moves on to smaller images.

Tips

• Speed: on the default 4 GB memory (1 CPU core), a 0.5 MP image takes about 15-20 s and a 2 MP image about a minute. More memory gives more cores (4 GB per core) and faster images; the price per image stays the same.
• Output size: 4x of a 2 MP image is 32 MP (e.g. 6528x4900). As PNG that is often 25 MB or more; use jpg or webp for smaller files. WebP allows at most 16383 px per side.
• Bigger inputs: images over 2 MP are already large; downscale them first if you really need to upscale them.
• Faces: Real-ESRGAN is a general model. It sharpens faces but doesn't reconstruct facial detail like dedicated face-restoration models do.

FAQ

Which model does it use? Real-ESRGAN realesr-general-x4v3 (BSD-3-Clause), a compact general-purpose model trained to remove real-world blur, noise and compression artifacts while upscaling. It runs on CPU with ONNX Runtime. Larger Real-ESRGAN models (x4plus) were measured at about 12x the CPU time for a modest quality gain, so they aren't offered.

Is there a face-enhancement option? No. The popular face models (GFPGAN, CodeFormer) are under licenses or trained on data that don't allow commercial use, so they aren't included.

Are my images sent to an AI provider? No. Everything runs inside the actor's container on Apify.

Can AI agents use it? Yes, through the Apify API, the Apify MCP server or any Apify integration (Make, Zapier, n8n, LangChain).

Disclaimer: You are responsible for having the right to process the images you submit.

Open-source software

This actor runs Real-ESRGAN model weights (BSD-3-Clause) with ONNX Runtime (MIT) and Pillow (MIT-CMU). See THIRD_PARTY_NOTICES.md in the source for all notices. It is not affiliated with or endorsed by the Real-ESRGAN authors, Tencent ARC or Microsoft.

Related actors

• Video Frame Extractor: Thumbnails, scene keyframes, contact sheets and GIFs from video files.
• Website Screenshot API: Full-page or viewport screenshots and PDFs of any URL.
• PDF to Markdown Converter: PDF, Word, PowerPoint, Excel and EPUB to Markdown or RAG chunks.