AI Image Upscaler: Upscale Images 4x to 8 MP in Bulk

Upscale and enhance photos 4x with Real-ESRGAN, up to 8 megapixels per image, for $0.01 per image. Paste image URLs, or point the Actor at another Actor's dataset to upscale every image a product scrape found.

Quick start

{"imageUrls": ["https://example.com/product-1.jpg", "https://example.com/product-2.png"]}

Upscale every image your product scrape returned:

{"datasetId": "<dataset ID of your scraper run>", "imageField": "image"}

Leave imageField empty to find image fields automatically, including lists of image URLs. To run it after another Actor finishes, open that Actor's Integrations tab and add this Actor. Keep the prefilled {"datasetId": "{{resource.defaultDatasetId}}"}: it selects the finished run's dataset and allows this Actor to read it.

Price

Event	Price
Image upscaled, any output size up to 8 MP	$0.01 per image
Actor start, once per run	Apify's standard start fee: $0.00005 per GB of run memory ($0.0001 at the default 2 GB)

The per-image fee is charged only after the upscaled image is saved. Failed images are never charged the per-image fee. The start fee is charged once per run, even when every image fails. If you set a maximum charge for a run, images beyond it are not processed or delivered.

Output

One dataset row per input image: sourceUrl, outputUrl, inputWidth, inputHeight, outputWidth, outputHeight, outputMegapixels, model, computePath (gpu or cpu; empty for failed images), msElapsed, status (succeeded or failed) and error. Upscaled files are saved in the run's key-value store, and outputUrl links to each one. Anyone who has an outputUrl link can download that file.

Limits

• Input: JPEG, PNG or WebP, up to 30 MB and 40 megapixels. Animated images use the first frame.
• Output: up to 8 megapixels (about 3264 x 2448). Larger 4x results are reduced to fit.
• Time on our GPU: in our test runs, each 8 MP image took 15 to 40 seconds from download to saved file, most of it waiting for our GPU worker to start. Whole runs of one or two images took 30 to 55 seconds.
• Time on CPU: when our GPU is busy or offline, the Actor upscales on its own CPU with a lighter model instead. In our tests at the default 2 GB memory, an 8 MP image took 40 to 60 seconds when it was the only image, and up to about 3 minutes each when two images shared a run. Images in a run take turns on the CPU, so large batches take several minutes. At 4 GB, an 8 MP image took about 23 seconds. The computePath field shows which route made each image.

How your images are handled

• Download. The Actor downloads each image URL from Apify's servers, with the user agent Mozilla/5.0 (compatible; ImageUpscaler/1.0; +https://cybernative.ai), following up to 5 redirects. Only submit images you are allowed to process.
• Isolation. Every image is decoded and processed in a separate process that has no network access and cannot read the Actor's credentials.
• GPU route. When our GPU is available, a copy of each image, resized to at most 2 megapixels and stripped of metadata, is uploaded to a private Amazon S3 bucket in the AWS us-east-1 region that CyberNative AI LLC controls. It is upscaled on a GPU computer that CyberNative AI LLC operates, and the result is written back to that bucket. The Actor deletes the copy and the result as soon as it has read the result. Anything left behind, for example after a crash, is removed by an S3 lifecycle rule set to 1 day. S3 applies these rules asynchronously, so removal can take longer than 24 hours. The GPU computer keeps no image. Its job log records only the job ID, how the run was started, the model, pixel counts, processing time and any error reason, and has no fixed deletion date.
• Your results. Upscaled files and dataset rows are stored in your run's default storage in your Apify account and kept under Apify's data retention for your plan.
• Run log. When the GPU route is configured, the Actor saves one record per run in the same S3 bucket so we can count usage: the run ID, a salted one-way hash of your Apify user ID, whether your Apify account is on a paid plan, how the run was started (for example API or schedule), how the dataset was chosen, the route and isolation status, charge counts, and for each image only its status, error type, route, file size, pixel counts and time. It contains no image, URL, file name or other input. It is set to expire after 400 days, applied asynchronously in the same way.
• No training. Your images are not used to train any model.

Model and licence

Real-ESRGAN by Xintao Wang et al., BSD 3-Clause licence (LICENSES/Real-ESRGAN-BSD-3-Clause.txt). Weights: RealESRGAN_x4plus.pth (photo), RealESRGAN_x4plus_anime_6B.pth (illustration) and realesr-general-x4v3.pth with realesr-general-wdn-x4v3.pth (CPU route), from the official Real-ESRGAN releases under the same licence, each pinned by SHA-256. No face-restoration model is used.

About

Built and run by CyberNative AI LLC, an AI-run company. Contact: hello@cybernative.ai