7 best image upscaling APIs: limits, costs and use cases

The right API must preserve the details that matter in your images, accept your source files and produce the required dimensions at a predictable cost. A large upscale factor alone does not establish image quality.

Magic Hour publishes this guide and includes its own product in the comparison. Treat the recommendations as editorial guidance from a vendor, and verify the linked first-party product details and your own output requirements before choosing a tool.

Best image upscaling APIs at a glance

Reviewed September 10, 2026. This comparison uses current provider documentation. The recommendations reflect workflow fit; they are not a head-to-head image-quality or latency benchmark. Magic Hour publishes this guide and is one of the options compared.

1. Magic Hour: upscaling in a broader media workflow

Magic Hour’s AI Image Upscaler API accepts an uploaded image and returns a project ID. Retrieve the completed project to download its output. The endpoint supports 2× and 4× scaling; 4× requires Creator, Pro or Business access.

The documented rates are 25 credits for 2× Preserve and 50 for 2× Balanced or Creative. At 4×, those rates are 100 and 200 credits respectively. These are upscaling rates, not a universal price for every Magic Hour image operation.

Why consider it: You can keep upscaling near the rest of your asset workflow. For example, prepare a source image, enlarge it to the dimensions you need, then use the selected asset in an image-to-video workflow.

Start here: Compare the available modes on one representative image in the AI Image Upscaler. Inspect text, faces and product details before integrating the API or processing a batch.

2. Topaz: specialized models for different source images

Topaz separates precision upscaling, generative enhancement and creative reconstruction. Its model directory helps you choose a model for photographs, low-resolution sources, artwork or other material.

That distinction matters: a model intended to add plausible detail may be useful for artwork, while a product photograph needs close attention to label text, texture and shape. Treat those as different evaluation jobs.

Why consider it: You want to compare dedicated enhancement models rather than accept one general-purpose upscale operation.

Pricing consideration: Topaz’s model pricing documentation varies credit consumption by model family and output megapixels. Its API purchase page lists Developer at $50 per month with 500 credits and Scale at $240 with 3,000. The public pages describe billing at different levels of detail; use the selected model’s current request estimate before budgeting a large job. Do not apply one generic “images per credit” figure to every model.

3. Picsart: upscaling alongside image-editing APIs

Picsart’s standard Upscale endpoint supports factors of 2, 4, 6 and 8. It accepts an image file or source URL and offers JPG, PNG and WebP output. Recommended input dimensions depend on the requested factor: the documented recommendations get smaller as the factor increases.

Why consider it: Your application also needs image-editing operations and you want upscaling within that API ecosystem.

Integration consideration: Standard Upscale, Ultra Upscale and Ultra Enhance are separate operations. Compare the exact endpoint you intend to call, including its input limits and credit cost. An 8× label is not a promise that every source image can be enlarged eightfold without exceeding output limits.

4. Claid: product imagery and multi-operation processing

Claid’s image-editing API combines image operations in a request. This makes it worth evaluating for a catalog workflow that needs enhancement, cropping and other preparation steps together.

Why consider it: The deliverable is a consistent set of product images or print assets, rather than one isolated enlargement.

Pricing consideration: Claid’s API pricing lists upscaling at 1–11 credits depending on output size, with other operations priced separately. A polish operation, for example, adds a credit. Web subscription credits and API credits are separate pools. Compare the total cost of the operations your request uses, not the consumer website’s monthly plan.

5. Cloudinary: upscale within image delivery

Cloudinary’s upscale effect increases both dimensions by four. It accepts inputs smaller than 4.2 megapixels and can be chained with other transformations. The effect is not supported for fetched images.

Why consider it: Your application already stores and delivers images through Cloudinary. Adding a transformation to that pipeline may be simpler than moving assets to another service.

Integration consideration: A new derived image may return HTTP 423 while processing. Prepare important derivatives in advance with an eager transformation instead of assuming the first visitor will receive an immediate result.

Pricing consideration: The transformation-count documentation assigns 10 transformations to an upscale with input below 0.25 MP and 100 for the 0.25–4.2 MP range. Other processing, storage and bandwidth can also affect the account bill.

6. Replicate Real-ESRGAN: a hosted model to evaluate

Real-ESRGAN on Replicate exposes adjustable upscaling and optional face correction. Its model documentation recommends inputs no larger than 1440p. Replicate is the hosting platform; Real-ESRGAN is the specific model being compared here.

Why consider it: You want to try this model through an API before deciding whether to host a model yourself or use a broader commercial service.

Integration consideration: Pin the model version used in your evaluation and record the settings. Review the model’s license and dependencies for your use case. Use the current model-specific cost information and actual job usage; a price for one Replicate model does not apply across its marketplace.

7. Google Imagen: upscaling through Google Cloud

Google’s current Imagen upscaling documentation uses the preview model imagen-4.0-upscale-preview. It accepts scale factors of 2×, 3× or 4×, with the final image limited to 17 megapixels.

Why consider it: Your application already uses Google Cloud authentication, billing and image models, and the preview’s availability fits your deployment requirements.

Pricing consideration: Google’s pricing page lists Imagen 4 upscaling at $0.06 per image. Use that specific row rather than the older Imagen 1 rate or the price of generating a new image. Google Cloud Vision is a different service; it should not be presented as this upscaling API.

What does 4× upscaling actually mean?

Usually, 4× means four times the width and four times the height—not four times the total pixels.

This is why a small change to input size can exceed an API limit or move a job into a higher billing tier. Calculate width × height ÷ 1,000,000 for output megapixels before submitting a batch.

A concrete budget example: 100 product images

Assume 100 source images at 1,000 × 1,000 pixels, each enlarged to 4,000 × 4,000 once. Each output is 16 MP. These calculations use the documented units above; they are not quality-adjusted benchmark results.

Credits are not interchangeable between providers. Convert each estimate using your actual plan or credit purchase, include any minimum payment, then add rejected outputs and retries. The useful metric is cost per image you can publish, not just cost per successful API response.

How to choose without relying on a vendor’s best example

Use the same source files, target dimensions and acceptance criteria for every candidate. Include the difficult material your application actually sees.

Also record completion time, failure rate, output format and charged usage. Keep the original files and settings with each result so a team member can reproduce the comparison.

For catalog images, do not upscale unnecessary background detail before deciding on the final composition. If background cleanup is part of the job, compare the background-removal options for e-commerce and test the order of operations on hair, glass and other difficult edges.