Three upscale APIs, three different ways to pay. This comparison lines up Topaz Labs credits per megapixel, Picsart's shared credit pool and Google's preview Imagen upscaler, with output limits, request shapes and a browser test bench so you can judge quality before writing code.
You have 3,000 product photos shot at 1,200 pixels wide, and a print partner asking for 4,800. Or a user-uploaded avatar that turns into a mosaic the moment it lands on a retina screen. Sooner or later every image pipeline needs an upscale call, and three names keep surfacing in developer threads: Topaz Labs, Picsart and Google. They all sell the same promise, more pixels without the mush, but through very different APIs, billing models and limits. This comparison is built from the public documentation and pricing pages of all three, and it focuses on the numbers that decide whether your pipeline stays cheap or quietly becomes a monthly surprise.
💡 Quick read: Topaz bills per output megapixel across three quality tiers. Picsart deducts a per-call rate from a credit pool shared across its tools. Google's Imagen upscale is a preview model with a hard cap of 17 megapixels on the output.
A normal resize stretches the pixels you already have. The result is bigger and blurrier. An upscale API sends your image to a trained model that predicts the detail the original never captured: the edge of a brick, the weave of a sweater, the line where hair meets a background. You send a small file and get back a larger one that looks like it was shot that way.
Choosing between providers is really a negotiation between three things: how faithful the output stays to the source, how large the output is allowed to get, and what each output pixel costs you.
Two families dominate the market. Precision models sharpen what is already there. They are conservative, they rarely invent anything, and they suit product shots, documents and any image a customer might compare against the real object. Generative models treat missing detail as a painting problem. They add plausible texture, which can look spectacular on a soft landscape and unsettling on a face, because a skin pattern that never existed is still a fabrication.
💡 Rule of thumb: if the photo is evidence of a real thing, such as a listing, an ID or a repair claim, pick precision. If the photo is mood, such as a hero banner or a poster, generative is fair game.
Almost every limit and every bill in this comparison is tied to the output size, not the file you upload. Doubling the scale factor quadruples the pixel count, so a harmless looking 4x setting asks for four times the pixels of a 2x one.
| Input size | Factor | Output size | Output megapixels |
|---|---|---|---|
| 1,000 × 1,000 | 2x | 2,000 × 2,000 | 4 MP |
| 1,200 × 800 | 4x | 4,800 × 3,200 | 15.4 MP |
| 1,000 × 1,000 | 4x | 4,000 × 4,000 | 16 MP |
| 2,000 × 2,000 | 4x | 8,000 × 8,000 | 64 MP |
Keep this table next to the pricing sheets below. A 16 megapixel result is already close to an A3 print at 300 dpi, and a 64 megapixel one is where credit costs and size caps start to bite.
Topaz Labs built its reputation on desktop photo software, and its developer API follows the same philosophy: pay for what you process. There are no subscriptions or minimums, credits are deducted automatically when a job finishes, and new accounts get free trial credits to test with.
Pricing is linear in output megapixels. The documentation puts it plainly: one credit buys a fixed number of output megapixels, and the number depends on the tier.
| Use case | Topaz model | Output MP per credit |
|---|---|---|
| Precision upscale | Gigapixel | 24 |
| Generative upscale | Wonder | 4 |
| Creative upscale | Bloom | 2 |
| Sharpen | Sharpen | 24 (GAN) or 20 (generative) |
| Denoise | Denoise | 24 |
Run the arithmetic on a 16 megapixel output. On the precision tier it costs less than one credit. On Wonder it costs 4 credits. On Bloom it costs 8. Same photo, same size, a spread of more than eight to one in price depending on how much creative freedom you buy. That spread is the single most useful fact in this article for anyone building a budget.
Topaz is the natural pick when the destination is paper or a large display and the source is a real photograph. The precision tier is cheap enough to run on whole catalogues, and the separate sharpen and denoise endpoints let you clean a file before or after enlarging it instead of paying for a generative pass that does everything at once.
If you want to see the look of that engine before touching the API, Topaz Image Upscale is available on PicassoIA with 2x, 4x and 6x factors and five specialised modes, which makes it a handy visual reference.
Picsart approaches the problem from the creative platform side. Upscaling is one tool inside a larger API catalogue that also includes background removal, effects and generation, so the pitch is one account for everything.
The endpoint is a single POST, and the required parameters are refreshingly short: a URL pointing at your image, an integer upscale factor (the documentation example uses 3) and the output format (the example uses PNG). Authentication runs through a custom request header that carries your Picsart credential.
curl -X POST "https://api.picsart.io/tools/1.0/upscale/ultra" \
-H "accept: application/json" \
-H "Content-Type: multipart/form-data" \
-F "image_url=https://example.com/photo.jpg" \
-F "upscale_factor=3" \
-F "format=PNG"
Add the Picsart authentication header with your own credential before running it. The factor limits, the size limits and the exact processing mode are spelled out on Picsart's reference page, so check it before you design retries and timeouts.
Picsart credits are described as the universal unit across every Picsart API and model, and each call deducts its own rate. At the time of writing, plans run from 1,000 credits for $7.50 one-time (or $5.00 a month) up to 20,000 credits for $100 one-time (or $90 a month). Confirm those figures on the live pricing page before you budget around them.
Two practical consequences follow. First, a shared pool is convenient when one product uses upscaling, background removal and generation together, because you top up once. Second, a per-call rate can behave very differently from per-megapixel billing: it can look generous on large files and expensive on thumbnails, so test it against your real size distribution.
Picsart's API platform also lists third-party upscalers as selectable models, including Topaz Image Upscale and the two Recraft options. When we checked, the Topaz model page said live pricing was not published yet. If you want to eyeball the Recraft look first, both are on PicassoIA: Recraft Crisp Upscale for faithful sharpening and Recraft Creative Upscale for added depth and detail.
Google's offering lives inside Vertex AI, which makes it either the easiest or the most awkward option depending on where your infrastructure already sits.
The model is named imagen-4.0-upscale-preview, and the word preview matters. It accepts three factors: x2, x3 and x4. The final output must not exceed 17 megapixels, which is the limit most likely to trip you up. Input can be sent as base64 or as a Cloud Storage URI, and output comes back as PNG by default or JPEG, where compression quality runs from 0 to 100 with a default of 75.
Do the math before you call. At x4, 17 megapixels means your input can be roughly 1,030 pixels square, no more. At x2 you get about 2,060 pixels square. A 2,000 pixel photo at x4 would land at 64 megapixels and break the cap, so you would downscale first or choose x2.
The upscale documentation page carries no price. Third-party pricing summaries list about $0.06 per upscaled image on Vertex AI, which is attractive for mid-size volumes, but confirm it on the Google Cloud pricing page because preview pricing can change. Preview status also means model names and behaviour may shift, so pin your version, log the model name with every result and keep a fallback.
The setup cost is real too: a Google Cloud project, IAM roles and ideally a Cloud Storage bucket. If you already live there, it is a small step. If you do not, it is the heaviest onboarding of the three.
To compare the look of Google's engine without any cloud setup, Google Upscaler on PicassoIA offers 2x and 4x output with an adjustable compression quality from 1 to 100.
Here is how the three line up on the facts that were verifiable from public documentation.
| Topaz Labs API | Picsart Ultra Upscale | Google Imagen Upscale | |
|---|---|---|---|
| Billing unit | Credits per output megapixel | Credits, fixed rate per call | Per image on Vertex AI |
| Quality tiers | Precision, generative, creative | One ultra upscale endpoint, plus hosted third-party models | One upscale model |
| Scale factors | Set per model | Integer factor (docs example: 3) | x2, x3, x4 |
| Output limit | Priced by size, no flat cap in the pricing docs | See the reference page | 17 MP maximum |
| Input | Image upload | Image URL | Base64 or Cloud Storage URI |
| Maturity | Pay as you go, trial credits | Credit plans, shared pool | Preview model |
| Best for | Print and photography | Mixed creative toolkits | Teams already on Google Cloud |
💡 Honest caveat: this comparison reads documentation and pricing pages. It is not a lab benchmark. Image quality is subjective and source dependent, so run twenty of your own files through each provider before you commit.
Before you write integration code, judge the look of the output with your own files. PicassoIA hosts Topaz Image Upscale and Google Upscaler in the browser, so you can test quality in minutes. At the time of writing, PicassoIA's own developer API lists image, image editor and video models, so treat this as a visual test bench rather than a replacement for the vendor APIs above.
| Source | Model and mode | Factor | Notes |
|---|---|---|---|
| Product photo | Topaz Standard V2 | 2x | Safe default for listings |
| Old low-resolution scan | Topaz Low Resolution V2 | 4x | Expect to retouch faces |
| Portrait | Topaz Standard V2 with facial sharpening | 2x | Creativity at 0 |
| Poster or infographic | Topaz Text Refine | 4x | Check small type |
| Quick web enlargement | Google Upscaler | x2 | Quality 80 is plenty |
Other upscalers are worth a quick test on the same file: Real-ESRGAN for a free 4x baseline, Clarity Pro Upscaler for photorealistic detail, Crystal Upscaler for portraits, P Image Upscale for speed and Increase Resolution for up to 4x.
💡 Cheap insurance: store the model name, factor and settings next to every upscaled file. When a customer complains about one image six months later, you can reproduce it.
Reading a comparison table only goes so far. The fastest way to settle the debate for your own images is to run them. Open Topaz Image Upscale on Picasso IA, push one of your softest files to 4x, then run the same file through Google Upscaler and compare. Within ten minutes you will know which look suits your catalogue, and which API deserves a real integration.
While you are there, try the other tools on the platform too: generate a fresh image from a text prompt, restore a damaged old photo, or push a favourite shot through an upscaler and print it. Browse everything at picassoia.com/en/all-models and make your own images with Picasso IA today.
