Topaz Labs Wonder 3.5 on the Yotta AI Gateway: Image Upscaling API, Pricing, and How to Call It

Topaz’s generative image upscaler, on the Gateway, one flat price up to 1440p output.

Topaz Labs is best known for Gigapixel and Topaz Photo, the desktop tools photographers have used for years. Wonder is the generative upscaling model family in Topaz’s catalog, and Wonder 3.5 is the current version. It is now live on the Yotta AI Gateway as the first and so far only model in the Image-Upscale category, alongside Topaz’s Starlight video upscalers in Video-Enhancement.

This post covers what Wonder 3.5 does, what it costs on the Gateway, how the three enhancement levels work, and the request format, so you can go from a low-resolution image to a print-ready one with one API call.

What Wonder 3.5 is

Wonder 3.5 is a generative upscaler. Instead of stretching pixels, it rebuilds the image at the target size, adding detail the source does not have. Topaz positions it for a wide range of inputs, from high-quality photos that need to go bigger to heavily degraded or compressed images that need to be rescued.

Topaz’s own description of what changed from Wonder 3:

• More detail and more natural realism across image types.
• Stronger enhancement on low-resolution and compressed photos, with better text clarity.
• Fewer repetitive patterns and better handling of visual noise, the two artifacts generative upscalers are known for.
• Better precision on tables, charts, and other structured graphics.
• Three enhancement levels, Low, Medium, and High, so you can pick how much the model is allowed to invent.

That last point is the one developers care about. Generative upscaling is a dial between faithful and impressive. Low keeps the model close to the source. High lets it add the most detail. Medium is the middle. On the Gateway the default is High.

What it costs on the Gateway

Wonder 3.5 is billed per image, and the rate depends on the output size. The console lists four tiers.

Three of the four tiers cost the same. Anything up to 1440p output is $0.0492 per image, and 4K is double that. So for most web and product work there is one price, and the only decision is whether the output needs to be 4K.

For scale: 100 product images upscaled to 1080p is $4.92. The same 100 images to 4K is $9.84. A thousand thumbnails to 720p is $49.20.

Your request specifies exact output dimensions in pixels rather than a named tier, so the console table is the reference for what each size range costs. The Gateway bills per image regardless of enhancement level; Low, Medium, and High cost the same.

The three enhancement levels

The enhancement_strength parameter takes low, medium, or high, and defaults to high.

Use high when the source is small, soft, or compressed and you want the model to do real reconstruction. This is the right setting for old photos, screenshots, and anything that was saved at low quality.

Use low when the source is already clean and you mainly need it larger. Product photos, renders, and anything where the client will compare the output to the original and notice invented detail.

Use medium when you are not sure. It is a reasonable default for batch jobs with mixed sources.

Since all three cost the same, the cheapest way to find the right level for your images is to run the same source at all three once and look.

How to call Wonder 3.5 on the Gateway

Wonder 3.5 uses the image-upscale endpoint. Submit a job, get a request ID, poll until it completes, read the output URL. Same pattern as every video and image model on the Gateway.

Endpoint, submit:

POST https://gateway.yottalabs.ai/api/maas/image-upscale/generations

Endpoint, poll:

GET https://gateway.yottalabs.ai/api/maas/image-upscale/generations/{request_id}

Authentication is an X-API-KEY header with your Gateway key. A successful submit returns code 10000 and a request_id. Poll until status is completed, then read output_url. Status can also come back as failed or cancelled.

Request parameters, all required:

• source_url. An HTTPS URL to a JPEG or PNG. The path has to end in .jpeg, .jpg, or .png.
• input_width and input_height. The source dimensions in pixels, each from 1 to 32,000, with the product no more than 256,000,000.
• output_width and output_height. The target dimensions, same limits, and the output pixel count has to be at least the input pixel count. You cannot use Wonder to shrink an image.
• enhancement_strength. low, medium, or high. Default high.
• output_format. jpeg or png. Default jpeg.

A request that takes a 640 by 427 source to 3840 by 2562, which is a 6x upscale:

{"model": "wonder-3.5", "parameters": {"source_url": "https://example.com/input.jpeg", "input_width": 640, "input_height": 427, "output_width": 3840, "output_height": 2562, "enhancement_strength": "high", "output_format": "jpeg"}, "providers": ["topaz"]}

The input dimensions are required fields, so read them from the file before you submit. Keep the aspect ratio the same between input and output unless you want the model to stretch.

The console model page for Wonder 3.5 has a full Python client with submit, poll, and a generate-and-wait wrapper. Open the API tab on the Wonder 3.5 page in the console to copy it.

Where it fits in a Gateway pipeline

The Gateway has two text-to-image models, Seedream 4.5 and Seedream 5.0 Lite. Wonder 3.5 is the step after, for when the output needs to be bigger than what the generator produced. Generate the image, take the output URL, send it to Wonder with the dimensions you need, and you have a 4K or print-size version on the same key.

The same applies to video thumbnails and poster frames. If you are already generating video with Seedance, Wan, Kling, or PixVerse on the Gateway, a frame pulled from that clip can go through Wonder for a hero image, and the clip itself can go through Starlight for the 4K version. One vendor, one bill.

Outside generation, the obvious use is catalogs. E-commerce teams with thousands of supplier images at inconsistent sizes can run them through Wonder at Low or Medium strength and get a uniform set at a fixed cost per image.

FAQ

Is Wonder 3.5 the same model as in Topaz Photo and Gigapixel?

Wonder is Topaz’s generative upscaling model family, and Wonder 3.5 is the current version. Topaz ships its models across its desktop apps, web apps, and API, with availability varying by product. On the Gateway you are calling Wonder 3.5 through the API.

Can I upscale more than 4x?

Yes. The limit is on dimensions, not on the ratio: each side up to 32,000 pixels and the output under 256 megapixels. The console example is a 6x upscale. Very large ratios ask more of the model, so test at your target ratio before running a batch.

Does enhancement level change the price?

No. Low, Medium, and High bill the same. Price depends only on output size.

What formats does it accept?

JPEG and PNG in, delivered by HTTPS URL. JPEG or PNG out.

Is there a face recovery option?

Wonder 3.5 handles faces as part of its general enhancement. There is no separate face parameter to set; the enhancement level is the only control.

What about the Starlight models?

Starlight Precise 2.6 and Starlight Fast 3 are Topaz’s video upscalers, live in the Gateway’s Video-Enhancement category. The comparison post covers which to use and what a clip costs.

Bottom line

Wonder 3.5 gives you Topaz’s generative image upscaler for $0.0492 an image at anything up to 1440p and $0.0984 at 4K, with three enhancement levels at the same price. It sits next to the image and video generation models on the Gateway, so the generate-then-upscale flow is one key and two calls.

Get an API key in the Yotta console, open the Wonder 3.5 page, and run one image through all three enhancement levels. That is the fastest way to see which one your sources need.