Topaz Wonder 4: Image Upscaling, API Pricing and What Changed

Topaz released Wonder 4 on November 3, combining image upscaling, denoising, sharpening and other repairs in one automatic pass. It is available in Topaz for Web and Mobile, with a Runware API option for developers who want to add the model to an image-processing workflow.

The main change is how much correction the model decides for itself. That can make a damaged image easier to restore, but it also makes reviewing the generated detail more important.

What changed from Wonder 3.5

Topaz's November update describes a mixture-of-experts architecture that analyzes image regions and applies different enhancements to different areas. A noisy background and a blurred face can receive different treatment rather than sharing one fixed sequence of corrections.

The company says Wonder 4 combines upscaling, denoising, deblurring, compression-artifact removal and face recovery without settings to tune. Those are product claims; the announcement does not supply a standardized independent comparison that establishes a win for every kind of image.

The launch lists Web and Mobile availability. It does not announce Wonder 4 for Topaz's desktop applications. That is a useful distinction for existing customers deciding whether their current workflow has actually changed.

Topaz also introduced Image HDR in the same update. It converts standard images into HDR JPEGs using a gain map, with the enhanced display experience requiring an HDR-compatible screen. It is a separate capability from increasing resolution.

Wonder 4 API pricing on Runware

Runware's model reference exposes Wonder 4 under topazlabs:wonder@4 using the upscale task type. The scale factor runs from 1 to 6, and output must stay within both 256 megapixels and 32,000 pixels per side.

Its listed rates depend on output size:

Output size	Listed price per image
1, 4 or 8 megapixels	$0.0428
16 megapixels	$0.0856
32 megapixels	$0.171
64 megapixels	$0.342
256 megapixels	$1.37

These are Runware's API rates, not Topaz subscription prices. The reference shows example completion times, but they vary and should not be treated as a latency guarantee.

Resolution changes compound quickly. Doubling width and height produces four times the pixels. A developer deciding between 2x and 4x therefore needs to consider the intended output, rather than automatically requesting the largest image.

Enhancement does not always require enlargement

Runware documents a useful smaller operation: setting the factor to 1 applies enhancement while retaining the original dimensions. That fits images already large enough for a page or listing but damaged by noise or compression.

This can simplify a publishing workflow. Instead of enlarging an image and then resizing it back down, the application can ask for cleanup at its target size. The result still needs inspection, especially where faces, logos or small text carry meaning.

For transparent assets, Runware's alpha-channel guide says transparent PNG input can retain transparency when returned as PNG. Choosing JPEG discards that property. Output format remains part of the task even when the model handles the visual corrections automatically.

Early reactions are mixed, and fidelity is the real test

In the official release discussion, one early user preferred Wonder 3.5's results, while another praised Wonder 4 and asked about desktop availability. That is a small set of initial impressions, not evidence of a broad quality verdict.

The thread also records a separate issue with strength controls for Wonder 3 and 3.5, which a Topaz representative said the team would fix. It should not be confused with Wonder 4's intentionally automatic workflow.

A cleaner-looking image is not necessarily a more faithful record. Generative enhancement can create plausible detail where the source does not contain enough information to recover it confidently. For a product listing, the comparison should include labels, colors and surface texture. For an archival image, retaining the untouched original is particularly useful.

Wonder 4 reduces the number of adjustments needed to attempt a restoration. The lasting value will come from how often that simpler process produces a result people can accept without spending the saved time checking and repairing details afterward.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn model launches and product updates into clear stories that developers and founders want to read and share.

Book a 30-minute call.