Topaz Labs has released Wonder 4, a one-click model that combines image enlargement and restoration without separate controls for each repair. Its mixture-of-experts architecture assigns regions of an image to specialized networks, aiming to reduce compounded artifacts and simplify workflows that previously required several models.

Routing repairs by region

Conventional restoration workflows often chain a denoiser, deblurring model, sharpener, and upscaler. Each stage receives the previous model’s output, creating opportunities for noise, halos, invented texture, and other errors to accumulate. According to Topaz, Wonder 4 divides an image into thousands of regions and selects the relevant specialists for each one. A face, smooth sky, blurred subject, and blocky text can therefore receive different processing within the same pass.

In machine-learning terms, a mixture-of-experts model uses a routing layer to choose specialized subnetworks for each input patch. Only the selected experts process that patch. This local routing gives the model a way to sharpen detailed subjects while applying gentler treatment to clean backgrounds or low-texture areas.

Five tasks share one pass

Wonder 4 combines five restoration tasks that commonly require separate tools:

  • Upscaling from the original resolution
  • Noise reduction
  • Deblurring and sharpening
  • Compression-artifact removal
  • Facial-detail recovery

Wonder 3.5 focused on generative upscaling, while Wonder 4 expands the workflow to cover enlargement and repair together. The model provides no controls for tuning individual restoration operations. Host services still expose choices such as enlargement scale, output format, and export size.

From browser to API

At launch, Wonder 4 is available through Topaz’s web and mobile apps, Runware, and ComfyUI. Pricing and output limits vary by route.

HDR arrives alongside Wonder

The release also adds a separate Image HDR model to the web and mobile apps. It converts standard photos into HDR JPEGs with a wider highlight and contrast range. The HDR conversion runs as its own model, separate from Wonder 4’s enlargement and restoration pass.

Best-fit workflows and limits

  • Batch photography: Cleaning, sharpening, and enlarging a shoot without tuning every image.
  • Photo restoration: Processing scans or compressed files that contain blur, noise, and JPEG artifacts in the same frame.
  • Generative-image pipelines: Enlarging outputs from tools such as Midjourney or Sora while applying repairs selectively across the frame.
  • Product integrations: Adding restoration through Runware or a ComfyUI workflow with explicit scale and export limits.

Zero-setting restoration provides limited control over grain, texture, sharpening strength, and face reconstruction. Teams with strict fidelity requirements should inspect faces, text, film grain, and repeated patterns at full resolution, then compare the results with Wonder 3.5 or Topaz’s specialist models. API integrations should also validate pixel dimensions before submission and account for each provider’s billing unit.