Wonder 4 Fixes a Photo for Half a Cent a Megapixel

Topaz Labs released two image models on November 3. Wonder 4 restores photographs: denoise, deblur, sharpen and upscale, in one click, up to six times the original. Image HDR converts standard images into high dynamic range. Both are sold separately, and both arrived about six weeks after Adobe completed its takeover of the company it agreed to buy in June.
That corporate detail is not gossip. Topaz spent a decade as the expensive utility people bought once and kept for years, the tool you opened when a client sent a photo taken in a dark room. Being inside Adobe changes who it is for and what it costs. The models are not bundled into Creative Cloud; they are still sold on their own.
The number that matters
On Runware, Wonder 4 costs $0.0428 for an output up to 8 megapixels, and $1.37 at 256 megapixels. The first figure is the one to sit with.
Forty-two thousandths of a dollar for eight megapixels of restored image works out to about half a cent per megapixel. For comparison, Krea 3 landed on Runware on November 4 at $0.025 per megapixel, and Krea 3 generates images. Nano Banana Pro lists image output at $0.00012 per unit on OpenRouter's API, which sounds cheaper until you notice it is billed per image token rather than per pixel and the totals converge somewhere far above a cent per final frame at full resolution.
The useful comparison runs between restore and regenerate rather than between two generators. Fixing an existing photograph costs a fraction of what generating a replacement costs, and the output is the actual photograph, with the actual person in it, at the actual moment it was taken.
That asymmetry explains a lot of what professional retouchers actually do now. The default move on a bad frame used to be reshoot or repaint. The default move is increasingly to run it through a restoration model, because it is cheaper than a regeneration and it does not invent a new face.
Run the arithmetic on a real job and the gap widens. A catalogue of 2,000 product photos at 12 megapixels each is 24 gigapixels of output. At Wonder 4's rate that lands under $130 for the whole catalogue, at a quality that used to require a retoucher and a week. Regenerating those same 2,000 shots with an image model, at the prices currently posted for a 4K generation, costs several thousand dollars and returns pictures of products that do not exist.
What six times actually means
Upscaling by six is a strange claim until you separate resolution from information. A 1,000-pixel-wide photo upscaled six times is 6,000 pixels wide and contains the detail of a 1,000-pixel photo, plus whatever the model plausibly invents. Whether that is acceptable depends entirely on the subject.
For product shots, architectural photography and anything with hard edges and known materials, a good upscaler is close to a free lunch. For faces with fine skin texture, the model starts making decisions about pores, and those decisions are guesses. Topaz's pitch is that it reproduces texture rather than smoothing it, which is the right pitch and also the hardest claim to verify without a reference.

The four operations being one click is the practical part. Denoise, deblur, sharpen and upscale interact: sharpening after denoising amplifies whatever the denoiser left behind, and upscaling before sharpening produces a different artifact profile than the reverse. A model that handles the sequence internally can avoid the compounding that a manual workflow creates, and most people's manual workflow is two sliders and hope.
There is a second-order effect on the review loop. Sharpen and denoise are the two operations most likely to be overdone, because they look like improvement at 100 percent zoom and like plastic at print size. A one-click model with a fixed recipe removes the temptation to push a slider too far, which is a quality argument rather than a speed one, and it is the argument Topaz has been making since before AI was in the name.
Image HDR is the quieter release
Converting a standard image to HDR is a smaller claim with a wider consequence. HDR is not "more saturated"; it is a wider range between the darkest and brightest parts of an image that can still be displayed as intended. SDR photos clip highlights and crush shadows by construction, and the information is gone.
A model that produces an HDR version is inventing the missing range. On a photo of a sunset it will look spectacular. On anything where the highlight carries meaning, such as a product label or a signature, it is manufacturing detail that was never captured. That is the same trade as AI fill, in a format that most viewers cannot inspect.
Both models are the kind of tool that gets used without comment and then shows up in the aggregate. A photographer's portfolio in 2027 will be mostly restored, upscaled and HDR-converted, and nobody will label it that way, because there is no accepted way to label it.
Why Adobe ownership complicates the pitch
Adobe agreed to acquire Topaz in June and completed the handover in September. In the same period it announced new AI features across Photoshop, Premiere and Lightroom, including a Relight tool, a generative restyle with a slider, and an Open Eyes feature that replaces a blink using another photo of the same person.
Topaz's models sit next to those features without being inside them. For a buyer that means a decision: pay Adobe for the suite, pay Topaz separately for the specialist models, or use a hosted endpoint on Runware and pay per image. Three routes to the same output, with very different cost structures.
The per-image route is the one that scales for teams. A retoucher doing a hundred product photos a month does not want a subscription per seat and does not want to think about GPU memory. They want a price per image and an API.
The practical takeaway
If your workflow involves a lot of existing photographs, restoration is now the cheapest operation in the stack. Running a batch of marginal frames through a model at half a cent per megapixel is cheaper than the coffee you drink while deciding whether they are worth keeping.
If your workflow involves generating new images, the interesting question is whether you still need to. A great many briefing documents ask for "a clean shot of the product on a neutral background", which is a restoration problem wearing a generation costume. The model that costs a fifth as much and returns the actual object is usually the better answer, and it is much easier to defend in a review.
The one thing to keep straight is what the model is allowed to invent. Upscaling invents texture, HDR invents range, deblurring invents edges. All three are fine for illustration and all three are a problem for evidence. Knowing which job you are doing is the entire skill.
Related articles
The Jensen Discount: What Nvidia Loses When Its Founder Walks Out
Nvidia lost roughly $600 billion in market value when its founder stepped down. The earnings did not change. The multiple did, and that says where the value really lived.
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.