Topaz AI: Pricing, Apps and API Options (2026)

Topaz AI is the shorthand most people use for the Topaz Labs apps that clean up and enlarge photos and video on your own computer: noise reduction, sharpening, and upscaling driven by neural networks trained on pairs of degraded and clean images. It is not a prompt box. You feed it a file you already have, and it gives you back a bigger or cleaner version of that same file. Understanding that single boundary saves a lot of wasted money, because half the people searching for Topaz AI actually need something else.

What is Topaz AI and what does it actually do?

Every Topaz AI app solves a variant of the same problem: your source file does not have enough good pixels in it. The software estimates what the missing detail should look like and paints it in. Three jobs cover almost all of it.

• Denoise. Separates sensor noise, the speckled grain you get from high ISO settings or dim light, from real texture. A good denoiser keeps the pores on a face and removes the colour blotches around them.
• Sharpen. Reverses blur. There is a real difference between motion blur, focus miss and lens softness, and the models are trained separately on each, which is why you pick a mode instead of dragging one slider.
• Upscale. Super-resolution, meaning the process of producing a larger image with more detail than simple interpolation would give you. Classic resizing averages neighbouring pixels and gets soft. A trained upscaler guesses plausible edges, hair and fabric weave, and can take a 1,000 pixel wide frame to 4,000 pixels that still look photographic.

The important technical fact about Topaz AI is where it runs. The models execute locally on your GPU, so a long video job can occupy your machine for hours, and the speed you get depends entirely on your hardware. The upside is that there is no per-file charge once you own the licence, and nothing leaves your drive. The downside is that you cannot call it from a server, you cannot batch a thousand jobs in parallel across cloud workers, and a laptop with integrated graphics will crawl.

It is also worth being honest about what “AI enhancement” means here. The detail Topaz AI adds is invented. It is a statistically likely reconstruction, not recovered information. On landscapes and textures that reads as a clean improvement. On small faces, licence plates and dense text, it reads as a confident hallucination, which matters if the footage is evidential or editorial.

Which Topaz AI apps exist, and what replaced Photo AI?

The naming has shifted more than once, which is why “is Topaz Photo AI discontinued” is one of the most common questions about Topaz AI. The short answer is that it was renamed, not killed. The older standalone tools, DeNoise AI and Sharpen AI, were merged into a single photo app, and that app is now sold as Topaz Photo. Gigapixel remains the dedicated upscaler with the most aggressive enlargement models. The video app, long known as Topaz Video AI, is now Topaz Video. Topaz Studio is the bundle that puts access to the apps in one place.

Practical consequences of that consolidation:

1. Guides written before the rename reference menus that no longer exist. Check the version number in the app before following a tutorial step by step.
2. Old licences for the retired standalone apps generally keep running, but they stop receiving new models. Model quality is where most of the year-over-year improvement lives, so an old licence slowly becomes a worse product than the thing you are comparing it against.
3. There is no official mobile version. The searches for a Topaz AI APK or an Android build lead to repackaged installers from third parties. Those are a credential and malware risk, and they cannot work anyway, because the desktop models expect desktop GPU memory.

How much does Topaz AI cost, and is there a free version?

Topaz sells per-app licences that include a window of free upgrades, plus a subscription that covers the suite. The exact figures move with promotions and with which apps you pick, so read them off the vendor’s own page rather than from any article, including this one. What does not change is the shape of the deal, and that shape is the thing to compare.

A Topaz AI licence is a fixed cost with unlimited local runs. Process ten files or ten thousand and you pay the same, minus electricity and your own time. That is excellent value if you have a large archive to work through and a capable GPU sitting idle. It is poor value if you process a handful of images a month, and it is simply the wrong tool if you need enhancement inside a product that other people use, because there is no hosted endpoint to call.

On the free question: there is a trial, and the trial is genuinely useful because you can run the real models on your own worst files before paying. It restricts export, typically with watermarks or blocked saves, so treat it as a quality test rather than a way to get work done. Anything advertising a permanently free full version of Topaz AI is a pirated build, and pirated media software is a well-worn malware delivery route.

The hidden cost most people miss

Time. A fifteen second 1080p clip upscaled to 4K with a heavy model is not a thirty second job on mid-range hardware, it can be tens of minutes, and a feature-length pass can run overnight. If your workflow needs a result in under a minute, or needs fifty results at once, the local model is the constraint and no licence tier fixes it.

What Topaz AI cannot do: generating new pixels from a prompt

Topaz AI starts from a file. If you do not have the shot, there is nothing to enhance. This is where a large share of the traffic around Topaz AI is really aimed at a different category of model: text-to-image and image-to-video generators, which produce an original frame or clip from a written description instead of improving an existing one.

The two categories are complements, not competitors, and the division of labour is clean:

Job	Right tool	Cost shape
Rescue a noisy archive photo	Local enhancer	Licence, unlimited runs
Enlarge a 2000px product shot to print size	Local upscaler	Licence, unlimited runs
Produce a product shot you never photographed	Image model via API	Per image
Animate a still into a short clip	Video model via API	Per second of output
Enhancement inside your own web app	Hosted API	Per request

A sensible 2026 pipeline uses both. Generate at a resolution the model handles well, then push the final selects through a local upscaler for the last stretch. Generating straight to 4K is usually slower and dearer than generating at 1080p and upscaling, and the upscaler often produces cleaner edges than the generator does at its maximum size.

Topaz AI plus a generation API: the practical workflow

On the generation side, the cost that matters is per output, and it varies by more than an order of magnitude between models. Unicorn API gives you one key and one bill across image, video, audio and chat models, priced in credits, which makes the arithmetic easy to do in advance. Credits are $0.01 each on a monthly plan (Starter $25, Creator $50, Studio $100, Pro $200), or $0.015 each as a top-up that never expires.

The models that pair best with a Topaz AI finishing pass:

• Nano Banana 2 Lite, Google’s cheapest Gemini image model, at 4 credits per image. Fast 1K images and edits, with up to 10 reference images. 1K output is exactly the size you would hand to an upscaler.
• GPT Image 1.5 from OpenAI, 3 credits at low quality, 4 at medium, 15 at high. Strong prompt adherence and legible text, with an optional transparent background and up to 4 reference images.
• Grok Imagine Video 1.5 from xAI, image-to-video only, 1 to 15 seconds, at 3 credits per second for 480p, 4 for 720p and 7 for 1080p. It can be busy at peak times.
• Grok Imagine Video 1.0, same per-second rate, but it accepts text-to-video as well and runs 6 to 30 seconds for longer takes.
• Vidu Q4 Image To Video, which goes up to 4K natively: 7.2 credits per second at 540p, 15.2 at 720p, 19.2 at 1080p, 30.4 at 2K and 62.4 at 4K. Vidu Q4 Reference To Video takes reference images and reference audio instead of a single start frame.

Note what that Vidu pricing implies. Generating 4K directly costs about 3x what 1080p costs per second. If you own a Topaz AI licence, generating at 1080p and upscaling locally is usually the cheaper route to a 4K deliverable, and you keep the ability to redo the upscale with a better model later without paying for the clip again.

Code: generate a frame, animate it, then enhance locally

Grab a key from /keys and export it as UNICORN_API_KEY. Media models use an asynchronous job pattern: you POST to /v1/jobs, get an id back, then poll GET /v1/jobs/{id} until the status is succeeded and read the file URL out of outputs[0].url.

The smallest possible image request:

curl -X POST https://api.unicornapi.net/v1/jobs \
  -H "Authorization: Bearer $UNICORN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano-banana-2-lite",
    "input": {
      "prompt": "A ceramic pour-over coffee dripper on a pale oak counter, morning window light, shallow depth of field, product photography",
      "aspect_ratio": "16:9"
    }
  }'

Then poll:

curl https://api.unicornapi.net/v1/jobs/JOB_ID \
  -H "Authorization: Bearer $UNICORN_API_KEY"

Here is the full two-stage pipeline in Python: generate a start frame, animate it to a 5 second 1080p clip, and print the URL you would then download and run through your local enhancer.

import os, time, requests

BASE = "https://api.unicornapi.net"
HEAD = {
    "Authorization": f"Bearer {os.environ['UNICORN_API_KEY']}",
    "Content-Type": "application/json",
}

def run_job(model, payload):
    r = requests.post(f"{BASE}/v1/jobs",
                      headers=HEAD,
                      json={"model": model, "input": payload},
                      timeout=60)
    r.raise_for_status()
    job_id = r.json()["id"]

    while True:
        job = requests.get(f"{BASE}/v1/jobs/{job_id}",
                           headers=HEAD, timeout=30).json()
        status = job["status"]
        if status == "succeeded":
            return job["outputs"][0]["url"]
        if status in ("failed", "canceled"):
            raise RuntimeError(f"{model} job {job_id}: {status}")
        time.sleep(3)

# 1. Start frame. 4 credits.
frame = run_job("nano-banana-2-lite", {
    "prompt": (
        "A ceramic pour-over coffee dripper on a pale oak counter, "
        "steam rising, morning window light, product photography"
    ),
    "aspect_ratio": "16:9",
})
print("frame:", frame)

# 2. Animate it. 5 seconds at 1080p = 35 credits.
clip = run_job("grok-imagine-video-1-5-preview", {
    "image": frame,
    "prompt": "Slow push in, steam drifting upward, light shifting gently",
    "duration": 5,
    "resolution": "1080p",
    "aspect_ratio": "16:9",
})
print("clip:", clip)

Total: 39 credits, about 39 cents on a plan. Download that clip, run your Topaz AI upscale pass on it, and you have a 4K asset without paying 4K generation rates. If you want the whole thing at 4K from the model instead, swap the second call for vidu-q4-image-to-video with "image_url", "duration": 5 and "resolution": "4k", which is 312 credits.

Check the price before you spend it

Never guess. POST /v1/models/price returns the exact cost of a request before it runs, with the same body you were about to send:

curl -X POST https://api.unicornapi.net/v1/models/price \
  -H "Authorization: Bearer $UNICORN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "vidu-q4-image-to-video",
    "input": {
      "prompt": "Slow push in, steam drifting upward",
      "image_url": "https://example.com/frame.png",
      "duration": 5,
      "resolution": "4k"
    }
  }'

Wire that into any user-facing feature and you can show a cost estimate before the user commits, which is the single best defence against surprise bills. Full request shapes are in the docs, and you can test prompts without writing code in the browser playground. If you drive your work from an editor, the MCP integration lets Cursor or Claude Code call the same models.

Writing the prompts and captions with a chat model

Batch work needs prompt variations, alt text and shot descriptions. Chat models use the OpenAI-compatible /v1/chat/completions endpoint on the same key, so you can generate the prompts that feed the image calls:

curl -X POST https://api.unicornapi.net/v1/chat/completions \
  -H "Authorization: Bearer $UNICORN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-4-argon",
    "messages": [{"role": "user", "content": "Write 5 product photography prompts for a ceramic coffee dripper, each one sentence, varying light and angle."}],
    "max_tokens": 400
  }'

Gemini 4 Argon costs 320 credits per million input tokens and 1,600 per million output tokens, so a prompt-writing call like that is a rounding error next to the images. Gemini 4 Argon OpenAI is the same model behind the OpenAI-shaped surface if your client expects that.

What a generated-then-enhanced clip costs in credits

Concrete numbers for a short social spot: one hero still plus three 5 second clips, generated at 1080p and finished locally.

Step	Model	Credits
Hero still, high quality	GPT Image 1.5 (high)	15
3 start frames	Nano Banana 2 Lite	12
3 clips, 5s at 1080p	Grok Imagine Video 1.5	105
Prompt and caption writing	Gemini 4 Argon	about 1
Total		133 (about $1.33 on a plan)

The local upscale adds no credit cost at all, which is the whole argument for owning a Topaz AI licence alongside an API account. The same job generated directly at 4K through Vidu Q4 would be 936 credits for the three clips. You are trading GPU hours for roughly 800 credits.

Where the trade flips: if you need fifty clips by tomorrow, your single GPU cannot upscale fifty clips overnight, and paying for native 4K generation is the sane choice. Volume and deadline decide, not principle.

Mistakes people make with Topaz AI

Patterns that come up repeatedly, worth checking against your own workflow:

• Upscaling before colour grading. Enhance first, then grade. A heavy grade applied after an upscale amplifies whatever artefacts the model introduced, and the sharpened edges fight the curves.
• Running the strongest model by default. Aggressive face recovery on a wide shot gives you plastic skin and shifted features. Match the model to the problem, and compare at 100% zoom, not fit-to-window.
• Enhancing a bad source instead of regenerating it. If the composition is wrong, no amount of Topaz AI processing fixes it. Four credits for a new frame from an image model beats an hour of rescue work.
• Expecting an API that is not there. Topaz AI is desktop software. If your feature needs to run for users on demand, build on a hosted endpoint and keep the desktop app for your own finishing passes.
• Chasing free downloads. The APK and cracked-installer results around Topaz AI are not the product, and the trial already answers the only question worth asking before you buy, which is whether the output quality is good enough on your files.
• Ignoring aspect ratio at generation time. Set aspect_ratio in the request. Cropping a 1:1 image to 16:9 afterwards throws away pixels you then ask an upscaler to reinvent.

Treated as an enhancer rather than a creative engine, Topaz AI holds up well: it removes noise, recovers sharpness and enlarges files with no marginal cost once you have paid for it. The gap it leaves is new material, and that gap is cheap to fill. One API key, a well-written prompt, 4 credits for a frame and 35 for a 1080p clip, then let the local GPU do the last mile.