Topaz Image Upscale: which upscaler to pick, and when

Open the upscalers on the models page and you will find four from Topaz alone, plus several from other makers. They sound alike and behave very differently. Topaz Image Upscale is the plain one, and that is its whole point: its description calls it non-generative, which means it makes your picture bigger and cleaner without painting new things into it.

What it takes and what it gives back

One input: an image, as JPEG, PNG, TIFF or WebP. There is no prompt, because there is nothing to describe. You choose the size of the result with a target resolution of 2K, 4K or 8K (the default is 4K), and the source aspect ratio is kept, so nothing gets cropped or stretched. You can also pick JPEG or PNG for the output; JPEG is the default.

The one setting worth learning is model, under Advanced. It is not a strength slider but a choice of engine, each tuned for a different kind of picture:

• Standard V2 (default): a general-purpose balance of detail, sharpness and noise. Start here.
• Upscale High Fidelity V3: for images that are already high resolution and need a careful lift.
• Low Resolution V2: recovery-focused, for small, low-resolution files.
• CGI: precise upscaling for artwork, illustrations and CG renders.
• Text Refine: clean results for text, graphics and shapes.

How it compares with its neighbours

The real decision is not about size, it is about how much freedom you give the tool. Three nearby models sit on that scale.

Topaz Image Upscale Creative is described as adding new detail to AI-generated images, and it has a prompt and a creativity setting. That is the one for a soft render that needs texture invented for it. Topaz Image Upscale Generative goes further and repairs as well as enlarges, with a list of generative engines to choose from. And Topaz Transparent Image Upscale exists for one job: transparent PNGs, with the alpha channel kept intact.

The plain upscaler is the right pick when the picture is evidence of something: a product shot where the stitching has to be the real stitching, a portrait where the face has to stay the same face, a scan of a document. For the longer version of this split, read faithful or creative AI upscalers.

A short checklist

1. 01Must the result stay true to the original? Use Topaz Image Upscale.
2. 02Is the source an AI render that looks flat and needs texture added? Try the Creative one.
3. 03Is the source damaged, blurry or badly compressed, and are you happy for the tool to repair it? Try the Generative one.
4. 04Is it a transparent PNG, such as a logo or a cut-out? Use the Transparent one, which keeps the alpha channel intact.
5. 05Is it text, a diagram or flat graphics? Stay with this model and set model to Text Refine.

Lifehacks

Match the engine to the subject before you touch anything else. Artwork and renders belong with CGI, a tiny thumbnail with Low Resolution V2, a page of text with Text Refine. Changing the model costs one more run, so it is the cheapest experiment you have. If the source is an old print, restore it first, as described in restoring and colorizing old photos, and enlarge last.

Where it falls short

Non-generative also means it cannot rescue what was never captured. A face that is a handful of pixels wide stays a handful of pixels, only larger and smoother, and a blurry photo stays blurry. If you need the missing detail to appear, you want a generative neighbour, and you accept that the new detail is a plausible guess rather than the truth. Output size is also capped by the three targets, so there is no custom number. Pick the file in the studio, open Advanced, and run it.