Topaz Gigapixel AI Pro
Version 8.4.4 · Topaz Labs · Windows 10 / 11 (64-bit) · ranked #44 in the library
An image enlarger that reconstructs detail rather than interpolating it, taking a small or soft source up to six times its size with edges and texture that survive the trip.
Read the installation section below before you start. The order of the steps matters on this release.
About this release
Overview
Traditional enlargement averages between the pixels you already have, which is why an upscaled image looks soft no matter which algorithm you pick. Gigapixel works differently: it runs the image through models trained on how detail actually behaves, and reconstructs plausible edges, texture and structure at the target size. On a photograph that means hair, fabric weave, foliage and brickwork come out looking like themselves rather than like a blur with sharpening applied.
Model choice matters more than any slider here. There are separate models for standard photographs, for compressed low quality sources, for images that are already sharp and just need size, for faces, for artwork and illustration, and for scanned material. Picking the wrong one produces a technically large image that looks wrong, and picking the right one is most of the job. The comparison view renders several models side by side on a crop so the decision takes seconds.
The professional build adds the parts that matter at volume: batch processing across a folder with per file model selection, command line operation so an enlargement step can sit inside an existing pipeline, and the higher output limits. Face recovery runs as a separate pass over detected faces, which is what stops a portrait enlargement from producing the waxy plastic look that a single pass model gives you.
What it does
Feature breakdown
Detail reconstruction rather than interpolation
Models trained on real image structure rebuild edges and texture at the target size instead of averaging between existing pixels.
Model set for different sources
Separate models for standard photographs, heavily compressed sources, already sharp images, artwork, illustration and scanned material.
Face recovery pass
Detected faces get a dedicated pass, which is the difference between a convincing portrait enlargement and a waxy one.
Side by side model comparison
Render several models on the same crop simultaneously and pick by eye rather than by guesswork.
Batch processing with per file settings
Queue a whole folder, assign a model and scale per file or per group, and leave it running.
Command line operation
Drive the enlargement from a script so it can sit inside an existing production pipeline rather than being a manual step.
Changed in this build
Version 8.4.4
- New high fidelity model producing noticeably better results on textile and foliage texture.
- Face recovery improved on profile and partially obscured faces, which the previous version often skipped.
- Faster processing on cards with limited video memory through better tiling.
- Batch queue rebuilt so a failed file no longer stops the rest of the queue.
- Output settings extended with more colour space and bit depth options for print destinations.
Installation
Follow these in order
- Go offline before installing.
- Extract the archive.
- Disable real time protection.
- Run the installer as administrator and complete setup without launching the application.
- Apply the included licensing step following the readme order.
- Open the application, confirm the professional features are available, then restore protection.
Release notes
From whoever packed it
- The models are bundled in this build so nothing is downloaded on first run.
- Block the application in your firewall. An online check reverts it to trial behaviour.
- Very large outputs need a lot of video memory. If processing fails, reduce the tile size in preferences rather than the output resolution.
System requirements
Minimum and recommended
| Operating system | Windows 10 version 22H2 or Windows 11, 64-bit |
|---|---|
| Processor | Modern x64 processor with AVX2 support |
| Memory | 16 GB minimum, 32 GB for very large output sizes |
| Graphics | GPU with 6 GB VRAM or more strongly recommended, processing works on CPU but is far slower |
| Storage | 6 GB for the application and the bundled model set |
Questions about this release
Answered before you ask
- How far can it enlarge?
- Up to six times the source dimensions in one pass. Beyond that, results depend entirely on how much real detail the source held.
- Does it work without a graphics card?
- Yes, on the processor, but expect minutes per image rather than seconds.
- Is it the same as the video upscaler?
- No, that is a separate product. This one is for still images.
Download
Pick a mirror