Overview
Enlarging an image used to mean choosing which kind of failure you preferred: soft, or crunchy with halos. Gigapixel takes a different route by using models trained on pairs of small and large versions of real photographs, so instead of interpolating between the pixels it has, it predicts what the missing detail probably was. On the right material the result at six times the original size holds up at full screen, which is not something any traditional resampling method achieves.
This is the Pro build with every model included in the archive rather than downloaded on first use, and with the licensing settled ahead of time. Batch processing runs across a folder unattended, output carries no watermark at any size, and the program works fully offline since nothing is processed on a server.
Choosing the right model
The model list is the whole skill of using this program. The standard model is the general purpose one and handles most photographs. The high fidelity model is more conservative and keeps existing texture rather than inventing it, which is the right pick for material that is already sharp and simply needs to be larger. The low resolution model is trained for small web sized sources with compression damage and is where a thumbnail sized image goes.
Beyond those there are specific models for compressed sources, for images that are soft from motion or focus, and for line art and rendered graphics where photographic texture would be wrong. The face recovery pass runs separately and rebuilds facial detail at a higher weight than the surrounding frame, which is worth turning down or off on group shots where it can over sharpen background faces.
Working in the program
The interface is a split preview with the original on one side and the result on the other, synchronised as you pan. Because the processing is heavy, the preview renders only the visible region, which keeps the feedback quick while you compare models. Comparison view can show up to four models against each other at once on the same crop, which is the fastest way to decide.
Alongside the scale factor there are manual controls for sharpening, noise suppression and detail recovery. The defaults are usually correct and the most common mistake is pushing these upward, which produces the artificial look people associate with upscaled images. If the result looks synthetic, reduce rather than increase.
Batch work and integration
A folder can be queued with one model and one set of parameters applied to everything, with the output written to a separate directory and a naming pattern of your choosing. On a machine with a capable graphics card this runs at a reasonable pace unattended, and the queue survives being paused. Format support covers the common raster formats plus raw files from most cameras.
The program also installs as a plug-in for the major raster editors, so an enlargement can happen inside an existing edit rather than as a separate round trip. The result comes back as a new layer at the enlarged size with the original preserved underneath.
What you get
- Multiple trained enlargement models for different source material
- Separate face recovery pass with adjustable weight
- Enlargement up to six times the original dimensions
- Split preview with synchronised panning and a four way model comparison
- Manual sharpening, noise suppression and detail recovery controls
- Batch processing across a folder with custom output naming
- Raw file support alongside the common raster formats
- Plug-in integration with the major raster editors
- Runs entirely locally with no server processing and no watermark
Inside the archive
- Gigapixel AI Pro installer
- Complete model pack, preinstalled rather than downloaded
- Plug-in components for the supported host editors
- Readme covering model selection by source type
System requirements
| Operating system | Windows 10 version 22H2 or Windows 11, 64-bit |
|---|---|
| Processor | Intel or AMD x64 with AVX2 |
| Memory | 16 GB recommended for large enlargements |
| Graphics | 6 GB VRAM strongly recommended, processor fallback available but slow |
| Storage | 8 GB free including the model pack |
| Display | 1920 by 1080 |
Installing it
- Unpack the archive to a local folder.
- Suspend real time protection for the duration of the install.
- Run setup as administrator and let the model pack install with it rather than downloading it later.
- Launch once and set the processing device in preferences. Pick the graphics card if one is available.
- Run a single test image through before starting a batch, to confirm the device selection is working.
Mirrors
| Route | Region | Note | State |
|---|---|---|---|
| Direct, primary | Europe | No wait, resumable | Online |
| Direct, secondary | North America | No wait, resumable | Online |
| Torrent magnet | Global | Optional at this size | Online |
Release history
- Face recovery weight now adjustable per image rather than globally
- Fixed batch queue stalling on files with unusual colour profiles
- Reduced memory use on enlargements above 200 megapixels
- New model added for heavily compressed web sources
- Comparison view extended to four models at once
- Raw file handling reworked for recent camera bodies
- Plug-in components updated for the current host editor releases
Questions about this release
Does it need an internet connection?
No. All processing runs on your machine and the model pack is installed with the program rather than fetched later.
Which model should I use?
Standard for most photographs, high fidelity for already sharp material, low resolution for small web sources. Compare them in the four way view on the same crop.
Why does my result look artificial?
Almost always because sharpening or detail recovery has been raised above the defaults. Reduce them rather than increasing them.
Is there a watermark?
No. Output is clean at every size, in single and batch processing alike.
Comments
Batch across 300 product shots ran overnight on the GPU. No watermark, no fuss.
Model pack installed offline which is why I picked this build. Works with no network at all.
Comments are read before they appear. If a build stops working, say so here and it gets rebuilt rather than quietly left up.
Scanned family photos at four times up and the faces came back properly. Turned the face weight down a bit as the notes suggest.