Overview
Topaz Video AI does one category of job and does it better than the alternatives: making old, small or damaged footage usable at a modern resolution. Standard definition archive material, phone clips, interlaced broadcast tape and compressed downloads all go in, and something you can put on a timeline comes out.
The model set covers the distinct problems separately rather than pretending one pass fixes everything. Enhancement models reconstruct detail while upscaling. Denoise models clear sensor grain and compression artefacts without smearing edges. Deinterlacing handles broadcast sources properly rather than throwing away a field. Frame interpolation generates intermediate frames for slow motion or for converting frame rates cleanly. Stabilisation removes camera shake with a rolling shutter correction pass.
This build includes every model weight in the package, so nothing is fetched on first use, and the Pro configuration removes the export watermark and unlocks batch processing across a folder of clips. Output goes to ProRes, DNxHR, H.264, HEVC or an image sequence.
What it does well
Upscaling with detail
Enhancement models reconstruct plausible detail while scaling, rather than interpolating a soft frame up to a larger soft frame.
Problem specific models
Separate models for denoise, compression artefacts, deinterlacing, sharpening and restoration, chained in one pass.
Frame interpolation
Generate intermediate frames for slow motion or frame rate conversion without the judder of a naive blend.
Stabilisation
Camera shake removal with rolling shutter correction and an auto crop that keeps the frame filled.
Batch processing
Queue a folder of clips with one recipe, with preview comparison before committing the render.
Changes in this build
- New enhancement model with better fine texture retention on faces.
- Faster processing on current GPUs with lower VRAM use.
- Improved deinterlacing on sources with mixed field order.
- Preview comparison rebuilt with a synchronised split view.
- Batch queue now retains per-clip settings rather than one global recipe.
What is in the package
- Video AI Pro installer, 64-bit
- Complete model weight set
- Export codec components for ProRes and DNxHR
- Batch queue and preset library
- Sample clips for model comparison
System requirements
| Processor | Intel or AMD 64-bit, 6 cores or more recommended |
| Memory | 16 GB minimum, 32 GB for 4K sources |
| Graphics | GPU with 8 GB VRAM strongly recommended, 6 GB minimum |
| Storage | 20 GB free plus output space, which is substantial |
| Display | 1920 x 1080 minimum |
| System | Windows 10 22H2 or Windows 11, 64-bit |
Installation
- Extract the archive to a local drive.
- Run the installer and let the model weights unpack completely.
- Apply the configuration step before first launch.
- Open a sample clip and run a short preview to confirm GPU processing.
- Set the output folder to a drive with plenty of space.
Before you start
Output files are large, especially in ProRes. Budget disk space before a batch.
Processing time scales with source resolution and model choice, test on a short segment first.
Chaining too many models softens the result, usually two is the right number.
Questions about this title
Are the models included?
Yes, the full weight set ships in the package with nothing fetched later.
Is the watermark removed?
Yes, exports are clean in the Pro configuration.
Does it need a strong GPU?
Yes. It runs on CPU but at a speed nobody would accept.
What formats does it export?
ProRes, DNxHR, H.264, HEVC and image sequences.
About this listing
This entry was checked on a clean install of Windows 10 / 11 (64-bit) before it was published, and it is rechecked whenever the package is rebuilt. The figures on this page come from the site index rather than from the publisher, so the download count is what people here have actually pulled.
If a mirror stops answering, use the contact form and it gets replaced. Reports about a specific title are handled faster than general messages because they name the file, the mirror and the point at which the transfer stopped.
Requests for a different version, a different language build or an older release go on the requests page. Older versions of a title are usually still held even when only the current one is listed.