Topaz Video AI Pro
Version 7.1.6 · Topaz Labs · Windows 10 / 11 (64-bit) · ranked #28 in the library
Enlarges, denoises, deinterlaces and reframes footage using trained models rather than interpolation, which is how old or low resolution material gets to a resolution a modern timeline can use.
Read the installation section below before you start. The order of the steps matters on this release.
About this release
Overview
Video AI runs trained models over footage to reconstruct detail that a conventional scaler cannot invent. Upscaling a standard definition archive tape to high definition, cleaning heavy grain and compression artefacts out of old digital footage, deinterlacing broadcast material properly, stabilising handheld shots and generating intermediate frames for slow motion are all separate models with their own strengths.
Model choice is most of the skill. There are several enhancement models tuned for different sources: heavily compressed web video, clean digital footage that just needs resolution, film scans with grain to preserve, and interlaced broadcast material. Each is previewable side by side on a section of your clip before you commit an hour of processing, which is the only sane way to work with this.
Frame interpolation covers the other half. Convert twenty four frames per second to sixty for smooth playback, or generate genuine slow motion from footage that was not shot at a high frame rate, with motion estimation that handles occlusion rather than smearing across it. Output goes to ProRes, H.264, H.265 or an image sequence, and the whole thing runs on the GPU with a batch queue for overnight work.
What it does
Feature breakdown
Model based upscaling
Reconstruct detail when enlarging rather than interpolating it, with models tuned for compressed, clean, film and interlaced sources.
Denoise and artefact removal
Remove grain, blocking and mosquito noise from old digital footage while keeping genuine texture.
Frame interpolation and slow motion
Generate intermediate frames for higher frame rate playback or true slow motion from normal speed footage.
Deinterlacing and restoration
Proper field handling for broadcast and tape sources rather than a blend or a field drop.
Side by side preview
Compare models on a section of the real clip before committing to a full length process.
Batch queue and GPU processing
Queue a folder and leave it running, with processing on the GPU and support for multiple cards.
Changed in this build
Version 7.1.6
- New enhancement model tuned for heavily compressed streaming sources.
- Faster processing on recent GPUs with better memory handling on long clips.
- Improved frame interpolation around fast motion and occlusion boundaries.
- Preview comparison rebuilt so more than two models can be compared at once.
- Export presets added for common intermediate codecs.
Installation
Follow these in order
- Go offline.
- Extract the archive.
- Disable real time protection.
- Run the installer and complete setup.
- Copy the included model files into the models directory if setup did not place them.
- Launch the application, confirm the models are listed, then re-enable protection.
Release notes
From whoever packed it
- The models are bundled here. Without them the application starts but has nothing to run.
- Block outbound access. An online check reverts the build to trial output with a watermark.
- Processing is slow by nature. A ten minute clip upscaled to 4K is an overnight job on most machines.
System requirements
Minimum and recommended
| Operating system | Windows 10 version 22H2 or Windows 11, 64-bit |
|---|---|
| Processor | Modern quad core or better |
| Memory | 16 GB minimum, 32 GB recommended for 4K output |
| Graphics | GPU with 6 GB VRAM minimum, 8 GB or more strongly recommended |
| Storage | 10 GB for the application and models, plus a lot of space for output |
Questions about this release
Answered before you ask
- Is the output watermarked?
- No, provided the application stays offline and the bundled models are in place.
- Which model should I use?
- Preview two or three on a difficult section of your actual clip. Source condition matters more than any general recommendation.
- Can it use two GPUs?
- Yes, multiple cards can be selected in preferences and work is split between them.
Download
Pick a mirror