Overview
Video AI does a small number of things and does them better than anything conventional. It enlarges footage by inferring detail rather than interpolating pixels, so old standard definition material becomes genuinely watchable at high resolution instead of merely larger. It generates intermediate frames for slow motion from footage shot at a normal rate. It stabilises without the aggressive cropping that traditional stabilisers need, and it removes noise and compression artefacts from material that has been through too many re encodes.
This build includes every model rather than a subset, exports at any resolution and length with no watermark, and runs entirely on the machine. All models are bundled in the archive, so nothing downloads on first use.
Upscaling
Several models are provided because footage varies enormously. There are models tuned for clean digital sources, for compressed material that has been through streaming, for interlaced broadcast footage, and for film scans with grain that should be preserved rather than smoothed away. Choosing the right one matters far more than any setting inside it.
The honest expectation for old material is a large improvement rather than a miracle. Detail that was never recorded cannot be recovered, and pushing the enlargement factor too far produces a smooth plastic look. Two times from a standard definition source is usually excellent. Four times is where the artefacts start showing.
Frame interpolation and stabilisation
Frame interpolation generates intermediate frames, which converts normally shot footage into slow motion or raises a low frame rate to something smoother. It handles moderate motion well and struggles with fast movement across a complex background, where it will occasionally produce a warped frame that needs a manual cut.
Stabilisation analyses the whole shot rather than a rolling window, so it can correct without the heavy crop that traditional stabilisers require. It also has a rolling shutter correction pass for footage shot on sensors that produced skew on fast pans.
Running it
This is a graphics card bound program to an extreme degree. On a current card a minute of standard definition footage upscaled to high definition takes a few minutes. On an older card the same job runs overnight. Check a ten second segment with your chosen model before committing a long piece.
The preview compares source and result side by side at the actual output resolution, which is the only reliable way to judge a model choice. Judging from a scaled down preview will mislead you every time.
What you get
- Machine learning upscaling with models tuned for different source types
- Frame interpolation for slow motion and frame rate raising
- Whole shot stabilisation without heavy cropping
- Rolling shutter correction
- Noise and compression artefact removal
- Interlaced broadcast footage handling
- Grain preserving models for film scans
- All models bundled offline, nothing downloaded on first use
- No watermark, resolution cap or length limit on export
Inside the archive
- Video AI Pro 7 application, offline installer
- Complete model set, preinstalled
- Preview and comparison tools
- Notes on model selection by source type
System requirements
| Operating system | Windows 10 version 22H2 or Windows 11, 64-bit |
|---|---|
| Processor | Multicore, 6 threads or more |
| Memory | 16 GB minimum, 32 GB recommended |
| Graphics | 8 GB VRAM strongly recommended, 6 GB is the floor |
| Storage | 20 GB for the install with models, plus output space |
| Display | 1920 by 1080 minimum for meaningful preview |
Installing it
- Update the graphics driver before installing.
- Suspend real time protection for the install.
- Run setup as administrator and let the model files finish unpacking.
- Open once and run the built in benchmark to confirm the card is being used.
- Test a ten second segment before queueing a long piece.
Mirrors
| Route | Region | Note | State |
|---|---|---|---|
| Direct, primary | Europe | No wait, resumable | Online |
| Direct, secondary | North America | No wait, resumable | Online |
| Torrent magnet | Global | Recommended at this size | Online |
Release history
- New model for heavily compressed streaming sources
- Frame interpolation artefacts reduced on fast horizontal motion
- Faster model loading between queued jobs
- Stabilisation reworked to reduce crop further
- Fixed memory exhaustion on cards with 6 GB during 4K output
- Grain preserving model added for film scans
Questions about this release
How much can it really improve old footage?
A large improvement, not a miracle. Two times enlargement from standard definition is usually excellent. Beyond four times the result starts looking artificial.
Which model should I use?
It depends entirely on the source. Test a ten second segment with two or three candidates and compare at full output resolution.
Are all models included?
Yes, bundled in the archive rather than downloaded on first use.
Why is it so slow?
It is graphics card bound. On a card below 8 GB, long jobs will take many hours.
Comments
Interpolation is great until something moves fast across a busy background, then you get a warped frame. Cut around it.
All models present, no download on first run. Benchmark confirmed the card straight away.
Comments are read before they appear. If a build stops working, say so here and it gets rebuilt rather than quietly left up.
Family tapes from a camcorder, two times upscale with the grain model. Genuinely watchable now.