How to Upscale Old Video Footage With Topaz Video AI
How to Upscale Old Video Footage With Topaz Video AI
Short answer: Open Topaz Video AI, drag in your clip, pick a model such as Proteus or Iris, set the output to 2x or 4x, preview a short range at low resolution, then export. A 720p clip upscaled to 1080p takes roughly real-time on a mid-range GPU; 4x on long footage can take many hours.
It is not magic, and it will not recover detail that was never captured. What it does well is clean up compression and add plausible grain.
What is Topaz Video AI?
Topaz Video AI is a desktop app that uses neural networks to upscale, deinterlace, denoise and frame-interpolate video. It runs locally on Windows and macOS, needs a modern GPU, and is licensed as a one-time purchase with a year of updates — around $299 for the full licence, though sales are frequent. After the update year ends, the app keeps working; you just stop receiving new models.
Local processing matters. Cloud upscalers charge per minute and upload your footage; this one does not.
Which model should you choose?
The model picker is where people freeze. A short decision path:
- Proteus — general purpose, auto-adjusting. Start here for most modern footage.
- Iris — tuned for faces and mid-quality sources. Best for interviews and talking heads.
- Gaia — sharpens crisp sources; poor with noisy ones.
- Theia — for compression artifacts; tends to look “clean” but slightly synthetic.
- Nyx — strength in denoising. Use before upscaling very noisy camcorder footage.
- Chronos / Apollo — frame interpolation, separate from upscaling entirely.
Grainy VHS needs Nyx then Proteus in sequence. A 2019 phone video at 1080p needs Proteus at 2x, nothing more.
How do you run an upscale?
- Drag the file onto the app window. Load a trimmed section rather than the full hour — the app trims with in/out markers.
- Choose the model from the Enhancement dropdown and set the multiplier. 2x is usually the honest ceiling for consumer footage; 4x on soft sources mostly adds invented detail.
- Set Output resolution explicitly. A 2x on 720p gives 1440p, which is awkward — override to 1920 and let the app scale.
- Add frame interpolation only if the source is cinematic 24fps and you want smoother motion. It doubles render time and can produce soap-opera motion.
- Generate a Preview of a 3–5 second range at low resolution. Check faces and text first.
- Export to ProRes or H.264 depending on the next step in your pipeline. ProRes is safer if you will grade afterwards.
That preview step saves hours. Text and signage are where models fail most visibly, and a still glance catches it.
How long does rendering take?
Rough figures from a mid-range desktop GPU:
| Source | Output | Approx. time |
|---|---|---|
| 1 min at 720p | 1080p, Proteus | 1–2 min |
| 1 min at 720p | 1440p, Iris | 3–4 min |
| 10 min at 1080p | 4K, Proteus | 40–60 min |
| 10 min SD (interlaced) | 1080p, Nyx + Proteus | 90+ min |
Chaining models multiplies the time. Run the chain on a short test clip before committing a full reel.
What are the real limits?
- Text and logos smear. Anything with lettering needs manual fixing or a hybrid workflow.
- No real detail recovery. Motion-blurred faces stay soft; the model invents plausible skin, not truth.
- GPU dependency. An older 4GB card will stutter and may refuse some models outright.
- Licence model confusion. Buying the app gets you one year of updates; many users report being surprised a year later when new models require a renewal.
Also worth knowing: interlaced broadcast footage needs deinterlacing first. Upscaling interlaced material produces combed edges that cannot be fixed afterwards.
Who is this for?
Archivists, editors working with legacy client footage, and anyone restoring family video. Also useful for game capture that was recorded at a lower resolution than the final deliverable.
Not for live streams, not for batch video-on-demand pipelines with thousands of files, and not a substitute for shooting properly the first time.
Quick Summary
- Price: roughly $299 one-time with a year of updates.
- Best general model: Proteus; Iris for faces; Nyx before upscaling noisy sources.
- Practical ceiling: 2x for consumer footage.
- Text handling: weak — expect to fix lettering manually.
- Speed: real-time to 6x real-time per pass on a mid-range GPU.
Last updated October 2026.
