AI Video Upscaling on Mac — Restore Old Family Videos This Weekend

Somewhere in a drawer at your parents’ house is a stack of MiniDV tapes or burned DVDs: a wedding, first steps, a Christmas where everyone in the frame is twenty years younger. The footage is 480 interlaced lines of comb-toothed, washed-out video that looks unwatchable on a 4K TV — which means, in practice, nobody watches it.

Last winter I restored my grandparents’ wedding video and three tapes of my own family’s footage from a 2003 Sony Handycam on my Mac Studio. The result isn’t magic — and I’ll be precise about where AI fails — but the difference between “blurry artifact of the past” and “video my mother watched twice and cried at” is entirely achievable in one weekend. Here’s the exact plan.

Step zero: get the footage off the dying media

This is the unglamorous step everyone skips past, and it’s the most urgent — tapes demagnetize and DVD dye layers rot whether or not you ever get around to the AI part.

MiniDV tapes: the best capture is the original digital stream over FireWire, not a re-recording. A used MiniDV camcorder with FireWire output (~1,500 CZK / $60 on local marketplaces), an Apple FireWire-to-Thunderbolt adapter chain, and capture via iMovie or ffmpeg gets you the pristine DV file. If the adapter chain defeats you, a local digitization service charges around $10–15 per tape — worth it for irreplaceable footage.

Old DVDs of home movies: rip them before they rot. HandBrake (free) handles unencrypted home-movie DVDs directly. Critical setting: in the Video tab set deinterlace to off — you want the raw interlaced stream, because Topaz’s AI deinterlacer (next section) is dramatically better than HandBrake’s, and you must only deinterlace once. Use a high-bitrate or near-lossless preset; this is an intermediate file, not the final product.

VHS: you’ll need an analog-to-digital capture device (the cheap USB ones are mediocre; a service is usually the better call) — but the downstream AI workflow is identical.

One rule above all: archive the untouched originals. The AI models of 2030 will be better than today’s, and you’ll want the source files again.

The premium tool: Topaz Video AI

Topaz Video AI (~$299, one-time with a year of updates, native Apple Silicon) is the tool this category is built around, and having tested the alternatives, the money is justified for a family-archive project. Three things it does genuinely well:

Deinterlacing old camcorder footage. Interlaced video — every DV tape and most home DVDs — stores two half-frames per frame, which is why fast motion shows comb teeth on modern screens. Topaz’s deinterlacing models reconstruct full frames using motion information from neighboring frames, and it’s the single biggest visual upgrade in the whole pipeline. Bigger than the upscale. Process: deinterlace first, then enhance.

480p → HD upscaling. The Proteus model (manual control) and Iris (tuned for faces and compressed sources) are the two I use for camcorder footage. Proteus at “Relative to Auto” with slight sharpening, 4x denoise restraint, upscaling 720×480 → 1920×1440 (keep the 4:3 aspect — don’t crop your family to fit 16:9) produces footage that reads as “decent early HD” rather than “VHS on a big TV.”

Frame interpolation. The Chronos/Apollo models can take stuttery footage to smooth 50/60 fps. Use sparingly — interpolated home video can acquire a soap-opera smoothness that feels wrong for the era. I interpolate only footage with broken cadence from bad prior conversions.

Realistic processing times for 480p → 1440p (Proteus), measured and extrapolated from my own runs — this is where expectations need calibrating:

Add a deinterlace pass and roughly double it. A 10-tape archive is therefore a queue-overnight project on any chip — which is exactly how the weekend plan below treats it.

Free and cheaper alternatives

If $299 isn’t happening, the open-source route is real, just less polished:

• Video2X (free, open source) wraps Real-ESRGAN and friends with Metal support — solid upscaling, no AI deinterlacing, so pair it with ffmpeg’s excellent bwdif filter: ffmpeg -i tape.dv -vf bwdif=mode=send_field out.mp4.
• Real-ESRGAN-ncnn via command line runs surprisingly fast on Apple Silicon GPUs for frame-by-frame upscaling.
• FCPX / DaVinci Resolve Studio users already own decent “Super Scale”-class upscalers — Resolve Studio’s is genuinely good, though its deinterlacing trails Topaz.

The honest gap: on clean progressive footage, free tools get you 70–80% of Topaz quality. On interlaced, noisy camcorder tape — i.e., the actual family archive — Topaz’s integrated deinterlace+denoise+upscale pipeline is meaningfully ahead.

For the photo prints and scans in the same drawer, the sibling tools matter: Topaz Photo AI (~$199) for batch restoration of scanned photos, or the free Upscayl app (Real-ESRGAN under the hood, native Mac build) which handles 80% of cases at $0. Faces in old photos benefit from GFPGAN-based restoration — built into Upscayl’s face-enhance option.

Realistic expectations: AI cannot invent your grandmother’s face

This section is the one that saves you disappointment, so I’m putting it before the workflow.

Upscaling is plausible hallucination, not recovery. The model paints detail that is statistically likely, not detail that was there. The practical consequences:

• Faces can go uncanny. Small or motion-blurred faces are the failure mode — the model “restores” them into subtly wrong, waxy versions of people you know intimately. You will notice immediately because you know these faces. When it happens, dial down enhancement strength (Proteus manual mode) or accept softness; soft-but-true beats sharp-but-wrong in family footage, every time.
• 480p becomes good 1080p-class, not 4K. Pushing camcorder footage past ~2x–3x upscale just enlarges the hallucinations. 720×480 → 1440p is the sweet spot; the TV’s own upscaler can do the last hop.
• Garbage in, modest out. A dark, smeary indoor shot from 2003 becomes a cleaner dark, smeary shot. The wins are biggest on decently-lit outdoor footage.

Treat the AI as a very good restoration technician, not a time machine, and you’ll be delighted instead of unsettled.

The weekend battle plan

Here’s the workflow that got me through 9 hours of footage in one weekend, structured around the fact that processing is slow but unattended:

Friday evening — triage and capture. Don’t restore everything; restore what people will watch. I sorted tapes into three piles: must-do (the wedding, first steps), nice-to-have, and skip (40 minutes of a 2004 aquarium). Start capturing/ripping the must-do pile — capture is real-time for tapes, so this runs while you watch a movie.

Saturday morning — test on 30-second clips. This is the step that separates a successful weekend from a wasted one. In Topaz, export 30-second test clips from three representative scenes (bright outdoor, dim indoor, fast motion) and try 2–3 model/setting combinations on each. Six minutes of processing per test instead of discovering on Sunday that 8 hours of output has over-sharpened faces. Pick one setting per tape type, write it down, stop tweaking.

Saturday afternoon → Sunday — queue overnight runs. Topaz has a proper batch queue: load every captured file, apply the preset you validated, set output to H.265 (or ProRes if you’ll edit afterward), and let the Mac grind. My M2 Max chewed through ~9 hours of deinterlace+upscale across Saturday night and Sunday. Plug in a laptop and run caffeinate -i so it doesn’t sleep. Power cost for the whole weekend on Apple Silicon: under 2 kWh — pocket change, and silent enough to run while the household sleeps.

Sunday evening — the payoff. Spot-check the output, then do the part that makes it matter: AirPlay the restored wedding video to the TV with the people who were in it. My grandfather narrated his own wedding for twenty minutes. That’s the actual deliverable — the H.265 file is just the medium.

Final checklist: originals archived untouched, restored versions shared to the family (iCloud Shared Album handles the codec headaches), and a note in your calendar to re-run the best tapes when the models improve. The drawer full of tapes has a half-life. The weekend you spend on this is the cheapest insurance against losing what’s on them — and the rare tech project whose end result is your mother in tears for the right reasons.