AI Video Upscaling: How to Restore Old Footage to 4K in 2026

A YouTuber I work with handed me a hard drive full of his 2014–2017 vlogs last winter. The plan was a "10 years on YouTube" retrospective. The footage was 720p at best, some of it 480p, all of it filmed on a Canon T3i with a kit lens and the camera's awful built-in noise reduction. After three weeks of restoration work — most of it AI-driven — we cut a 22-minute remaster that read as 4K on modern displays. The original raw clips would have been embarrassing at fullscreen. AI video upscaling made the project economically possible. This is the working guide to doing the same on your own archive.

What AI video upscaling actually solves

Image upscaling has been good since 2023. Video is harder because every frame has to be temporally consistent with the frames around it. A naive frame-by-frame image upscale produces flicker — fine details that change frame-to-frame because the upscaler hallucinates them differently each time. AI video upscalers solve this with temporal attention, looking at multiple frames simultaneously to enforce stable detail.

Two leaders in May 2026:

• Topaz Video AI — The professional default since 2022 and still the safe choice. Multiple neural models (Proteus, Iris, Artemis, Nyx) for different content types. Mature ecosystem, NLE plugins, batch processing.
• Bytedance Upscaler Video — Newer, faster, particularly strong on faces and human subjects. Slightly more aggressive on detail recovery, occasionally over-sharpens textiles.

Versely runs both behind a single endpoint in the AI video generator and content workflows, so you can route a clip to the right model per content type without leaving the dashboard.

VHS and SD footage restoration

VHS is the hardest case. The source has 240–333 effective lines of resolution, severe analog noise, chroma bleed, and time-base instability. Pushing it to 4K is roughly a 10x linear upscale — far beyond what any single model can do cleanly.

The pattern that actually works:

1. Capture clean. Use a properly calibrated capture card. Garbage-in is unsalvageable.
2. Stabilize first. Time-base correct (TBC) the analog signal before any AI processing. Topaz can sometimes correct TBC artifacts; better to do it right at capture.
3. Denoise. Run a heavy denoise pass before any upscale. Topaz's Nyx model is built for this.
4. Deinterlace. VHS is interlaced. Deinterlace properly (Topaz QTGMC mode) before upscaling.
5. Two-step upscale. Go SD to 1080p first, evaluate, then 1080p to 4K. One-step 10x upscales produce visible texture invention.
6. Frame-rate interpolation last. If you want 60fps from 30fps source, interpolate after upscaling, not before.

For the 2014 vlog footage I mentioned, the source was already digital 720p so VHS-style preprocessing was unnecessary. The restoration was simpler: denoise, then 720p to 4K via Topaz Proteus, then color regrade.

The denoise-before-upscale rule

This is the single most-violated principle in amateur restorations. Upscalers amplify whatever is in the source — including noise. Run an upscale on noisy footage and you get crisp, beautiful, sharp noise. The detail you wanted is lost in the texture.

Order operations:

1. Denoise (Topaz Nyx, or DaVinci's neural noise reduction)
2. Stabilize (if handheld)
3. Upscale (Topaz Proteus / Bytedance Upscaler Video)
4. Frame interpolation (Topaz Apollo or Chronos)
5. Color grade and final sharpening

Reverse any of these and you get worse results. Color grade after upscale because you want to grade the final detail level. Frame interpolate after upscale because interpolating low-res frames and then upscaling produces compounded artifacts.

Topaz Video AI vs Bytedance Upscaler Video

Tested both on the same 12 source clips in February 2026 — handheld smartphone footage from 2018, a DSLR vlog from 2015, and one VHS-captured family video from 1994.

Scenario	Topaz Video AI	Bytedance Upscaler Video
720p digital to 4K	Excellent (Proteus)	Excellent
1080p smartphone to 4K	Very good	Excellent (faces)
SD VHS to 1080p	Excellent (multi-pass)	Good
Heavy compression artifacts	Excellent (Iris)	Fair
Talking-head faces	Very good	Best
Fabric / texture detail	Best	Sometimes over-sharp
Frame interpolation built in	Yes (Apollo, Chronos)	No
Cost (May 2026)	$299 license + GPU	~$0.20–0.50 per minute
Speed (4K render, 1 min source)	5–15 min on RTX 4090	2–5 min cloud

For old vlog restoration with talking-head shots, Bytedance is my default. For VHS or heavily compressed archival, Topaz with model switching per content type. For mixed projects, Versely's routing layer picks per-clip.

Frame interpolation alongside upscaling

Old footage is often 24 or 30fps. Modern displays handle 60fps and the eye reads it as smoother, more "current." Frame interpolation models (Topaz Apollo, Topaz Chronos, RIFE for the open-source crowd) generate intermediate frames.

Two cautions:

• Do not interpolate everything to 60fps reflexively. Cinematic 24fps is part of the cinematic look. Interpolating a film to 60fps produces the "soap opera effect" most viewers find unpleasant.
• Interpolation hates motion blur. If the source has heavy motion blur, the interpolated frames blend awkwardly. Reduce motion blur before interpolating, or skip interpolation for those clips.

For vlog restoration, 30fps to 60fps interpolation is appropriate — vlogs were always video-look, not film-look. For a music video remaster, leave the original frame rate.

Archival workflows for content creators

If you have a backlog of old content and you are thinking about a remaster series, here is the operational plan I use for clients.

1. Inventory. List every clip, source resolution, source format, runtime. Most creators are surprised by how much usable archive they have.
2. Tier by salvageability. Tier A (1080p source, clean): light enhancement only. Tier B (720p, some noise): full denoise + upscale. Tier C (480p or VHS): heavy multi-pass restoration. Tier D (severely damaged): consider replacing with new b-roll.
3. Pick model per tier. Tier A and B: Bytedance for faces, Topaz Proteus for everything else. Tier C: Topaz exclusively, with manual model switching.
4. Batch where possible. Versely's batch endpoint accepts an entire archive folder and routes per-clip.
5. Color match across the archive. A 2014 clip and a 2024 clip should grade to a consistent look in the remaster. Use a single LUT applied post-upscale.
6. Generate replacement b-roll for irrecoverable clips. Versely's AI b-roll generator fills gaps where original footage is unsalvageable.

For the 22-minute vlog retrospective, total restoration time was about 18 hours of operator work plus 9 hours of GPU rendering, spread across three weeks.

When AI video upscaling fails

Honest list:

• Sub-240p source. There is not enough information for the model to recover faces or fine detail without inventing it. Use as low-resolution texture or replace with regenerated content.
• Severe rolling-shutter or jello effect. Upscaling does not fix it. Stabilize and de-jello first; some artifacts remain.
• Aggressive in-camera sharpening. 2010s smartphone footage often had heavy in-camera sharpening that produces ringing artifacts. Upscalers amplify these. Use Topaz Iris which de-emphasizes existing sharpening before upscaling.
• Very fast motion in low light. Motion blur plus high noise is the worst case. Result is usable but not crisp. Frame interpolation actually makes this worse.
• Compressed VHS captures from late-generation copies. Generation loss compounds. There is a floor.

Versely's video upscale stack

End-to-end, the workflow inside Versely:

1. Upload source clips (up to 4K input, any common codec).
2. Auto-classify content type (talking head, action, animation, archival).
3. Route per clip to Topaz Video AI or Bytedance Upscaler Video based on classification.
4. Optional denoise pre-pass and frame interpolation post-pass.
5. Output 4K at 24/30/60fps as specified.

For creators thinking about remastering and republishing old content, the AI content creation 2026 complete playbook maps out the broader cadence of new versus restored, and the best AI video editing tools 2026 breakdown covers the NLE side of integrating restored clips into new edits.

FAQ

Can AI upscale a VHS tape to 4K?

Yes, with caveats. Capture cleanly, denoise heavily, deinterlace, and use a two-step upscale (SD to 1080p, then 1080p to 4K). Expect the result to read as a polished SD or low-end 1080p look at 4K display, not native 4K. The detail floor of the source caps what is recoverable.

Topaz Video AI vs Bytedance Upscaler Video — which is better?

Topaz wins on archival, VHS, and texture-heavy footage. Bytedance wins on talking-head faces and modern 720p–1080p sources. For mixed projects, route per clip. Versely does this automatically.

How long does AI video upscaling take?

On Topaz running locally with an RTX 4090, roughly 5–15 minutes per minute of 4K output. Cloud-based Bytedance is 2–5 minutes per minute on Versely's infrastructure as of May 2026. Long archives plan in hours and days, not minutes.

Is AI video upscaling worth it for old YouTube videos?

For creators with 5+ years of archive and an audience that responds to retrospective content, yes. The cost is operational time more than money. For one-off old clips, the time investment is harder to justify unless the clip has specific value.

Can I upscale copyrighted footage I do not own?

Upscaling does not change ownership. If you do not have rights to the original footage, you do not have rights to the upscaled version. Restoring your own old vlogs and home videos is fine; restoring someone else's commercial film for redistribution is not.

Will AI upscaling hide watermarks or DRM?

It is not designed to and you should not try. Aggressive denoising can sometimes reduce visible watermarks but distorts the rest of the image and is a trademark and copyright issue regardless.

Should I upscale before or after editing the video?

Upscale the master, then edit at 4K. Editing in low-res and upscaling the final cut is more expensive (every cut is a separate clip to upscale) and produces worse results because each clip is processed without the context of what comes before and after.

Bottom line

AI video upscaling in 2026 is operationally mature — Topaz Video AI and Bytedance Upscaler Video both produce results that hold up at 4K display sizes from 720p and even 480p sources. The discipline that matters is order of operations: denoise, deinterlace, upscale, then interpolate. Get the order right, route the right model per clip, and an archive that was unwatchable on a modern TV becomes the most differentiated content on your channel — old footage with a remastered presentation is the kind of catalogue play almost no competitor is making.