How to improve old video quality and upscale it to 4K

Why a 480p video looks bad on a 4K TV

A 480p video has 854 × 480 pixels if it was shot in 16:9, or 640 × 480 if it came from an older 4:3 phone or camcorder: roughly 0.3 to 0.4 megapixels per frame. A 4K television shows 3840 × 2160 pixels, about 8.3 megapixels. To fill that screen, the TV has to produce between 20 and 27 pixels for every single pixel you recorded. A 640 × 360 clip, a common size for early smartphone video, needs 36 new pixels for each original one.

The scaler inside a TV or a media player does this by interpolation: it looks at neighbouring pixels and places intermediate values between them (bilinear or bicubic filtering, sometimes with a sharpening pass on top). That spreads the existing information across a larger grid. It does not add any. Edges get softer, a thin halo appears around high-contrast lines, and surfaces that should have texture, such as a jumper or a lawn, become smooth gradients. Up close on a 140 cm (55-inch) screen, the picture reads as blurry rather than small.

The second problem is older than the screen size: compression. Phones and camcorders of the 2000s and early 2010s recorded at low bitrates, and the encoder (usually H.264 or an older MPEG variant) throws away fine detail first, because fine detail costs the most bits. Hair turns into a flat brown shape with a few painted strands, skin loses its pores and becomes a plastic surface, and anything that moves quickly dissolves into 16 × 16 pixel blocks. The information about individual hairs was never written to the file. No amount of stretching can bring back what was discarded at the moment of recording.

So the honest starting point when you want to improve old video quality is this: the frame is small, and what is in the frame is already simplified. A conventional upscale makes the first problem bigger without touching the second. Anything that is going to look better on a modern screen has to rebuild texture, not just enlarge it. The same applies to the photos from that era, which is why the tool to restore old photos uses a reconstruction model rather than a resize.

What an AI video enhancer actually does differently

Generative video restoration is a different operation from scaling. Airpic uses SeedVR2, a model published by ByteDance Seed under the Apache-2.0 licence, trained to take a degraded video and output a clean, higher-resolution version. Instead of interpolating between pixels, it has learnt what hair, beards, skin, fabric, brick and foliage look like at high resolution, and when it meets a smeared brown patch where hair should be, it draws plausible strands. That is why the output of an AI video enhancer looks sharper than the source rather than just larger: when you enhance a video with AI this way, it adds texture that the file never contained.

The part that matters for video, as opposed to a single photo, is consistency between frames. If you ran a photo upscaler on every frame independently, each one would invent slightly different strands and pores, and played back at 25 frames per second the result would shimmer and crawl. A video restoration model looks at a group of neighbouring frames together, so the texture it generates in frame 40 agrees with frame 39 and frame 41. Hair moves as hair, not as noise.

For 4K, Airpic does not ask the model to generate 8 megapixels per frame directly. The generative pass runs at 1080p, where the model does its reconstruction, and a final 2× step mixes AI scaling with classic scaling to reach 3840 × 2160. The mix is deliberate: a second pure AI pass on top of a generative one tends to look painted, with over-regular texture everywhere. The blend keeps the detail the model recovered without inventing another layer on top of it.

Now the limits, because they decide whether this is worth doing on your clip:

• Very small faces. If a face is 30 pixels high in a 480p frame, the model has a few blobs to work from. It will produce a sharp face, but not necessarily the right face: eye spacing, the line of a nose or a smile can shift. On a group shot from across a room, people can come back looking subtly unlike themselves. A face that fills a quarter of the frame comes back recognisable; a face the size of a fingernail is a gamble.
• Extreme motion blur. A hand that is a streak across three frames stays a streak. The model restores texture on surfaces it can identify; it does not reconstruct the shape of something that was never captured in one place.
• Sources that were already upscaled. If a phone app once "enhanced" a 480p clip to 1080p, the tool sees a 1080p file, offers only the 4K option, and will faithfully restore the smearing the app introduced as if it were real detail. Always start from the original file, even if it looks worse.
• Interlaced footage. PAL camcorders recorded 720 × 576 interlaced (576i), with alternating lines taken at two different moments. Airpic has no deinterlacing step, so deinterlace in your editing software before uploading, or the comb lines on moving subjects will be treated as detail to keep.

Everything else, which is most family and holiday footage, is a good candidate: static or slow shots, people reasonably close to the camera, and a source that is soft and blocky but not destroyed. That is exactly what the tool to upscale a video to 4K is built for.

Choosing the clips: shots of up to 20 seconds

The tool accepts one clip of up to 20 seconds at 25 frames per second, which works out at about 16-18 seconds for 30 fps footage. The frame rate is capped at 30 fps on output, so a 60 fps clip comes back at 30. Those limits exist because every second of video means 25-30 frames of generative work. Rather than fighting them, use them as an editing discipline: you are not restoring a 40-minute tape, you are restoring the dozen moments on it that anyone will watch twice.

Go through the original and write down timecodes for the shots that matter: the toast, the first steps, the goal, the moment someone turns round and laughs. A shot is one camera position without a cut. Keep each piece under the limit and cut on the shot change, not in the middle of a shot. The model uses neighbouring frames to keep texture consistent, and a hard cut in the middle of a clip gives it nothing to agree with across that boundary. One shot per clip gives the cleanest result.

Trim before you upload, not after:

1. On the phone, open the video in the gallery app, tap Edit, drag the handles to the in and out points and save as a new clip so the original stays intact.
2. On a computer, any free editor will do: mark in and out, export the range.
3. Export at the original resolution and frame rate. This is the step people get wrong. If the trimming app offers to export at 1080p or to "improve" the clip, say no. You would be handing the restoration model a bicubic upscale instead of the real pixels, and it would then only offer you 4K because the file already reads as Full HD.
4. Keep the audio. The upscaled MP4 carries your original audio track, so trim with sound, not from a muted copy.

Formats: MP4, MOV and WebM all work. If the file is already 1080p, the tool only offers 4K; if it is already 4K there is nothing to do. Old low-resolution files are small, so the upload itself is quick; the processing is the slow part.

Do this for the five or ten shots you actually want, and you have a set of clips ready for the next decision: how far to increase the resolution.

1080p or 4K: how far to increase video resolution

Full HD is 1920 × 1080; 4K is 3840 × 2160, four times the pixels. Which one you need depends on where the clip will be watched, not on which number sounds better.

1080p is enough when the destination is a phone or social media. Most phone screens have around 1080 pixels across the short side, so a 4K file is scaled down before anyone sees it, and every platform re-encodes what you upload anyway. For a reel, a WhatsApp send or a clip people will watch on the sofa holding a phone, Full HD already carries all the detail the model recovered. It also finishes sooner and costs fewer credits per clip.

4K is worth it when the clip will be shown on a 4K television or a projector, when you are cutting it into a 4K timeline alongside modern footage, or when you plan to crop and reframe. A 4K frame lets you punch in 2× on a 1080p timeline without losing sharpness, which helps with old footage that was framed too wide. A memorial or anniversary film for a big screen is the typical case.

Real processing times, measured in our production tests on 15 October 2026: a 5-second clip at 640 × 360 px upscaled to 1080p in 309 seconds and to 4K (3840 × 2160 px) in 356 seconds. Two things follow from those figures. First, the planning rule is 5-7 minutes for every 5 seconds of video, so a full 20-second clip in 4K takes around 25 minutes. Second, 4K adds only about 15% to the time of 1080p, because the heavy generative work happens at Full HD and the final 2× step is comparatively quick. If you are undecided and the screen is large, the extra wait for 4K is small.

This is slow by the standards of a photo tool because it is generative AI running on every frame, not a resize. Keep the browser tab open while it works. The GPUs are in the European Union (Sweden), the source is deleted when the job finishes and the result one hour after you download it, 24 hours at most, and nothing is used for training. If that matters to you, there is more detail on AI image processing in the EU.

Step by step: upscale video to 4K in Airpic

Here is the exact sequence inside Airpic, with what you see at each step. There is no account to create and nothing to install.

1. Open the tool to upscale a video to 4K and drop your trimmed clip onto the upload area, or tap to choose it. MP4, MOV and WebM are accepted. A file card appears with the name, duration and resolution of the clip.
2. Read the line under "Resolution". If the clip is over the limit, it says so straight away: "This clip is too long: the maximum is 20 s. Trim it before uploading." The button stays disabled and nothing is charged. If the source is already Full HD, the same line tells you to choose 4K.
3. Choose the resolution. Two options: "1080p (Full HD)" and "4K (2160p)", each labelled with the AI credits it costs. Pick using the rule from the previous section: phone and social, 1080p; television, projection or a 4K edit, 4K. The hint updates with an estimate in minutes for your clip and a reminder that the credits are refunded if the job fails.
4. Press "Upscale to 4K". Airpic first checks the AI credits saved in this browser. If there are none, or not enough, a message tells you the tool is prepaid and how many credits it needs, and asks you to paste a code or buy a pack. The pack gives you a bearer code that is stored in the browser, so you only do this once. This tool is not part of Airpic's free monthly download; current pack sizes are on the tool page.
5. Processing starts and the credits are deducted at that moment. If the job fails for any reason (an unreadable file, a clip too large for the GPU memory, a timeout), the credits come back to your code automatically and the error message says so.
6. Wait with the tab open. A progress bar runs for the estimated time. When it finishes, the result plays directly in the page, so you can judge faces and texture before downloading. The file is an H.264 MP4 with your original audio.
7. Download. The clip is already paid for, so you can download it as many times as you like within the next hour. After that it is gone from the server.

Repeat for each shot on your list, then put them together. The simplest route is to edit the video with AI in Airpic: upload the upscaled clips and any photos, choose the format (16:9 for a television, 9:16 for phones), set a target length and a pace, and the editor proposes the order, cuts and transitions, with optional word-by-word subtitles you can correct before export. The preview is free in the browser and you pay only when you export. That export is a 1080p MP4, so if your goal is a 4K master for a big screen, assemble the 4K clips in your desktop editing software and use Airpic's editor for the phone-friendly version. For a vertical cut of the single best moment, you can also make shorts from a video with automatic captions.

What you end up with is not the footage you wish you had shot in 4K. It is the footage you did shoot, with the texture that compression took out put back in a way that holds together from frame to frame. For a 480p clip that has sat on an old hard drive for fifteen years, that is the difference between people glancing at it and people watching it again. Start with your most important shot and upscale the video to 4K to see how it holds up.

FAQ

Keep reading