Hedra AI All articles
Experiments

Pixels on Credit: What Upscaled AI Video Actually Delivers on a TV and a Phone

Hedra AI
Pixels on Credit: What Upscaled AI Video Actually Delivers on a TV and a Phone

Photo: PhotoPicker13, CC BY-SA 4.0, via Wikimedia Commons

Upscaling has a very good pitch: your generated clip is a little soft, a little small, and a little cheap-looking, so pay once and get 4K. In 2026, that pitch is everywhere in generative video tooling. The problem is that upscaling a synthetic frame is not the same as upscaling a photograph or a clean camera plate. Generative video already contains guesses. An upscaler can add more guesses, and then charge you for the storage.

We wanted to know what actually survives. Not which upscaler wins a spec sheet, but which artifacts make it to the couch, which ones disappear on a phone feed, and which ones quietly blow up the delivery budget.

The bench: four models, four upscalers, one ugly spreadsheet

We pulled twelve clips from four 2026 text-to-video and image-to-video models. Each clip was six seconds at either 720p or 1080p. The prompts were deliberately boring: a person talking to camera, a storefront sign with small lettering, a slow dolly through a kitchen, and a fast pan across a crowd. Boring is useful. It gives the upscaler fewer places to hide.

Each source went through four pipelines: a plain Lanczos resize, two hosted AI upscalers, and one open-source model running locally. Everything landed at 3840x2160. We measured frame-to-frame differences in static regions, inspected edge crops, ran OCR on the sign shots, and logged file sizes and encode settings. Then we watched the outputs on a 65-inch TV at about nine feet and on a phone at arm's length, because those are the two places most US viewers will actually meet this footage.

This was not a leaderboard. It was a survival test.

Bigger frames, softer facts

The first surprise was how often the plain resize looked more honest. Lanczos made the frame larger without inventing much. Edges stayed soft. Faces stayed faces. The file got bigger, but the image did not get weirder.

The hosted upscalers were more ambitious. On slow shots, they added convincing texture to fabric, hair, and countertops. On a static wide shot, one pipeline produced a genuinely better-looking kitchen. Then the camera moved, and the same pipeline started to boil. Fine details shimmered from frame to frame. A jacket collar changed shape. A background window developed a nervous flicker that was not in the source.

In our static-region frame-difference check, the plain resize added roughly 8 percent more variation than the source. One hosted upscaler added closer to 22 percent on a slow pan. That number is not a universal score, but it matches what your eye does: the upscaler is not just enlarging the frame, it is re-deciding the frame every 1/24th of a second.

The artifacts that survive the resize

A few failure modes showed up again and again.

Temporal boiling is the big one. Textures that look crisp in a still become unstable in motion. It reads as heat shimmer around hair, fabric, and foliage.

Edge halos are the second. Upscalers like to sharpen boundaries, especially around faces and product edges. On a phone, that can look like punch. On a TV, it looks like a cutout.

Texture confetti is the third. The model adds plausible detail in places where the source had none, then changes its mind. Skin pores, brick patterns, and background leaves all become tiny improv performances.

Text hallucination is the fourth, and it is the one that should scare anyone shipping ads. The upscaler often made small signs more legible. It also changed letters. A storefront that read CLOSED in the source came back as CLOSEO in one pass. OCR confidence went up while accuracy went down. That is a very specific kind of confidence.

What people saw on the couch and on the feed

We showed the clips to eleven colleagues in a casual viewing test. Nobody knew which pipeline was which. On the TV, the upscaled versions won on slow, static shots. The extra texture was visible and sometimes nice. But when the camera moved, viewers pointed at the shimmer before they pointed at the resolution. Several people described faces as looking 'wet' or 'scan-y.' That is not a technical term, but it is a clear product review.

On a phone, the whole argument mostly collapsed. At arm's length, a 1080p source and a 4K upscale looked close enough that nobody reliably picked the 4K version. Compression, screen size, and scrolling did more damage than the upscaler did good. The failures that survived were the obvious ones: a warped mouth, a sign that changed wording, a halo around a head. Resolution was the least interesting variable.

The bitrate bill

The source clips were mostly 6 to 12 Mbps. The 4K upscales landed between 35 and 60 Mbps before delivery encoding. That is a real line item. You pay to store the big file, pay to move it, pay again to transcode it, and then often deliver something closer to 12 to 16 Mbps so it actually plays nicely on a US living-room connection.

If the final delivery bitrate is going to be modest anyway, the upscale has to earn its keep before compression. Sometimes it does. Often it just gives the encoder more expensive noise to throw away.

How I would budget an upscale in 2026

Treat upscaling as a finishing decision, not a resolution checkbox. Upscale the shots that need it, not the whole timeline. Keep the source. Cap the multiplier at 2x unless you have a specific reason. Test at delivery size, not at 400 percent in a review app. Watch on a phone and a TV before you approve anything. And be extra careful with text, faces, and fast pans, because those are where the added confidence gets caught.

The 4K label is not the interesting number. The interesting number is how much of the original performance survives the trip. In our test, the pixels arrived. The detail sometimes did. The new failure modes always did, and they came with an invoice.

All Articles

Related Articles

Pixel Borders, Real Budgets: Testing Ten Image Editors Where the Mask Ends

Pixel Borders, Real Budgets: Testing Ten Image Editors Where the Mask Ends

Your TTS Sounds Great Until the Feed Normalizes It: A Twelve-API Loudness Audit

Your TTS Sounds Great Until the Feed Normalizes It: A Twelve-API Loudness Audit

Aspect Ratio Is a Contract: Auditing Ten Image APIs for True Vertical Output

Aspect Ratio Is a Contract: Auditing Ten Image APIs for True Vertical Output