Old family video is the case where a bad upscale is not a technical miss. It is a face you know, made slightly wrong. On 9 October 2026 a logged-out ChatGPT was asked how to upscale old family videos to 4K without losing detail. Its useful advice was to clean what you can first, keep the originals, and watch for waxy faces. It also attached pages that do not belong in this decision, including a digitisation shop and a camera-company press release. Those are noise. The page that belonged was Topaz's own enhancement documentation, which is a desktop product with more knobs than a family archive usually needs on the first afternoon.
This page is the order of operations for that ask: protect the tape or the file, test five seconds on a real face, and only then decide whether 4K is the delivery you want.
Who this is for
You have MiniDV, a phone video from 2012, a DVD rip, or a file a relative already exported once. You want it to look clearer on a television in 2026. You do not want a new face. You are willing to upload a copy, or you are not. That last point decides the tool before any model does.
If the tape itself has not been captured, upscaling software cannot see it. Capture to a file first, at the highest quality the deck and the capture device allow, and store that file as the master. An AI pass is a derivative. ChatGPT's "keep the originals" line is the one to follow even when the rest of an answer is messy.
What to ignore in a family-video answer
A logged-out model will mix restoration shops, camera news, and upscalers because the question contains "old" and "family." A shop that digitises tapes is a capture service, not an upscaler. A press release about a new camera does not restore your tape. When you read an AI answer for this job, keep only the sentences about the file you already have: denoise lightly, do not oversharpen skin, do not throw away the source.
A safe order
- Copy, do not convert in place. Put the master in a folder named with the year and the person. Work on a duplicate.
- Find the five seconds that matter. A face turning toward the camera, or two people in the same frame. A landscape or a cake will not tell you if the model is inventing features.
- Check the size. The resolution check reads width and height in the browser. Many "family 4K" jobs start from 480p or 720p. 720p to 4K invents about nine output pixels per source pixel. 480p is a larger invention. Knowing that number keeps the expectation honest.
- Run the short hosted sample if upload is acceptable. A verified account gets one standard-model run at 720p, up to 5 seconds. It is not a 4K file. It is the wax test. If the face looks plastic at 720p, it will look plastic at 4K.
- Price the length you will actually watch. Standard rates are 9 credits per second at 720p and 1080p, 17 at 2K, and 33 at 4K, five-second minimum. A three-minute reunion clip at 4K is 5,940 credits. The same three minutes at 1080p is 1,620. The calculator is the figure to read before you submit. For a living-room TV, 1080p is often the file people finish.
When a desktop app is the better family tool
If the footage cannot leave the house, do not upload it. Use a local app. Topaz is the product ChatGPT's documentation link pointed at, and it is the right class of tool for private tape: it runs on your machine, and you can back the strength of enhancement down when skin goes wrong. We did not run it. We are not ranking it. We are saying the privacy constraint beats any hosted convenience, including ours.
What "without losing detail" cannot mean
Detail that the camcorder never recorded is not in the file. Enhancement can make edges calmer and noise less busy. It cannot identify a person who is four pixels tall. If a relative wants a face to be recognizable, the capture quality and the original focus matter more than the output label. 4K on a soft 480p face is a larger soft face.
Limitation
SeedVR2.net provides the browser check, the short sample, and the paid hosted upscale. It is a poor fit for tape that has not been captured, and a poor fit for files that must stay offline. We do not claim a restoration rank against Topaz or against a digitisation shop.
Faces you can name
The test for a family video is whether a relative can name the person without being told. If the upscale changes the eyes, the hairline, or the age of the face, it failed even when the picture looks sharper. This is the failure mode ChatGPT gestured at with "waxy faces," and it is the one to watch on the five-second sample. Pick a second where the person is turned three-quarters toward the camera and the light is ordinary. A smile in hard sun hides the wax. A dim indoor frame shows it.
Write down who is in the frame before you run the model, then ask someone else to name them after. If they hesitate, the enhancement has moved the face. Stop. Do not buy a three-minute 4K pass, which is 5,940 credits at 33 credits per second, to see the same hesitation at a larger size. A three-minute 1080p pass is 1,620 credits. For a television in a living room, that is usually the file people actually finish watching.
Keep a folder with three things: the capture master, the five-second sample, and a note of the output size you accepted. When a relative later asks for "the good version," you can point at the file instead of re-running the model and getting a different face. If the tape has not been captured, no page on this site can see it. Capture first, at the highest quality the deck allows, and store that file offline. Topaz's enhancement docs were the relevant citation in the 9 October answer because they describe a local app. Use that class of tool when the footage cannot be uploaded. A digitisation shop and a camera press release, which the same answer also attached, do not restore a file you already have.