HOW IT WORKS
What to expect from browser-based video upscaling
Video is the same model as images, applied frame by frame — which makes it dramatically heavier. Knowing the constraints up front saves you a long run that was never going to finish.
A ten-second clip at 30 frames per second is three hundred images. Every one of them is decoded, passed through the AI model, and re-encoded into the finished file. That is the whole reason video upscaling takes minutes where a photo takes seconds, and why it is far more sensitive to how much memory your machine has free.
Because all of it happens inside your browser tab, there is no queue, no upload, no file size ceiling imposed by a server, and no copy of your footage sitting on someone else's disk. The trade is that your own hardware sets the limits, and the tab has to stay open for the duration.
Start with a short clip. Ten or fifteen seconds tells you what your machine can do, how the settings affect the result, and roughly how long the real job will take — long before you commit to a run measured in tens of minutes.
Formats and browser support
Video upscaling needs both WebGPU and WebCodecs, which in practice means a Chromium-based browser. Image upscaling has a CPU fallback and works nearly everywhere; video does not.
| Requirement | Supported | Notes |
|---|---|---|
| Browser | Chrome 113+, Edge 113+ | Other Chromium browsers of the same vintage generally work. Safari and Firefox cannot yet run this pipeline. |
| Container | MP4, WebM, MOV | The container is unwrapped in your browser; nothing is sent anywhere to be converted. |
| Video codec | H.264, VP8, VP9 | The codecs the browser itself can decode. Exotic or professional codecs will not open. |
| Memory | 8 GB or more recommended | 4 GB is workable for short, small clips. HD footage on 4 GB will struggle. |
Target resolution and processing mode
Two controls shape the run. The target resolution — 1080p, 2K or 4K — is the ceiling you want the output to reach. The processing mode trades time against reconstruction quality.
- Fast — the quickest pass. Useful for checking that a clip decodes correctly and for previewing what the settings will do before committing.
- Balanced — the default, and the right answer for most footage. Clearly better than Fast without the time penalty of the heavier modes.
- Quality — more reconstruction per frame. Worth it for short clips where the result matters more than the wait.
- Ultra — the heaviest mode. Reserve it for short, important clips on a capable machine; on long footage the runtime becomes impractical.
Pushing a heavily compressed 480p source to 4K on Ultra is the most common way to be disappointed. The model can only reconstruct from what survived the original encode, and a low-bitrate source simply has less to work with. Aim one step above your source rather than at the largest number available.
Checking the result before you commit
While a video is processing, keyframe previews appear as they are produced. Click one to see the before-and-after for that frame. This is worth doing on the first few frames of any long job: if faces look waxy or text has turned to mush at frame ten, it will look the same at frame three thousand, and you have saved yourself the rest of the run.
You can pause a job and resume it, or cancel it outright. Pausing is useful when you need the machine for something demanding — processing competes for exactly the resources a video call or a build needs.
How long it actually takes
Honest answer: it depends enough on your hardware that any single number would be misleading. What reliably drives it is frame count (length × frame rate), source resolution, target resolution, and processing mode — roughly in that order.
A short clip on Balanced on a machine with a discrete GPU is a matter of minutes. The same clip on Ultra, or a longer one on integrated graphics, can run considerably longer. This is why the short test clip matters: it gives you a real measurement from your own machine instead of an estimate from ours.
Keep the tab open and the machine awake while it runs. Closing the tab, or letting the laptop sleep, ends the job — there is no server holding your place, which is the same property that keeps your footage private.
Audio and what comes out
Veedz AI carries the original audio track through to the output where the browser can copy it. If a particular track cannot be passed through, the tool tells you rather than silently returning a silent video. The finished file downloads directly to your device — like everything else here, it never touches a server.