How it works
The whole pipeline, including the parts that are limitations rather than features.
The pipeline, step by step
The free tools run entirely inside the browser tab. Nothing they process is sent to a server, so a file you drop in never leaves your machine. The optional cloud models are the one exception and say so before they run: your file is uploaded for that run, and the result is kept for 30 days unless you delete it.
- Decode. For video, the container is demuxed and its encoded samples are handed straight to the browser’s video decoder, frame by frame — no player element, and no waiting for playback. For images, the browser decodes the file into a bitmap.
- Upload to the GPU. Each frame is copied into GPU memory as an 8-bit RGBA texture.
- Resample. Two shader passes enlarge the frame — horizontally first, then vertically — using a Lanczos-3 windowed-sinc filter. Intermediate results are kept in 16-bit float so the filter’s overshoot is preserved rather than clipped.
- Sharpen. A third pass applies an unsharp mask clamped to the local minimum and maximum of each pixel’s neighbours. The clamp is what prevents halos along high-contrast edges.
- Encode. Video frames are re-encoded with your device’s hardware H.264 (or VP9) encoder and written into a new MP4 in memory. Images are written as PNG so the result is not immediately degraded by lossy compression.
What “AI upscaling” means here — and what it does not
Plenty of tools in this category describe any enlargement as AI. We would rather be exact: the pipeline described above is classical signal processing — a high-quality reconstruction filter plus adaptive sharpening. It is a genuine improvement over the bilinear scaling a browser or video player does by default, particularly on edges, and it needs no model download, so it starts instantly and works offline.
It is not a neural network hallucinating detail that was never in the source. Trained networks do sit alongside it — 13 of the 27 methods in the catalogue are one, from Anime (Anime4K, 924 KB) up to the Real-ESRGAN photographic models. Each is labelled as generating detail rather than recovering it, and none of them is a default: the classical pipeline still is, which is why the tool starts instantly and works offline.
Honest limits
- Detail cannot be invented. Upscaling 480p by 4× gives you a clean 1920×1080, not a 4K master. Sharp-but-small sources improve most; heavily compressed or motion-blurred sources improve least.
- Output size drives the time. Frames are demuxed and decoded directly, so the job is not tied to the clip’s running time — but every doubling of the output width quadruples the work.
- Only codecs WebCodecs can decode. MP4, WebM, MOV and MKV containers all open; the codec inside is what matters. H.264 and VP9 are safe everywhere, H.265 and ProRes usually are not.
- Audio is copied, not re-encoded. The encoded audio moves from the source container to the output as-is, so it loses nothing — unless the output container cannot hold that codec.
- HDR shifts colour. 10-bit sources are tone-mapped by the browser before we see them.
- Phones are constrained. Mobile GPUs have far less memory; long clips and 4K output often fail. The page tells you what your device can handle before you choose a file.
Requirements
WebGPU is required for the upscaling itself, and WebCodecs additionally for video. Desktop Chrome and Edge have both today. Safari 26 has WebGPU. Firefox is still rolling WebGPU out. If your browser is missing either, the page says so before you pick a file rather than after you have waited for an upload.