AI Video Upscaler — HD to 4K Super-Resolution

A regular resize (lanczos / bicubic / nearest) stretches the existing pixels — the result has the same amount of information, just bigger. Real-ESRGAN is a deep learning model trained on millions of images to generate plausible missing detail: sharp edges, fine textures, realistic skin and foliage. Side-by-side, AI upscale looks dramatically crisper, especially on faces and text.

Because a neural network is run on every single frame. Most tools here decode, change something cheap and re-encode; Real-ESRGAN reconstructs each frame at four times the size, which is a genuine piece of computation per image and does not get faster with a better container or a smaller file. That is also why a ten-second clip is the sensible way to try it before committing a long one.

WebGPU is the API the AI model needs to run on your GPU. It ships in Chrome 113+, Edge 113+, Opera 99+, and recent Yandex Browser — on desktop and Android. Safari has WebGPU behind a flag on macOS. Firefox is shipping it gradually. If your browser doesn't have it, the upscaler falls back to CPU inference — slower, but it works.

Yes — every source frame is upscaled, regardless of frame rate. A 60 fps clip just takes twice as long as a 30 fps clip of the same length. The output preserves the source frame rate exactly.

HDR (HDR10, HLG, Dolby Vision) is automatically tonemapped to SDR rec.709 before the AI sees it — Real-ESRGAN was trained on SDR images and would clip wide-gamut highlights. The result is a clean, correctly-exposed SDR video.

The model runs on your own device through the browser, which is why the speed depends on your hardware rather than on a queue: a machine with GPU acceleration will be several times faster than one without, and a phone will be slower again. The work is per frame, so time scales with length and with the target scale — doubling to 2× is roughly a quarter of the work of 4×. We record anonymous performance events (frame count, scale, inference time) to find bugs and speed things up; what those events contain is described in the Privacy Policy. A practical hint: try thirty seconds first and read the estimate before starting a long clip.

Real-life footage with clear detail benefits the most — faces, text, foliage, fabric textures. Compressed sources (YouTube downloads, phone videos) gain dramatically. Already-sharp 4K sources gain less. Heavily compressed or motion-blurred frames produce less impressive results — the AI can't invent detail that was never captured.

The model is fixed at ×4. From a 1080p source that gives you 4320p (above 4K) — which the tool keeps with the "×4" scale option. From 1440p you'd get 5760p. The "Up to 4K" option caps the final size at 4K for sources where ×4 would be excessive.