SeedVR Video Upscaler
Upscale any video up to **10x its original resolution**, or straight to a target like 1080p or 4K.
AI Video Upscaler: Upscale Video to 1080p, 2K, and 4K with SeedVR2
Take any video up to 10x its original resolution, or straight to a target like 1080p or 4K. Upload a clip and get back a sharper, cleaner version with the same motion and timing.
It runs on SeedVR2, ByteDance’s diffusion-based video restoration model. It rebuilds each frame at the new size while keeping frames consistent with each other, so edges stay sharp across motion and you don’t get the shimmering or flicker that frame-by-frame upscalers produce.
What can you do with it?
- Upscale AI-generated video. Most video models output 720p or lower. Bring a generation up to 1080p or 4K before publishing.
- Restore old footage. Phone clips from 2012, screen recordings, archive material: SeedVR2 cleans compression artifacts and noise while it enlarges.
- Deliver in 4K. Client wants a 4K master and you shot in 1080p? Upscale the edit and ship it.
- Fix heavily compressed downloads. Recover a watchable version of a clip that only survives as a low-bitrate copy.
- Sharpen zoomed or cropped shots. Punch in on footage, then upscale the crop back to full delivery resolution.
How to use it
- Upload your video. MP4, MOV, WebM, any common format.
- Pick your size. Set an upscale factor from 1x to 10x (2x is the default). Or switch to target mode in the settings and choose 720p, 1080p, 1440p, or 4K directly.
- Run it. SeedVR2 restores and enlarges every frame while keeping them temporally consistent. Longer clips take longer, so expect a short wait.
- Download the result, or send it into your next step right here.
Settings
- Upscale factor: how much to multiply the original dimensions. 2x quadruples the pixel count of every frame.
- Target resolution: name the output size directly instead (720p up to 4K).
- Noise scale: how much freedom the model gets to invent detail. The 0.1 default stays faithful; raise it for rough, heavily compressed footage.
- Output format: MP4 (H.264) by default. WebM, ProRes 4444 for editing pipelines, and GIF are also available.
- Quality and write mode: trade encode speed against file size, from fast-and-big to slow-and-small.
- Seed: set one to make a run reproducible.
Tips
- 2x is the workhorse. 720p to 1440p, or 1080p to 4K with target mode, both look great. Bigger jumps work but give the model more to invent.
- Trim first. Upscaling processes every frame, so cutting your clip down to the part you need saves time and money.
- Use ProRes for editing. If the upscale goes back into a timeline, ProRes 4444 avoids a second round of compression.
- Rough source? Raise noise scale. Around 0.2 to 0.3 helps the model rebuild texture on badly compressed footage.
- Chain it after generation. Generate video at native resolution, then upscale here for the final export. Cheaper than rendering large, and often sharper too.