SeedVR Image Upscaler
AI Image Upscaler: Upscale Any Image to 4K with SeedVR2
Enlarge any image up to 10x its original size, or straight to a target resolution like 1080p or 4K. Upload a photo, screenshot, or AI-generated image and get back a sharper, cleaner, bigger version.
It runs on SeedVR2, ByteDance’s diffusion-based restoration model. Instead of stretching pixels the way classic upscalers do, it rebuilds the image at the new size: compression artifacts get cleaned up, soft edges come back sharp, and fine texture gets restored.
What can you do with it?
- Upscale AI-generated images. Most models output around 1 megapixel. Take a generation to 2K or 4K before publishing or printing.
- Rescue old or low-res photos. Small originals, heavy JPEG compression, blurry scans: SeedVR2 restores detail while it enlarges.
- Prepare images for print. A web-res image rarely survives print DPI. Upscale it first and it will.
- Sharpen screenshots and web grabs. Enlarge a small asset you found or captured without the mush.
- Fix thumbnails and social exports. Recover a usable version when the original file is gone and all you have is the compressed copy.
How to use it
- Upload your image. Any common format works: PNG, JPG, WebP.
- 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 the image in one pass, usually within seconds.
- Download the result, or feed it into your next step right here.
Settings
- Upscale factor: how much to multiply the original dimensions. 2x quadruples the pixel count.
- Target resolution: skip the math and name the output size instead (720p up to 4K).
- Noise scale: how much freedom the model gets to invent detail. The 0.1 default stays faithful to the original; raise it for heavily degraded inputs that need more reconstruction.
- Output format: JPG by default, with PNG and WebP available.
- Seed: set one to make a run reproducible.
Tips
- 2x to 4x is the sweet spot. It’s enough for most print and publishing work, and detail stays believable. Past that, the model has to invent more.
- Feed it your best source. Restoration works with what’s there; a slightly bigger original beats a tiny one every time.
- Degraded input? Nudge noise scale up. Around 0.2 to 0.3 gives the model room to rebuild texture on very compressed or blurry images.
- Chain it after generation. Generate at the model’s native size, then upscale here for the final export. It’s faster and cheaper than forcing generators to render large.