![[Project] useknockout - a open SOTA background removal + super resolution + face restore API (BiRefNet + Swin2SR + GFPGAN), MIT, Modal deployed](https://external-preview.redd.it/L2udfYmJC4b69uEdZcMxRoMtGdiEowW46fbXUymIK8Q.png?width=1080&crop=smart&auto=webp&s=cecfb76319c68d3574442283cb43c9c86af39af1)
[Project] useknockout - a open SOTA background removal + super resolution + face restore API (BiRefNet + Swin2SR + GFPGAN), MIT, Modal deployed
Built useknockout as a single FREE FastAPI service on Modal that bundles a few SOTA vision models with sane defaults so you do not have to wire them up yourself.
Models:
- Background removal: BiRefNet (ZhengPeng7/BiRefNet, MIT) + pymatting closed form foreground estimation for clean alpha edges
- Super resolution: Swin2SR (caidas/swin2SR-realworld-sr-x4-64-bsrgan-psnr) for photo content, Real-ESRGAN as opt-in for graphics
- Face restoration: GFPGAN v1.4
Endpoints:
- POST /remove, /remove-url, /replace-bg, /remove-batch, /upscale, /face-restore
Infra:
- Modal L4, scale to zero (60s window), weights baked into the image for fast cold starts
- 200 to 300ms per image warm for /remove, 13 to 17s for x4 upscale at 1024px input
- Tiled inference (256px tile, 32px overlap, triangular blend) for arbitrary input sizes
Live: https://useknockout.com
Repo: https://github.com/useknockout/api (MIT)
SDKs: /useknockout/node, /useknockout/react, /useknockout/cli, useknockout (PyPI)
Please try it out for free on the playground and let me know what you think. it does take a second to warmup if it hasnt been ran recently but its been getting good traffic