NanoVSR

Pretrained checkpoints for NanoVSR, a lightweight video super-resolution (VSR) model for real-time inference on edge devices. Accepted to ECCV 2026.

NanoVSR uses a bidirectional recurrent design with reparameterizable multi-branch blocks that collapse into plain 3×3 convolutions at inference — no custom ops, ONNX/TensorRT compatible out of the box.

Code: github.com/filippawlicki/nanovsr

Models

All models perform 4× upscaling.

Model	Params	REDS4 (PSNR/SSIM)	Vid4 (PSNR/SSIM)	Vimeo-90K-T (PSNR/SSIM)	Orin NX 25W (FPS)
nanovsr_226k.pth	226k	28.23 / 0.8057	25.26 / 0.7252	34.31 / 0.9130	43.86
nanovsr_644k.pth (baseline)	644k	28.64 / 0.8215	26.05 / 0.7761	35.00 / 0.9226	27.20
nanovsr_1.7m.pth	1.7M	29.15 / 0.8364	26.44 / 0.7964	35.49 / 0.9294	19.58
nanovsr_5.4m.pth	5.4M	29.73 / 0.8526	26.76 / 0.8089	35.85 / 0.9335	8.66