ANVIL — Accelerator-Native Video Interpolation

Pretrained weights and ONNX models for the ANVIL video frame interpolation system.

ANVIL achieves 1080p INT8 real-time frame interpolation on mobile NPUs by replacing learned optical flow with codec motion vector prealignment, producing an NPU-friendly inference graph dominated by standard convolutions.

Models

Model	Params	Vimeo90K PSNR	Xiph 1080p PSNR	HTP V75 INT8 (avg)
ANVIL-S	855K	33.45 dB	29.65 dB	12.8 ms
ANVIL-M	2.66M	33.66 dB	29.74 dB	16.7 ms
NAFNet-ceiling	17.1M	34.58 dB	30.30 dB	—

Capacity sweep models (D-tiny through D-unet-l, 1.8K–289K params) are included under checkpoints_capacity/.

File Structure

checkpoints/
  D-unet-v3bs-nomv/best.pt     # ANVIL-S (10 MB)
  D-unet-v3bm-nomv/best.pt     # ANVIL-M (31 MB)
  D-nafnet-nomv/best.pt         # NAFNet ceiling (197 MB)
checkpoints_capacity/
  D-tiny-nomv/best.pt           # 1.8K params
  D-mini-nomv/best.pt           # 10.6K params
  D-mid-nomv/best.pt            # 33K params
  D-unet-s-nomv/best.pt         # 129K params
  D-unet-l-nomv/best.pt         # 289K params
onnx/
  D_unet_v3bs_nomv_1080p.onnx   # ANVIL-S 1080p (BN fused)
  D_unet_v3bm_nomv_1080p.onnx   # ANVIL-M 1080p (BN fused)

Usage

git clone https://github.com/NihilDigit/anvil && cd anvil
pixi install
pixi run download-weights   # downloads to artifacts/
pixi run reproduce-main-quality

Links

• Paper repository: github.com/NihilDigit/anvil
• Android demo: github.com/NihilDigit/mpv-android-anvil

License

MIT