What it does FLAVR generates intermediate frames between existing video frames, turning 30 FPS footage into 60, 120, or 240 FPS. It offers pretrained models for 2×, 4×, and 8× interpolation, plus a ready-made script to slap a “slomo filter” on arbitrary videos with OpenCV.

The interesting bit Most frame interpolation methods estimate optical flow first, then warp and blend. FLAVR is “flow-agnostic” — it uses spatio-temporal convolutions with channel gating to learn motion directly from the raw frames. The authors claim this sidesteps the usual flow-estimation bottleneck without sacrificing quality; on their Vimeo-90K benchmark they hit 36.3 PSNR versus SuperSloMo’s 32.9.

Key highlights

• Single-shot multi-frame prediction: one forward pass generates all intermediate frames, not just one at a time.
• Inference speed is middle-of-the-pack at 3.10 FPS on 512×512 (faster than DAIN and QVI, slightly behind SuperSloMo).
• Pretrained models available for 2×, 4×, and 8× interpolation via Google Drive links.
• Includes Colab notebook and interpolate.py script for casual video experiments.
• WACV 2023 best-paper finalist; code is heavily based on Facebook’s PyTorch video repo and CAIN.

Caveats

• Baseline model links from other methods (AdaCoF, QVI, DAIN, SuperSloMo) are currently dead — the author lost Google Drive access in April 2024 and warns “don’t count on” replacements.
• Motion-magnification models are promised but not yet available.
• Training dependencies are pinned to fairly old versions: PyTorch 1.5.0, CUDA 10.1, Python 3.7.4.

Verdict Worth a look if you need research-grade frame interpolation with a simpler architecture than flow-based alternatives, or if you want pretrained slomo models that actually run. Skip it if you need the missing baseline comparisons or motion-magnification features today.