Five submissions, four rejections, one ECCV paper, and a lot of slow motion

What it does RIFE is a PyTorch implementation of a video frame interpolation network that synthesizes intermediate frames between two existing images or video frames. It supports arbitrary-timestep interpolation—meaning it can generate frames at any point between A and B, not just the midpoint—to produce 2×, 4×, or even 16× frame rate increases. The authors report it can run at 30+ FPS for 2× 720p interpolation on an RTX 2080 Ti.

The interesting bit The paper endured rejections from CVPR 2021, ICCV 2021, AAAI 2022, and CVPR 2022 before finally landing at ECCV 2022, and the project has since become foundational infrastructure for an entire ecosystem of video tools. Beyond the core research code, the community has spun off practical variants—including anime-tuned models and a lightweight release specifically for post-processing diffusion-generated video—plus third-party integrations ranging from VapourSynth plugins to commercial GUI wrappers.

Key highlights

• Real-time performance: claims 30+ FPS for 2× 720p on an RTX 2080 Ti.
• Arbitrary-timestep interpolation between frame pairs, enabling flexible slow-motion and upscaling factors.
• A sprawling derivative ecosystem: Practical-RIFE adds user-facing features, while third-party tools like FlowFrames and ncnn/Vulkan ports extend reach.
• Specialized model variants: v4.7–4.10 target anime scenes using SAFA research, and 4.22.lite is tuned for diffusion-model video cleanup.
• Can also be repurposed for optical flow estimation and video stitching, according to issue discussions linked in the README.

Caveats

• The pretrained HD model provided for easy inference is not the exact paper model; the README directs users to separate evaluation weights for benchmark reproduction.
• macOS support currently depends on an unmerged pull request, so first-class cross-platform compatibility is unclear.
• Applying slow-motion effects strips audio from the output, which the README notes in passing.

Verdict Researchers and video-restoration tinkerers—especially those upscaling anime or cleaning up AI-generated footage—will find a mature, well-trafficked base to build on. Casual editors who just need to re-encode a clip should probably stay in ffmpeg territory.