open-svpflow

open-svpflow is a drop-in open-source replacement for SVPFlow's VapourSynth plugins in Rust, reverse engineered from version 4.3.0.168. It runs up to 3x faster on CPU and up to 2.6x faster on GPU, adds a few things the original doesn't have, and works in places where the original doesn't.

It started as a side project, mostly to see if I could get couleur's smoothie-rs running in a browser. That became the demo, which runs open-svpflow client-side as WASM with WebGPU (about 2-3x slower than native, since WASM can't use AVX2). It was only meant as a proof of concept, but it grew into another project, interpolini, which is meant as a showcase of how much faster you can get without VapourSynth and Python in the way.

Parity with SVPFlow

• svpflow1: Super and Analyse, including super-frame pyramids, multi-level predictors, SAD/SATD and chroma costs, and hex2, UMH, and exhaustive motion search.
• svpflow2: SmoothFps rendering for algorithms 1, 2, 11, 13, 21, 22, and 23, with scene handling, masks, CPU rendering, and the GPU path.
• AviSynth+ support with the original function names.

Output matches the original byte for byte in everything I've tested, on both CPU and GPU. GPU algorithms 1 and 2 are the exception, since the original reads uninitialized memory there and doesn't even match its own previous run (±1 on about 0.01% of pixels, same as ours). So it should work anywhere SVPFlow already does, whether that's smoothie, SVP itself, mpv or your own scripts.

Benchmarked on a 1080p 200 fps Counter-Strike 2 clip interpolated to 1920 fps with smoothie's settings (Ryzen 5 7600X, RTX 4060, Linux). You can run it yourself with scripts/bench.py:

The speedup mainly comes from AVX2/AVX-512 motion search, running part of that search on the GPU (when using GPU mode), and caching work the original redoes for every frame.

On top of SVPFlow

• VapourSynth API 4 support, so it works on current VapourSynth (R80 and newer can't load the original at all, since it only has API 3).
• Native I444 input and output.
• SmoothFpsBlend, which interpolates and frame-blends on the GPU in one go, so it's about 2x faster than blending separately (build with --no-default-features if you don't want it).
• Halve and DoubleVectors, which let the motion search run on a half-size frame, a lot faster for a small quality loss.
• A C API for running it without VapourSynth or AviSynth (cargo build --release -p svpflow-capi).

SmoothFps options

All of these are off by default.

• mask:{peel:true} learns the HUD and its opacity for each pixel from how the scene moves behind it, takes it off before the warp and puts it back after. Tune it with mask:{peel:{limit:6}}.
• mask:{still:true} finds text, logos and other things that stay in place and keeps them from smearing. Tune it with mask:{still:{limit:4.5,edge:5,tolerance:30}}. svp2.Still(clip, source) does the same for frames from any other interpolator.
• dedup:0.5 skips a frame that repeats the one before it and stretches the next motion over both frame times. The number is how different two frames can be and still count as a repeat.

Install

pip install open-svpflow

It installs with an assumption that you already have VapourSynth installed in the same Python environment (we deliberately not include it as a dependency). Wheels are available for Windows x64, Linux x64 and arm64, and macOS x64 and arm64.

AviSynth

The plugins also load in AviSynth+ 3.6 or newer:

LoadPlugin("svpflow1_vs.dll")
LoadPlugin("svpflow2_vs.dll")
super = SVSuper("{gpu:1}")
vectors = SVAnalyse(super, "{}")
SVSmoothFps(super, vectors, "{rate:{num:5,den:2}}", mt=8)

Releases

Download the latest nightly builds for Windows, Linux, and macOS from Releases. Versioned releases follow semver and are published to PyPI.

Releases are reproducible and immutable once published. To reproduce the binaries, check out the release commit in a clean directory and use the matching build environment from the build workflow. Run build.sh with the target from the release's BUILD.txt, then compare the files in dist with the extracted release.

Build

Rust 1.88 or newer:

cargo build --release -p svpflow1 -p svpflow2

The plugins end up in target/release: svpflow1_vs.dll and svpflow2_vs.dll on Windows, libsvpflow1_vs.so and libsvpflow2_vs.so on Linux, .dylib on macOS.

Licensed under Apache-2.0.