Video2X is a machine-learning video super-resolution and frame-interpolation framework that has been running since a 2018 hackathon and now sits at 20,899 stars and 1,811 forks, back on GitHub Trending this week. Version 6.0.0 rewrote the entire project from Python into C/C++, which is what turned it from a fragile script into a cross-platform desktop application — the author's own summary of the rewrite is that it "genuinely works this time."

The framework runs two modes. Filtering upscales, backed by Anime4K v4 plus any MPV-compatible custom GLSL shader, and by Real-ESRGAN and Real-CUGAN through ncnn and Vulkan. Frame interpolation raises framerate through RIFE, with all published RIFE models supported; the current default was moved to rife-v4.26 in the latest commits. Because everything runs on ncnn/Vulkan rather than CUDA, it works across NVIDIA, AMD, and Intel GPUs equally, and the pipeline needs zero additional disk space during processing — only room for the final output, which matters when a two-hour upscale would otherwise stage terabytes of intermediate frames.

Distribution is unusually thorough for a project of this kind. Windows gets a signed installer and a Qt6 GUI localized into six languages (English, Simplified Chinese, Japanese, Portuguese, French, German). Linux gets AUR packages, archlinuxcn packages, and a universal AppImage. There are container images on the GitHub Container Registry for one-command use on Linux and macOS, and a maintained Google Colab notebook for anyone without a capable GPU, borrowing a T4, L4, or A100 for up to 12 hours a session. Full documentation lives at docs.video2x.org covering building, installing, running, and developing.

Hardware requirements are modest but firm: the precompiled binaries need AVX2 (Intel Haswell / AMD Excavator or newer), and the GPU must support Vulkan — NVIDIA Kepler, AMD GCN 1.0, or Intel HD Graphics 4000 and newer all qualify. The project bundles or depends on FFmpeg, Tencent's ncnn, Anime4K, and the ncnn-vulkan ports of Real-CUGAN, RIFE, and Real-ESRGAN, with every dependency license enumerated in the NOTICE file.

Two caveats are worth stating plainly. First, the license is AGPL-3.0, not the permissive MIT or Apache terms most of this catalogue carries — network-facing deployments inherit source-disclosure obligations, so it is a poor fit for a closed hosted service even though it is ideal for local and desktop use. Second, development has slowed considerably: the newest tagged release is 6.4.0 from January 2025, and although the main branch received real work in early 2026 — a nullptr-dereference and resource-leak fix, the RIFE default bump, CI upgrades — anyone installing from the releases page is running a build well over a year old. The models it wraps are also anime-and-illustration-leaning by heritage; Real-ESRGAN generalizes to live action, but Anime4K and Real-CUGAN are explicitly tuned for animation.