// the find
nihui/waifu2x-ncnn-vulkan
waifu2x converter ncnn version, runs fast on intel / amd / nvidia / apple-silicon GPU with vulkan
A C++ command-line image upscaler and denoiser built on the ncnn inference framework and Vulkan, so it runs on any GPU with a working Vulkan driver and doesn't need CUDA. It bundles the original waifu2x models for anime-style art and photos and ships a portable binary, which suits people who want batch upscaling without a Python or CUDA stack.
One Vulkan code path covers Intel, AMD, NVIDIA and Apple Silicon (via MoltenVK on macOS), so there's no CUDA or Caffe runtime to install. The -t tile size and -j load:proc:save thread counts give real control over the VRAM-versus-throughput trade-off, and -g accepts multiple GPU IDs. Model choice (-m), per-level denoise (-n -1 to 3) and TTA (-x) map directly onto the original waifu2x design, so the knobs are visible rather than hidden.
The benchmark tables come from a 2019-era rig (GTX 1070, Windows 10 1809, driver 419.67) and compare against waifu2x-caffe-cui, which is no longer maintained. They say nothing about current GPUs or the current ncnn version, so treat the speed numbers as historical. The models are the original waifu2x architectures, and output on photographs is generally softer than newer super-resolution models like Real-ESRGAN; the README offers no quality comparison to justify the choice. It is CLI-only, with no GUI and no video pipeline, and building from source needs recursive submodule init and the Vulkan SDK on macOS. The crash guidance is essentially 'update your GPU driver', with no diagnostics beyond that.