// the find
nihui/realcugan-ncnn-vulkan
real-cugan converter ncnn version, runs fast on intel / amd / nvidia / apple-silicon GPU with vulkan
A Vulkan-based command-line port of Real-CUGAN, the anime-style image super-resolution model, built on Tencent's ncnn inference framework. It runs without CUDA or PyTorch on Intel, AMD, Nvidia and Apple GPUs, so it suits people who want to upscale anime art or scans from a shell script or a batch folder without installing a Python ML stack.
- The release is a self-contained folder of binaries and model files. There is no Python or CUDA runtime to install, which is the main reason to pick this over the PyTorch version of the model.
- Vulkan is the GPU path, so one binary covers Intel iGPUs, AMD and Nvidia cards, and Apple Silicon via MoltenVK. Most upscalers of this quality lock you to CUDA.
- Tile size (-t), multi-GPU device lists, and separate load/proc/save thread counts (-j) are exposed as flags, so you can trade VRAM use against throughput instead of accepting one default.
- The model variants (models-se, models-pro, models-nose) and the denoise and scale levels are selected with flags, not code changes, which makes it easy to compare outputs on the same input.
- The last push was 2023-03-12, more than three years ago, and the README still carries an 'early development stage, it may bite your cat' warning. Nothing in the repo shows that current driver and ncnn releases have been tested against it.
- The README's answer to crashes is to upgrade your GPU driver. It describes no fallback beyond '-g -1' for CPU, and says nothing about how slow CPU inference is.
- The interface is the command line. The README documents no library API, so embedding it in another program means shelling out and parsing exit codes and stdout.
- The build section is short. It assumes you already have a Vulkan SDK and the git submodules (ncnn, libwebp) sorted, and it gives no Windows build steps.