stable-diffusion.cpp: Run Stable Diffusion & FLUX in Pure C++ — No Python Required

Most image-generation stacks force you into a heavyweight Python environment: PyTorch, CUDA wheels, virtualenvs, and gigabytes of dependencies. stable-diffusion.cpp takes the opposite approach — it is a pure C/C++ inference engine built on the ggml tensor library that runs Stable Diffusion, FLUX, Wan, Qwen Image and Z-Image models as a single native binary, with no Python runtime at all.

Why C++ matters for diffusion inference

The ggml foundation gives stable-diffusion.cpp three decisive advantages. First, quantization: GGUF models run at q4_0, q5_0 or q8_0 precision, cutting memory use dramatically with minimal quality loss. Second, portability: the same binary runs on x86 and ARM CPUs, plus CUDA, Vulkan, Metal, SYCL and MUSA GPUs. Third, startup speed — no interpreter warm-up, which makes it perfect for serverless and edge deployments.

Model support

Beyond classic Stable Diffusion 1.5, 2.x, SDXL and SD3.5, the project tracks the fast-moving open-model landscape: FLUX.1, Wan 2.1 video, Qwen Image and Z-Image are all supported. Features include txt2img, img2img, ControlNet, LCM for fast sampling, and LoRA adapters — a genuinely production-ready feature set.

With 6.9K stars and an MIT license, stable-diffusion.cpp is the pragmatic choice for anyone who wants image generation in containers, on bare-metal servers, or on resource-constrained devices — without the Python tax.