The project merged a CUDA optimization for supported INT8 linear layers. Adding the bias in the same kernel that finishes the matrix multiplication avoids a separate pass over the intermediate output.
stable-diffusion.cpp
Also known as: stable diffusion cpp · sd.cpp · sd-cli
Plain English
stable-diffusion.cpp runs image and video generators from the command line or an embedded library. It supports model families such as Stable Diffusion, FLUX, and Wan.
Technical Overview
The C and C++ runtime executes diffusion and flow models using ggml backends. It can load supported GGUF weights and run model components such as text encoders, denoisers, and autoencoders. Which models fit and how fast they run depend on the backend, precision, and available memory.
Chronological Entries (3)
The project merged support for Wan 2.2 S2V. A reference portrait and audio clip can condition video generation, with the input audio included in the output container.
The project merged a Vulkan backend for its ggml operations. This gave GPU deployments a path through Vulkan on compatible hardware alongside platform-specific backends.
Direct Relations (2)
- Part ofInference Engines
- UsesGGUF