zjwfufu/smd is a Python repository collecting distillation methods that reduce the sampling steps of diffusion models. A diffusion model updates samples several times over to produce a result; distillation is a training method that reduces the number of those updates. This repository covers image, video, audio-video and 3D generative models in one place, and the README states that the default setup uses four sampling steps. The license is Apache-2.0, and the repository was created on October 7, 2026.
The repository description introduces this approach as a simplification and improvement of on-policy distillation. The README organizes model-specific training setups and inference procedures for FLUX, Qwen-Image, Wan, MiniMax H3, Hunyuan3D and others. It lists setups separately for image generation and editing, video generation, audio-video generation, and generating 3D shapes from images. Installation, inference and training procedures assume a Linux x86_64 environment with a CUDA-compatible NVIDIA GPU, and guidance on exporting model weights is included as well. Links to data preparation documents and code, and to public model and dataset resources, are provided too. It has 222 stars.
However, the public README alone does not allow experimental reproducibility or a performance advantage to be confirmed. The research record scores are novelty 8.0, technology 7.0, quality 7.0 and momentum 5.0. Technology and quality were rated conservatively. There is no formal release yet, and no performance comparison figures are presented. For now, it is better to approach it as reference material for training and inference setups and code.