StreamDiffusion

StreamDiffusion turns diffusion image generation into a real-time, interactive pipeline. Instead of optimizing a single model, it restructures the whole inference loop: a stream batch technique processes denoising steps for consecutive frames together, Residual Classifier-Free Guidance cuts redundant guidance computation, and a stochastic similarity filter skips frames that barely change, saving GPU work on static scenes. Combined with TensorRT acceleration and pre-computed caching, the pipeline reaches over 100 frames per second for text-to-image with SD-Turbo on an RTX 4090, with image-to-image close behind. The open-source project, which originated as academic research, installs via pip, conda, or Docker and supports both txt2img and img2img modes. It is used by developers building live creative tools such as webcam filters, interactive art installations, and other applications where a diffusion model must respond continuously rather than render one image at a time.