StreamDiffusion

StreamDiffusion is an open-source pipeline-level solution that turns diffusion image models into real-time, interactive generators. By rethinking how consecutive frames are batched and denoised, it lets Stable Diffusion and SDXL respond to changing prompts and inputs at interactive frame rates — fast enough for live installations, streaming, and creative tools.

Overview

The framework’s key innovation is Stream Batch: instead of generating each image independently, it staggers denoising steps across a batch of frames so every pass produces a finished frame, dramatically improving throughput. A residual autoencoder (R-CFG) cuts the cost of classifier-free guidance, stochastic similarity filtering skips redundant computation on near-identical frames, and an IO queue with pre-computation keeps the GPU fed. Combined with TensorRT acceleration and tiny autoencoders, StreamDiffusion can reach tens of frames per second on a single high-end GPU for image-to-image generation. The project comes from researchers at the University of Tsukuba and collaborators, has more than 10,800 GitHub stars, and was last updated in December 2024. It is Apache-2.0 licensed.

Key Features

• Real-time interactive image generation with Stable Diffusion and SDXL
• Stream Batch scheduling overlaps denoising for far higher frame rates
• Residual CFG and similarity filtering reduce wasted computation
• TensorRT and tiny-autoencoder support for maximum throughput
• Works for live img2img, webcam effects, and interactive installations
• Research-grade code with demos and reproducible configurations

💳 Pricing

StreamDiffusion is free and open source under the Apache-2.0 license. You only need your own GPU hardware to run it.

Who Is It For

Creative technologists, live-stream producers, and researchers building real-time generative visuals, interactive exhibits, or AI video effects.