SANA-Streaming: Real-time Streaming Video Editing with Hybrid Diffusion Transformer
Researchers have developed SANA-Streaming, a framework for real-time video editing on consumer GPUs. It utilizes a hybrid diffusion transformer architecture with attention mechanisms for improved local modeling and efficiency. The system also incorporates a novel cycle-reverse regularization technique to enhance temporal consistency without needing paired long videos. Optimized for NVIDIA Blackwell architecture, SANA-Streaming achieves 24 FPS editing at 1280x704 resolution on a single RTX 5090 GPU. AI
IMPACT Enables real-time, high-resolution video editing on consumer hardware, potentially impacting live broadcasting and gaming applications.