SANA-Streaming

### TL;DR

SANA-Streaming is an open-source, real-time video-to-video editing model developed by NVIDIA (NVLabs) and hosted via Reactor Technologies. Powered by a hybrid Diffusion Transformer, the framework enables continuous, text-guided edits on both pre-recorded video clips and live webcam feeds without interrupting the playback flow. Whether performing minor adjustments like watermark removal or major overhauls like background replacements, SANA-Streaming preserves the original video's motion and geometry perfectly. Optimized to run locally on NVIDIA’s RTX 5090 GPU architecture, the model delivers crisp 1280×704 resolution outputs in 24-frame chunks with a near-instant latency of 1–1.5 seconds. For users seeking an immediate, hassle-free deployment, Reactor provides comprehensive documentation alongside a managed cloud API service priced at a flexible $6.12 per hour (17 credits/sec).

Key Insights & Metrics

Key Features

• Real-time streaming video editing
• High-resolution, minute-level video-to-video editing
• Hybrid Diffusion Transformer architecture
• Optimized for NVIDIA RTX 5090 GPU
• Achieves 24 FPS at 1280 x 704 resolution
• Open-source with code and demo available

Discussion