JoyAI Releases Real-Time Video Editing Model

JoyAI-Video-Edit is a real-time video editing model that processes frames as they arrive, applying natural language instructions without pre-loading the entire video. It operates causally without referencing future frames, making it suitable for indefinite-length streams such as live camera feeds.

The system consists of the following components:

• MLLM-based conditional encoder
• Causal video VAE
• 16B parameter MM-DiT
• Chunk-wise autoregressive adaptation
• Source-anchored distribution matching distillation (SA-DMD)
• Long-horizon autoregressive distillation (LHAD)

For deployment, the system utilizes bounded KV cache, FP8 quantization, and an optimized VAE path. Reported deployment benchmarks show a full pipeline speed of 30.19 FPS at 720×1280 resolution. In comparisons on OpenVE-Bench, it achieved 37.21 FPS with an editing quality score of 3.60, demonstrating high quality among streaming editing models.

The code, technical report, checkpoints, online demo, and real-time editing live demo have been released. The key achievement is not necessarily superior peak quality compared to offline editing models, but rather the balance between quality and speed under real-time processing constraints.