JD.com Open-Sources JoyAI-Video-Edit: 24 FPS Real-Time Streaming Video Editing
新智元 · wechat · 2026-08-07
JD.com has open-sourced the JoyAI-Video-Edit model, the first in the video editing track to simultaneously achieve a streaming architecture, real-time speed, and usable quality. The model delivers 30 FPS inference at 720P resolution and supports stable streaming editing for unlimited durations.
- Technical Breakthroughs: Employs a 16B-parameter multimodal diffusion Transformer, using the SA-DMD distillation method to compress per-frame iteration to just two steps. It introduces 'bounded KV state inference' to limit memory usage, solving error accumulation and visual drift in long-video autoregressive inference.
- Performance: Achieves an end-to-end latency of 226ms and 30.1 FPS throughput on a single B200 GPU. Scores 3.60 on OpenVE-Bench, not only dominating streaming models but also matching mainstream offline commercial models like Runway Aleph.
- Applications: Transforms post-production from 'waiting for renders' to real-time on-the-spot operations. Highly applicable for live-stream try-ons and real-time stylization. Furthermore, it efficiently converts human-hand operation videos into robotic arm training data, mitigating the severe scarcity of high-quality physical interaction data for embodied AI.
More from Embodied
- Robot Autonomously Makes Blizzard Ice Cream, Solving Fragile Cup Grip with Five-Finger Hand — Scobleizer · 2026-08-07
- Unitree Accused of Relying on MIT Open Source, But Praised for DTC Strategy — kscottz · 2026-08-07
- Open-Source Web 3D Robot Viewer: Supports URDF and MuJoCo Formats — tom_doerr · 2026-08-07
- KiWear Smart Ring Hands-on: A Universal Remote Reshaping AI Interaction — 智东西 · 2026-08-07
- Deep Dive into LeHome Challenge RL Solution for Vision-Language-Action Models — philfung · 2026-08-07
- Unitree Reportedly Set for IPO at $9B Valuation — zephyr_z9 · 2026-08-07