JD.com launches JoyAI-Echo for long video generation, fixing character consistency and speed issues

JD.com has launched JoyAI-Echo, a framework for generating long videos. It solves common problems like characters changing appearance, voice instability, and slow generation. The system has a memory bank to keep character looks and voice consistent across multiple shots. Tests show stable identity and voice in a 5-minute video.

The framework uses a memory-driven training method with SFT, cross-modal RLHF, and DMD technology. DMD alone speeds up generation by about 7.5 times. It also includes a Director Agent that can break down natural language requests into scripts, characters, scenes, and shots.

JoyAI-Echo has a real-time upscaling module that supports two resolution upgrades, up to 1472×2560. JD.com says this puts the company in the global top tier for long video generation. The project is open source.