JD.com open-sourced JoyAI-Echo, a long video generation model that produces 5-minute videos with strong cross-shot consistency. The standout: the model is the first to solve the "cross-shot consistency" problem at 5-minute length, and uses DMD (Diffusion Model Distillation) to achieve 7.5× speedup over the base model.
The "cross-shot consistency" problem: when generating a long video, the model must switch between "shots" (different camera angles, different scenes), and the switch must be consistent — objects and characters must look the same across shots. Previous video models lose consistency across shot transitions, leading to "shifts" in appearance.
The JoyAI-Echo fix: a "scene graph" representation that tracks the appearance of all objects and characters across the video. The scene graph is updated at every frame, and the generation model conditions on the scene graph, ensuring consistency. The "5-minute" length is achieved by a hierarchical generation process — the model first generates a "coarse" 5-minute video, then refines each segment.
The "DMD distillation" highlight: the base JoyAI-Echo model is 30B parameters, with 7.5× slower inference than real-time. The DMD distillation produces a 4B distilled model that runs at 7.5× speed (real-time for 5-minute videos), with quality loss of less than 5%.
The benchmark: on the "long-video quality" benchmark, JoyAI-Echo hits 81.2, on par with Sora 2. The "cross-shot consistency" score is 87.4, the highest among all evaluated models. The distilled 4B model hits 76.8 — still higher than most 5-minute video models.
The bigger takeaway: "scene graph + DMD distillation" is the right architecture for long video generation. The "one big model" approach is too slow, and the "graph + distilled" approach is both high quality and fast. For the industry, this signals that "long video generation" will move to hierarchical + distilled architectures, and the next round of competition will be in "how smart the scene graph is."