DyCraft: An Agentic System for Long-Horizon 4D Scene Generation
DyCraft checks key 3D poses before building the motion between them, reducing physical errors in long animations.
The system turns a prompt into a detailed plan and static scene, then builds a 3D storyboard whose poses can be debugged individually. Motion tools connect the verified start and end states, avoiding the difficulty of scripting an entire animation in one pass. Across 24 long-horizon, multi-shot tasks, the authors report fewer physical violations and higher-quality scenes than the tested agent baselines. Human raters also preferred DyCraft’s results more often.
HF Daily Papers' note
The system turns a prompt into a detailed plan and static scene, then builds a 3D storyboard whose poses can be debugged individually. Motion tools connect the verified start and end states, avoiding the difficulty of scripting an entire animation in one pass. Across 24 long-horizon, multi-shot tasks, the authors report fewer physical violations and higher-quality scenes than the tested agent baselines. Human raters also preferred DyCraft’s results more often.
HF Daily Papers' note
score 5