🎬 This project generates realistic, text-driven crowd movement animations. It uses trajectory extraction from trace data, applies root masks, and leverages diffusion models (MDM) to generate coherent and diverse paths, producing high-quality animated results.
- 🎯 Extract Root Trajectories (Trace): The system sets the root mask by extracting agent trajectories from real-world/simulated crowd data (Trace). This defines the basic movement path for each agent.
- ✨ Motion Generation (MDM/Diffusion Models): With the root masks fixed, the MDM diffusion model generates detailed human motion along each path, ensuring realistic and coherent agent behaviors matching the crowd context and text prompts.
- 🎞️ Output & Visualization: The resulting motions are animated and visualized as GIFs for easy inspection, with each output demonstrating natural group movement and scenario alignment driven by your text or config input.
Below you can see examples of generated crowd motions and activities:
| 🏃 Running Crowd | 👐 Raising Hands Crowd | 🤼 Fighting Crowd | 🧍 One Person (Reference) |
|---|---|---|---|
 |
 |
 |
 |
- 🎯 Root Mask (from Trace): Sets the spatial path for each agent.
- ✨ MDM Diffusion Model: Fills in realistic, coherent human movement details onto each path, maintaining scene context coherence.
- 🎞️ Animations: Synthesized as shareable GIFs for qualitative assessment.
- Install the requirements (Python 3.x, see code for
numpy,matplotlib,h5py, andmoviepyif you want to generate new GIFs yourself). - Run the pipeline via
integration/src/generate_traces.pyto generate new crowd trajectories and prepare new visualizations.
- TRACE & PACE (NVIDIA):
- TRACE (Trajectory-aware Control for Realistic Animation) is a planner for plausible pedestrian and crowd motion paths that accounts for interactions, group behaviors, and the environment. PACE (Pedestrian Animation with Controllable Environment) synthesizes lifelike human movement along these paths. Learn more
- PriorMDM (Guy Tevet et al.):
- PriorMDM (Prior Motion Diffusion Model) uses diffusion models and a learned prior to generate coherent, realistic motion for complex agents from textual descriptions or other constraints. This provides scene- and prompt-consistent, group-aware movement.
This project: Sets root masks using TRACE-like trajectory data, then generates full agent motion with PriorMDM-like diffusion, resulting in synchronized, semantically-driven crowd animations as illustrated above.
All components referenced above are original; folders such as priorMDM/ and trace/ are provided as external dependencies only.