repo_data/README.md

---
title: Movimento
emoji: �
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: 6.14.0
python_version: '3.12'
app_file: app.py
pinned: true
license: apache-2.0
short_description: Text-driven multi-character motion generation with Qwen LLM planning
---

# Movimento: Multi-Character Motion Generation

**Text-driven interactive motion synthesis** powered by **Qwen LLM planning** and **Kimodo diffusion models** on AMD hardware.

## Features

- 🎭 **Multi-character orchestration**: Synchronize motion for multiple characters in a single scene
- 🧠 **Qwen LLM planning**: Convert natural language prompts to structured motion scripts
- 💾 **BONES-SEED dataset**: Pre-trained motions for realistic human movement
- ⚡ **Real-time visualization**: Viser-based 3D motion preview with playback controls
- 🎯 **Interactive constraints**: Hands-on guidance for character interactions (hand pose, foot contact)
- 🔄 **Smooth transitions**: Automatic blending between motions with multiple transition policies (cut, overlap, hold, smooth)

## Usage

1. **Enter a multi-character scenario**: e.g., "Two characters walk together, then one sits down while the other stands."
2. **Review the motion script**: Qwen parses your prompt into character segments with timing and transitions
3. **Adjust parameters**: Configure transition styles, FPS, and postprocessing options
4. **Generate motion**: Diffusion models synthesize realistic human motion in real-time
5. **Visualize and download**: Preview in 3D, adjust playback speed, export as BVH/FBX

## Architecture

```
User Prompt
    ↓
Qwen LLM (7B/3B/1.5B) - Script Planning
    ↓
DeterministicLoop Scheduler - Multi-character collision detection & resolution
    ↓
CharacterKimodoPlan - Segment→Motion mapping with transition policies
    ↓
Kimodo Diffusion Models - Per-character motion generation with constraints
    ↓
Viser Viewer - 3D preview & pose refinement
```

## Technical Stack

- **LLM Planner**: Qwen2.5-7B via HuggingFace Inference API
- **Motion Model**: Kimodo (NVIDIA Labs) - text-to-motion diffusion
- **Dataset**: BONES-SEED (comprehensive human motion capture)
- **Scheduler**: Deterministic RNG-based conflict resolution for multi-character scenes
- **Infrastructure**: HuggingFace Spaces with AMD GPU support

## Documentation

- [GitHub Repository](https://github.com/RydlrCS/kimodo)
- [Kimodo Paper & Model](https://research.nvidia.com)
- [BONES-SEED Dataset](https://huggingface.co/datasets/bones-studio/seed)

## Citation

If you use Movimento in your research, please cite:
```bibtex
@software{movimento2026,
  title={Movimento: Multi-Character Motion Generation with LLM Planning},
  author={Ted Iro Opiyo},
  year={2026},
  url={https://huggingface.co/spaces/lablab-ai-amd-developer-hackathon/movimento}
}
```

---

**Built for the lablab.ai × AMD Developer Hackathon 2026**