AI-powered MMD pose generator - Transform natural language into expressive 3D character animations
PoPo uses fine-tuned LLMs to generate MMD character poses from natural language descriptions. Instead of training on raw quaternions, we use MPL (MMD Pose Language) - a semantic, MMD-specific pose description language that helps AI understand and generate anatomically correct poses.
🌐 Live demo: popo.love
Demo model: 深空之眼 三相·梵天「无间玩伴」
- Natural Language Input: "wave right hand with big laugh, inviting me for dinner"
- LLM-Generated Poses: Fine-tuned models output semantic MPL code for precise pose control
- Real-time Rendering: Instant pose creation with smooth bone animations
- MMD-Specific: Built for anime characters with proper bone constraints and physics
PoPo fine-tunes LLMs with MPL: MPL is a semantic pose description language designed specifically for MMD. This approach provides:
- Better training convergence - Structured, human-readable pose descriptions
- Consistent outputs - Same prompt generates reliable pose code
- Anatomically correct - Built-in constraints prevent impossible movements
- Debuggable results - Generated MPL code can be read and modified
{
"messages": [
{ "role": "system", "content": "Generate MMD Pose Language (MPL) script from description." },
{ "role": "user", "content": "Description: arms down" },
{ "role": "assistant", "content": "arm_l bend forward 40;arm_r bend forward 40;" }
]
}- Frontend: Next.js, shadcn/ui, TypeScript
- 3D Engine: Babylon.js with babylon-mmd
- Pose Language: MPL (MMD Pose Language) for semantic pose description
- AI Model: Fine-tuned GPT-4o-mini for natural language → MPL generation
- Deployment: Vercel
- MiKaPo: Camera → MediaPipe → MMD bones (real-time capture)
- PoPo: Text → Fine-tuned LLM → MPL code → MMD bones (AI-generated poses)
By using semantic MPL as the training target instead of raw quaternions, we achieve better consistency and allow the AI to learn the "grammar" of human movement.
GPL-3.0 License - see LICENSE for details.