Motion Generation Demo

Local FastAPI demo for text-to-motion visualization. The local app owns only the browser UI, topology endpoints, and WebSocket forwarding. Actual G1 motion generation runs in the remote Kimodo API.

Run

Start or tunnel the remote Kimodo API first, then run the local demo:

KIMODO_G1_API_URL=http://127.0.0.1:9001 \
FLOOD_DIFFUSION_API_URL=http://127.0.0.1:9002 \
FLOOD_DIFFUSION_CHECKPOINT_DIR=/mnt/data/cpfs/haiyang/FloodDiffusion-Dev/outputs/20260526_003412_chunk_dit_201d_concat_mix_l2_l3_random_mini16_beat_test128 \
KIMODO_G1_SEED=11 \
KIMODO_G1_DIFFUSION_STEPS=20 \
uvicorn app:app --host 127.0.0.1 --port 7861

Open:

http://127.0.0.1:7861/

Current Code Shape

app.py
  # Local FastAPI entrypoint
  # Serves the page
  # Provides /api/config, /api/g1/topology, /api/smplx/topology
  # Provides WS /api/offline
  # Calls remote Kimodo G1 / Flood Diffusion APIs and forwards binary motion frames

templates/index.html
  # Single demo page

static/css/style.css
  # Page styling

static/js/main.js
  # UI state
  # Reads online/offline text input and offline_audio uploads
  # Opens /api/offline websocket
  # Receives motion frames
  # Handles playback, canvas/audio recording, status, latency, seed, smooth

static/js/avatars/
  # Three.js avatar display code
  # G1 mesh avatar
  # SMPL-X avatar

renderers/
  # Python renderer adapters
  # G1 topology + binary frame encoding
  # SMPL-X topology + compact params/binary frame encoding

Local API

Full API details are in docs/api.md.

GET /api/config
GET /api/g1/topology
GET /api/smplx/topology
POST /api/recordings
WS  /api/offline

/api/offline expects the first WebSocket message to contain a schedule and config:

{
  "renderer": "g1",
  "config": {
    "seed": 11,
    "smooth": 0
  },
  "schedule": [
    { "text": "walk forward", "start": 0 },
    { "text": "dance", "start": 4, "end": 8 }
  ]
}

For audio-driven offline motion, use input_mode=offline_audio, renderer=smplx, and schedule rows with audio plus optional text, text_start, and text_end. The server derives duration from the uploaded audio. Audio longer than 10 seconds is split into multiple chunks before forwarding to the Flood Diffusion chunkdit_concat_201d_audio API. The browser plays the uploaded local audio itself and starts it when the first returned motion frame is applied, so audio frame 0 and motion frame 0 share the same playback origin.

{
  "input_mode": "offline_audio",
  "renderer": "smplx",
  "config": {
    "seed": 11,
    "smooth": 0
  },
  "schedule": [
    {
      "audio": {
        "name": "clip.wav",
        "mime_type": "audio/wav",
        "data": "<base64-or-data-url>"
      },
      "text": "wave during the middle of the clip",
      "text_start": 1.2,
      "text_end": 4.8
    }
  ]
}

The WebSocket returns JSON lifecycle events plus binary motion frames.

Assets

Large model and mesh assets are not committed.

• Unitree G1 local renderer assets: .g1_cache/g1skel34/
• SMPL-X local renderer model: .smplx_cache/SMPLX_NEUTRAL_2020.npz
• Existing SMPL-X copies on the motion server: /mnt/data/cpfs/motion_data/smplx_models/smplx/SMPLX_NEUTRAL_2020.npz /mnt/data/cpfs/motion_data/smplx_models/smplx/SMPLX_NEUTRAL.npz
• Remote Kimodo G1 checkpoint path: /mnt/data/cpfs/haiyang/kimodo-api/checkpoints/Kimodo-G1-RP-v1
• Flood Diffusion offline audio checkpoint directory: /mnt/data/cpfs/haiyang/FloodDiffusion-Dev/outputs/20260526_003412_chunk_dit_201d_concat_mix_l2_l3_random_mini16_beat_test128

Requirements

Install Python deps with:

pip install -r requirements.txt