managarten/services/mana-video-gen/CLAUDE.md

CLAUDE.md - Mana Video Generation Service

Service Overview

AI video generation microservice using LTX-Video via HuggingFace diffusers:

• Port: 3026
• Framework: Python + FastAPI
• Model: LTX-Video (~2B params, Lightricks)
• Backend: diffusers + PyTorch CUDA
• Target Hardware: NVIDIA RTX 3090 (24 GB VRAM)

Features

• Fast generation: 10-30 seconds per clip on RTX 3090
• Text-to-video: 480p-720p, up to ~6 seconds
• Low VRAM: ~10 GB — leaves room for other GPU services
• Lazy model loading: Model loads on first request, stays in VRAM
• VRAM management: POST /unload to free GPU memory for other services
• MP4 output: Direct video file serving

Commands

# Setup (installs PyTorch CUDA + diffusers + LTX-Video)
chmod +x setup.sh && ./setup.sh

# Development
source .venv/bin/activate
uvicorn app.main:app --host 0.0.0.0 --port 3026 --reload

# Test
curl http://localhost:3026/health
curl -X POST http://localhost:3026/generate \
  -H "Content-Type: application/json" \
  -d '{"prompt": "A cat walking in a garden"}' | jq

# Free VRAM (e.g. before running image generation)
curl -X POST http://localhost:3026/unload

File Structure

services/mana-video-gen/
├── app/
│   ├── __init__.py
│   ├── main.py              # FastAPI endpoints
│   └── ltx_service.py       # LTX-Video diffusers pipeline
├── setup.sh                 # Setup script (CUDA + Python deps)
├── requirements.txt
├── .env.example
└── CLAUDE.md

API Endpoints

Endpoint	Method	Purpose
/health	GET	Health check + GPU info
/models	GET	Model info
/generate	POST	Generate video from text prompt
/videos/{filename}	GET	Serve generated video
/videos/{filename}	DELETE	Delete video
/unload	POST	Unload model, free VRAM
/cleanup	POST	Clean old videos

Generate Request

{
  "prompt": "A timelapse of a flower blooming",
  "negative_prompt": "blurry, low quality",
  "width": 704,
  "height": 480,
  "num_frames": 81,
  "fps": 25,
  "steps": 30,
  "guidance_scale": 7.5,
  "seed": null
}

Generate Response

{
  "success": true,
  "video_url": "/videos/abc123.mp4",
  "prompt": "A timelapse of a flower blooming",
  "width": 704,
  "height": 480,
  "num_frames": 81,
  "fps": 25,
  "duration": 3.24,
  "steps": 30,
  "seed": 42,
  "generation_time": 18.5
}

Environment Variables

Variable	Default	Description
PORT	3026	Service port
LTX_MODEL_ID	Lightricks/LTX-Video	HuggingFace model ID
DEVICE	cuda	PyTorch device
DEFAULT_WIDTH	704	Default video width
DEFAULT_HEIGHT	480	Default video height
DEFAULT_NUM_FRAMES	81	Default frame count (~3.2s)
DEFAULT_FPS	25	Default framerate
DEFAULT_STEPS	30	Default inference steps
DEFAULT_GUIDANCE_SCALE	7.5	Default CFG scale
GENERATION_TIMEOUT	600	Timeout in seconds
MAX_PROMPT_LENGTH	2000	Max prompt chars
MAX_FRAMES	161	Max frames (~6.4s)
CORS_ORIGINS	(production URLs)	CORS config

Model Details

LTX-Video

• Parameters: ~2 billion
• License: Lightricks Open License (commercial use allowed)
• Download size: ~4 GB (auto-downloaded on first use)
• VRAM usage: ~10 GB
• Optimal settings: 704x480, 30 steps, 7.5 guidance
• Speed on RTX 3090: 10-30 seconds per clip

VRAM Management

The GPU server runs multiple AI services. LTX-Video uses ~10 GB VRAM:

• Model loads lazily on first /generate request
• Use POST /unload to free VRAM when not generating videos
• Other services (mana-image-gen, mana-stt, mana-tts) share the same GPU
• enable_model_cpu_offload() moves unused layers to CPU automatically

Performance (RTX 3090)

Resolution	Frames	Steps	Time
512x320	41	20	~8s
704x480	81	30	~20s
704x480	41	20	~10s
1280x720	41	30	~45s

Integration

Used by:

• Picture App — video generation alongside images
• Chat App — inline video generation

Example (TypeScript)

const response = await fetch('http://192.168.178.11:3026/generate', {
  method: 'POST',
  headers: { 'Content-Type': 'application/json' },
  body: JSON.stringify({
    prompt: 'Ocean waves crashing on rocks at sunset',
    width: 704,
    height: 480,
    num_frames: 81,
  }),
});

const result = await response.json();
const videoUrl = `http://192.168.178.11:3026${result.video_url}`;