ai-media - AI Media Generation

Full-stack AI media generation powered by GPU server (RTX 3090/3080/2070S).

Capabilities

1. Image Generation — Photorealistic images via ComfyUI (z-image, Juggernaut XL)
2. Video Generation — Video synthesis via ComfyUI (AnimateDiff, LTX-2)
3. Talking Heads — Animated talking faces via SadTalker
4. Voice Synthesis — Natural TTS via Voxtral (whisper.cpp)

GPU Server

• Host: ${GPU_USER}@${GPU_HOST}
• SSH Key: ~/.ssh/id_ed25519_gpu
• ComfyUI: /data/ai-stack/comfyui/ComfyUI/ (port 8188)
• SadTalker: /data/ai-stack/sadtalker/
• Voxtral: /data/ai-stack/whisper/
• Output: /data/ai-stack/output/

Usage

Generate Image

./scripts/image.sh "lady on beach at sunset" realistic
./scripts/image.sh "cyberpunk cityscape" artistic

Arguments: - $1: Prompt text - $2: Style (realistic|artistic) — optional, default: realistic

Output: Path to generated image (e.g., /data/ai-stack/output/image_001.png)

Generate Video

./scripts/video.sh "waves crashing on shore" animatediff 4
./scripts/video.sh "city traffic timelapse" ltx2 8

Arguments: - $1: Prompt text - $2: Model (animatediff|ltx2) — optional, default: animatediff - $3: Duration in seconds — optional, default: 4

Output: Path to generated video (e.g., /data/ai-stack/output/video_001.mp4)

Generate Talking Head

./scripts/talking-head.sh "Hello, I'm Agent" gentle input.jpg
./scripts/talking-head.sh "Welcome to the future" neutral photo.png

Arguments: - $1: Speech text - $2: Voice style (gentle|neutral|energetic) — optional, default: gentle - $3: Avatar image path — optional, generates default if not provided

Output: Path to talking head video (e.g., /data/ai-stack/output/talking_001.mp4)

Generate Audio

./scripts/audio.sh "This is a test message" en male
./scripts/audio.sh "Bonjour le monde" fr female

Arguments: - $1: Text to speak - $2: Language code (en|fr|es|etc) — optional, default: en - $3: Voice gender (male|female) — optional, default: male

Output: Path to audio file (e.g., /data/ai-stack/output/audio_001.wav)

Models Available

Image Models

• z-image — 6B params, S3-DiT, photorealistic (downloading, 43% complete)
• Juggernaut XL v9 — SDXL-based, versatile (7.1GB, ready)

Video Models

• AnimateDiff — SD 1.5 motion module (512x512, working ✅)
• LTX-2 — 19B params, high quality (14GB checkpoint ready, Gemma encoder ready)

Talking Head Models

• SadTalker — Audio-driven head animation (working ✅)

Voice Models

• Voxtral — whisper.cpp-based TTS (installed)

Dependencies

All dependencies are pre-installed on GPU server: - ComfyUI with custom nodes (AnimateDiff-Evolved, VideoHelperSuite) - SadTalker with face enhancer - Voxtral with whisper.cpp - FFmpeg for video encoding

Error Handling

Scripts will: - Check SSH connectivity before execution - Validate GPU server is running - Return meaningful error messages - Clean up failed generations automatically

Performance

• Image: ~10-20s for 1024x1024
• Video (AnimateDiff): ~20-30s for 512x512, 16 frames
• Video (LTX-2): ~60-90s for 768x512, 4s @ 24fps
• Talking Head: ~30-40s for 10s video
• Audio: ~2-5s for 30s speech

Future Enhancements

• [ ] Batch generation support
• [ ] Style transfer capabilities
• [ ] Video upscaling (spatial + temporal)
• [ ] Multi-language voice cloning
• [ ] Real-time preview streaming

Status: Active development Maintainer: Agent GPU Server: ${GPU_USER}@${GPU_HOST}