EcoVideo Studio 🌿

Energy-Efficient Video Generation on Apple Silicon

EcoVideo Studio generates videos from images using Stable Video Diffusion, optimized for Apple Silicon M3 Pro with a focus on sustainability and Green AI principles.

Features

• Video Generation: Transform static images into 14-frame videos using SVD
• Three Quality Presets: Eco (fast/efficient), Balanced, Quality (best output)
• Sustainability Metrics: Real-time energy, carbon footprint, and cloud savings tracking
• ACSA Algorithm: Adaptive Compression Selection using CLIP features + hardware profiling
• Local Processing: 100% on-device, no cloud dependency
• History & Analytics: Track all generations with quality and efficiency metrics

Installation

cd ecovideo-studio
python3.11 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python setup_app.py  # Downloads model & initializes database

Usage

streamlit run frontend/app.py

Navigate to http://localhost:8501 in your browser.

Architecture

ecovideo-studio/
├── frontend/          # Streamlit UI
├── backend/           # Inference orchestration
├── ml_components/     # SVD model, compression, quality
├── intelligence/      # ACSA adaptive preset selection
├── sustainability/    # Energy, carbon, impact metrics
├── storage/           # SQLite persistence
├── utils/             # Shared utilities
├── data/              # Models, benchmarks, configs
├── tests/             # Test suite
└── notebooks/         # Research notebooks

Sustainability Methodology

Energy is estimated using M3 Pro power profiles (22W GPU load, 5W idle) multiplied by generation time. Carbon emissions use regional grid intensity factors (default 475 gCO2/kWh). Savings are computed against a cloud baseline of ~150 Wh per video.

ACSA Algorithm

The Adaptive Compression Selection Algorithm uses a small MLP classifier that takes CLIP image embeddings (768-dim) concatenated with hardware features to predict the optimal preset for each image. This achieves ~85% accuracy on validation data.

Performance (M3 Pro)

Preset	Time	Peak VRAM	Energy
Eco	90-150s	~8GB	~8 Wh
Balanced	180-240s	~10GB	~11 Wh
Quality	300-420s	~12GB	~15 Wh

Limitations

• MPS backend does not support INT8/INT4 quantization (bitsandbytes)
• CPU offloading designed for CUDA is not available
• Power measurement is estimated, not direct hardware reading
• Neural Engine not accessible through PyTorch

References

• Stable Video Diffusion (Blattmann et al., 2023)
• Green AI (Schwartz et al., 2020)
• CLIP (Radford et al., 2021)

License

MIT

BTP