FLUX.1-schnell 4-bit Quantized for MLX

This is a 4-bit quantized version of the FLUX.1-schnell model optimized for use with MLX and flux.swift. The model size has been reduced from ~24GB to 9.2GB while maintaining excellent image generation quality.

Quantized using flux.swift, a Swift implementation of FLUX models for Apple Silicon.

Model Details

Quantization: 4-bit with group size 64
Total Size: 9.2GB
Original Model: black-forest-labs/FLUX.1-schnell
Framework: MLX (Metal Performance Shaders)
Components: Transformer, VAE, CLIP text encoder, T5 text encoder

Usage

This model requires the flux.swift implementation. Please refer to the repository for installation and usage instructions.

Quick Start

# Load and use the quantized model
flux.swift.cli \
  --load-quantized-path /path/to/this/model \
  --prompt "Your prompt here" \
  --output output.png

Recommended Parameters

Steps: 4 (optimized for speed)
Guidance Scale: 3.5
Quantization: Ideal for rapid prototyping

Example with Parameters

flux.swift.cli \
  --load-quantized-path /path/to/this/model \
  --prompt "A majestic mountain landscape at sunset" \
  --steps 4 \
  --guidance 3.5 \
  --width 512 \
  --height 512 \
  --seed 42 \
  --output mountain_sunset.png

License

This model is a quantized version of FLUX.1-schnell, which is licensed under Apache 2.0. The quantized weights maintain the same license as the original model.

Performance

Memory Usage: Reduced from ~24GB to 9.2GB
Quality: Excellent preservation of generation quality
Platform: Optimized for Apple Silicon Macs

Citation

@misc{flux-schnell,
  author = {Black Forest Labs},
  title = {FLUX.1-schnell},
  publisher = {Black Forest Labs},
  year = {2024},
  url = {https://huggingface.co/black-forest-labs/FLUX.1-schnell}
}

@software{flux-swift,
  author = {mzbac},
  title = {flux.swift: Swift implementation of FLUX models},
  url = {https://github.com/mzbac/flux.swift},
  year = {2024}
}