Skip to content

Latest commit

 

History

History
79 lines (59 loc) · 2.31 KB

README.md

File metadata and controls

79 lines (59 loc) · 2.31 KB

Local Diffusion

Flutter GUI wrapper for stable-diffusion.cpp - Run Stable Diffusion locally on Android

Features

  • Pure Flutter implementation using FFI bindings to stable-diffusion.cpp

  • Android support (more platforms coming soon)

  • Super lightweight with minimal dependencies

  • Supports multiple model architectures:

    • SD1.x
    • SD2.x
    • SDXL
    • SD3/SD3.5
    • Flux/Flux-schnell
    • SD-Turbo and SDXL-Turbo

  • Performance optimizations:

    • OpenBLAS acceleration
    • TAESD for faster decoding (reduces latent decode time from ~48s to ~2s)

  • Advanced features:

    • LoRA support
    • On-the-fly model quantization (q8_0, q6_k, q5_0, q5_1, q5_1k ,q4_0, q4_1, q4_k, q3_k, q2_k)
    • Negative prompts
    • Token weighting

  • Sampling methods:

    • Euler A
    • Euler
    • Heun
    • DPM2
    • DPM++ 2M
    • DPM++ 2M v2
    • DPM++ 2S a
    • LCM

Flash Attention Support

When initializing a model, you can choose to load it with or without Flash Attention:

RAM Usage Benchmarking for SD 1.5

Precision Level Without Flash Attention (MB) With Flash Attention (MB)
FP16 2035 1969
Q8_0 1683 1618
Q6_K 1637 1572
Q5_0 1543 1477
Q4_0 1496 1431
Q3_K 1498 1433
Q2_K 1462 1397

These measurements were taken using SD 1.5 model, benchmark on larger models is coming soon. Other Quantization method can be implemented if really needed.

Roadmap

  • Vulkan backend for GPU acceleration
  • ControlNet support
  • img2img generation
  • inpainting/outpainting
  • RealESRGAN upscaling
  • iOS support

Building

Run in release mode

flutter run --release

Release APK

flutter build apk --release

The APK will be available at build/app/outputs/flutter-apk/app-release.apk

Credits

Based on stable-diffusion.cpp