v0.2

v0.2:

• New UI: Thanks to flutter-shadcn-ui
• New Bindings: Added support for Clip_L, Clip_G, T5XXL, Embeddings, VAE, and VAE Tiling (Flux/SD3.5 model should work in theory)
• Feature: Added ControlNet support with image reference and Canny preprocessing
• Feature: Added Image-to-Image mode
• Feature: Added 4x Upscaler via RealESRGAN (although it sometimes produces oversaturated results; may fix or consider another model)
• Feature: Added Photomaker V1 (V2 is unlikely to be supported due to insightface dependency)
• Feature: Added Scribble-to-Image via control_v11p_sd15_scribble

Upcoming:

• Update libstable-diffusion.so to the latest stable-diffusion.cpp commit
• Add inpainting support with image segmentation for mask creation
• Add outpainting
• Sync with the latest ggml commit.
• Add experimental backend for testing: Vulkan, OpenCL, OpenBLAS, KleidiAI
• Add face restoration and a better upscaler

v0.1

• Generated images are now saved to the gallery
• LORA now works correctly. Please refer to this guide for instructions on how to use the LORA tag in the prompt field

EDIT :

Add a precompiled version for testing and benchmarking purposes that includes OpenBLAS v0.3.29 as a backend. In my testing, I didn’t notice any improvement, but it may be useful for some scenarios. If so, an option to toggle BLAS on/off will be added to the UI.

Local Diffusion v0.1 preview

A preview version for benchmarking purpose, tested with SD 1.5/2.1/SDXL/SDXL Turbo

• LORA implementation is currently broken
• TAESD is correctly implemented, I strongly encourage you to use it to speed up decoding process