Z-Image Turbo and FLUX.2 Klein 4B
Huge thanks to my friend happyyzy for the deep optimization work on the Hexagon path in stable-diffusion.cpp. It is genuinely impressive — and it is what makes these two DiT models run on Snapdragon 8 Elite and newer.
- Z-Image Turbo, 8 steps. FLUX.2 Klein 4B, 4 steps.
- Width and height are set independently, anywhere from 512 to 2048. These models generate the requested shape directly, so a non-square image is no longer cropped out of a padded square canvas.
- The models only appear on supported devices (SM8750 and newer). Weights download from their original repositories; the engine ships inside the app.
SDXL on CPU/GPU, and a VAE fix — thanks to @Mr-J-369 (#306)
- SDXL MNN packages can now be imported and run without an NPU, with tiled VAE decoding and low-RAM staged loading.
- Long prompts are split into chunks for SDXL MNN and SD1.5 MNN.
- Fixed: VAEs with floating point inputs are now supported. Generation could previously finish every denoising step and still return nothing.
Draw on your source image — thanks to @0xf1 (#302)
A new drawing canvas for Img2Img and Inpaint. Paint, erase, or clone directly on the source image before generating — handy for masking an area for Inpaint, or sketching rough color guidance for Img2Img, without leaving the app.
Variants :
- LocalDream_xxx_with_filter: Block NSFW Results. Same as Google Play version.
- LocalDream_xxx: No filter.
Co-Authored-By: Claude Opus 5