Folders and files
| Name | Name | Last commit date | ||
|---|---|---|---|---|
Repository files navigation
# SD1.5_rknn_3588_euler Stable Diffusion 1.5 inference runtime optimized for Rockchip RK3588 / RK3588S using RKNN and Euler scheduler. This repository contains only code and runtime logic. All heavy assets (RKNN models, Super-Resolution, GFPGAN weights) are hosted separately on Hugging Face. Since it is NOT LCM but Euler- takes much more time to generate -approximately 4-5 minutes , depending on number of steps etc. but also high quality images. Right now only 512x512 generation is possible but can upscale 4X using the built in fucntions. For faster but a little not lifelike images LCM method can be used . url : https://docs.radxa.com/en/rock5/rock5b/app-development/rknn_toolkit_lite2_stable-diffusion --- ## Hardware target - Rockchip RK3588 / RK3588S - Tested on: Rock 5B / Rock 5B+ - Runtime: RKNN Toolkit Lite 2 - Scheduler: Euler - Precision: pre-converted RKNN binaries (FP16 / INT8) --- ## What this repository contains - RKNN-based Stable Diffusion 1.5 runtime - Command-line interface (rkimg.py) - Image-to-image and inpainting support - Real-ESRGAN RKNN upscaler - Lightweight Flask-based WebUI - Hugging Face asset downloader This repository does NOT store large binaries. --- ## Repository structure (after setup) sd15_rknn_euler/ ├── scripts/ │ ├── fetch_assets.py │ ├── txt2img_rknn_sd15_euler.py │ ├── img2img_inpaint_unet.py │ └── upscale_realesrgan_rknn.py ├── models/ (downloaded from HF) │ └── business_rknn/ │ └── unet/ │ └── model.rknn ├── model/ (downloaded from HF) │ └── sr/realesrgan/ │ └── realesrgan_x4plus_tile128_fp16.rknn ├── gfpgan/ │ └── weights/ (downloaded from HF) │ ├── GFPGANv1.4.pth │ ├── detection_Resnet50_Final.pth │ └── parsing_parsenet.pth ├── images/ ├── outputs/ ├── out/ ├── webui/ │ ├── app.py │ ├── templates/index.html │ └── data/ │ └── masks/ ├── rkimg.py ├── requirements.lock.txt └── README.md --- ## Installation Clone the repository: git clone https://github.com/Mojo24x7/SD1.5_rknn_3588_euler cd SD1.5_rknn_3588_euler Create and activate virtual environment: python3 -m venv .venv source .venv/bin/activate pip install -r requirements.lock.txt --- ## Download models (mandatory) All RKNN models and weights are hosted on Hugging Face. You must run this once before using the runtime. python3 scripts/fetch_assets.py Assets are downloaded from: https://huggingface.co/datasets/Mojo24x7/SD1.5_rknn_3588_euler --- ## Usage examples Text to Image (Euler): python3 scripts/txt2img_rknn_sd15_euler.py \ --prompt "a cinematic ultra realistic portrait photo" \ --steps 30 Image to Image / Inpainting: python3 scripts/img2img_inpaint_unet.py \ --init-image images/input.png \ --mask webui/data/masks/mask.png \ --prompt "In masked area ONLY: lush green grass" \ --negative-prompt "snow, white" \ --steps 30 Upscale (Real-ESRGAN RKNN): python3 scripts/upscale_realesrgan_rknn.py \ --rknn model/sr/realesrgan/realesrgan_x4plus_tile128_fp16.rknn \ --in images/input.png \ --out outputs/upscaled.png \ --tile 128 \ --overlap 64 \ --scale 4 IMPORTANT: Upscaling works ONLY at x4. Other scales are not supported. --- ## WebUI python3 webui/app.py Open in browser: http://<rock-ip>:7860 --- ## Model details - Base architecture: Stable Diffusion 1.5 - Main checkpoint style: Realistic Vision (SD1.5) - Scheduler: Euler - UNet: single-batch RKNN - Target hardware: RK3588 NPU --- ## Credits and acknowledgements This project builds upon the work of: - Stable Diffusion 1.5 by CompVis and Stability AI - Realistic Vision (SD1.5 checkpoint) - Real-ESRGAN - GFPGAN - RKNN Toolkit and RKNN Toolkit Lite 2 by Rockchip - Hugging Face Hub All model weights remain under their original upstream licenses. This repository provides pre-converted RKNN binaries for convenience only. --- ## Disclaimer This project is intended for research, experimentation, and personal use. The author is not affiliated with Stability AI, CompVis, Rockchip, Real-ESRGAN, GFPGAN, or Hugging Face.