Stable Diffusion 1.5 Euler

This document explains how to run the Stable Diffusion 1.5 Euler model for text-to-image inference on a host device equipped with Radxa AX-M1.

tip

This example is based on the solution provided by community user SHANKAR_THAKUR, thanks for their contribution.

Environment Preparation

Before starting, please ensure:

• Radxa AX-M1 is properly installed and connected to the host
• The host (supporting Rock 5B/5B+ or Raspberry Pi 5) has the Radxa official image or other compatible operating system installed
• Network connection is normal and GitHub and HuggingFace are accessible

tip

Raspberry Pi 5 can also be used as the host, as long as AX-M1 is connected and the AXCLRT runtime can run on that operating system.

caution

Please ensure the AX-M1 hardware connection is stable. If inference fails, first check the hardware connection status and device temperature.

Clone the Project Repository

Host

git clone https://github.com/Mojo24x7/SD1.5_AXM1-AX8850_Euler && cd SD1.5_AXM1-AX8850_Euler

Create Virtual Environment

Host

python3 -m venv .venv

Activate Virtual Environment and Upgrade pip

Host

source .venv/bin/activate && pip install -U pip

Install Project Dependencies

Host

pip install -r requirements.txt

Download and Install pyaxengine

pyaxengine is the official Python API library provided by AXERA for running inference on AX-M1.

Host

wget https://github.com/AXERA-TECH/pyaxengine/releases/download/0.1.3.rc2/axengine-0.1.3-py3-none-any.whl
pip install axengine-0.1.3-py3-none-any.whl

Download Pre-trained Model

First, install git-lfs:

Host

sudo apt update && sudo apt install git-lfs
git lfs install

Clone the model repository from HuggingFace:

Host

git clone https://huggingface.co/Mojo24x7/sd15-axm1-euler512-axmodels hf_axmodels

Copy model files to the specified directory:

Host

mkdir -p axmodels
cp -v hf_axmodels/*.axmodel axmodels/

tip

The model files are large and download may take a long time, please be patient.

Run Text-to-Image Inference

Host

python3 scripts/txt2img_axengine_euler.py \
  --weights_dir ./axmodels \
  --tokenizer_dir ./support/tokenizer \
  --scheduler_dir ./support/scheduler \
  --vae_config ./support/vae/config.json \
  --prompt "a cinematic ultra realistic portrait photo" \
  --steps 30 \
  --out ./out/txt2img_euler_512.png

Parameter Description

Parameter	Description
--weights_dir	Model weights directory
--tokenizer_dir	Tokenizer directory
--scheduler_dir	Scheduler configuration directory
--vae_config	VAE config file path
--prompt	Input text prompt
--steps	Inference steps
--out	Output image path

Expected Output

After successful inference, the output image will be saved at ./out/txt2img_euler_512.png.

Run Web UI

The project also provides a Web UI for interactive image generation:

Host

python run_ui.py --base_dir ./ --out_root ./out --venv_py ~/SD1.5_AXM1-AX8850_Euler/.venv/bin/python --token_max_len 100

tip

Please ensure the virtual environment path matches the actual path. If the path is different, modify the --venv_py parameter to point to the correct Python interpreter path.

FAQ

How to Troubleshoot Inference Failure?

1. Check hardware connection: Ensure the AX-M1 module is properly connected to the M.2 interface
2. Check device temperature: High temperature may cause performance degradation or inference failure
3. Check virtual environment: Ensure you are using the Python interpreter from the correct virtual environment
4. Check log output: Detailed logs usually indicate the specific error cause

What if Model Download Fails?

• Check network connection to ensure HuggingFace is accessible
• Try using a proxy or mirror site
• You can manually download model files and copy them to the corresponding directory

Slow Inference Speed?

• Check AX-M1 heat dissipation, add a heatsink if necessary
• Confirm the NPU is working properly (can be checked with axcl-smi command)