Setup

Setup

Two ways to get the backend running.
Recommended – Automated script (GPU ⬅︎→ CPU auto-detect).
Alternative – Manual install for locked-down networks or custom tuning.

---

1) Automated install (recommended)

1.1 Create & activate the conda env

conda create -n easy_art python=3.10 -y
conda activate easy_art

1.2 Run the script

Platform	Command	What the script does
Linux / macOS / WSL	bash setup_unix.sh	• Detect GPU → install cu121/cu118 wheels, else CPU wheels \n• Clone Automatic1111 repo \n• Install pinned deps \n• Download default SD-1.5 (v1-5-pruned.safetensors)
Windows (PowerShell)	Set-ExecutionPolicy -Scope Process RemoteSigned; .\setup_windows.ps1	Performs the same steps, CUDA-aware or CPU-only

Scripts live in the repo root (or scripts/ if you moved them). Both scripts print ✅ Done and a “Next steps” block.

1.3 Launch & smoke-test

# Start local A1111 (GPU or CPU-only)
python program_main/serve_local_sd.py --port 7860

# Unified backend smoke test
python program_main/backend_main.py
# Enter a description; the generated image will open automatically.

---

2) Manual install (full control)

Follow this if scripts are blocked or you need different versions.

2.1 Base environment

conda create -n easy_art python=3.10 -y
conda activate easy_art

2.2 PyTorch — choose one

Scenario	Command
CUDA 12.x GPU	pip install --index-url https://download.pytorch.org/whl/cu121 torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2
CUDA 11.x GPU	pip install --index-url https://download.pytorch.org/whl/cu118 torch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2
CPU-only	pip install --index-url https://download.pytorch.org/whl/cpu torch torchvision torchaudio

2.3 Pinned runtime dependencies

pip install numpy==1.26.4 scipy==1.11.4 pandas==1.5.3 \
            matplotlib==3.7.2 kiwisolver==1.4.5 orjson==3.9.10 \
            fastapi gradio==3.41.2 uvicorn pillow==9.5.0 tqdm \
            transformers==4.39.3 diffusers==0.27.2 accelerate==0.27.2 \
            pyyaml addict safetensors==0.4.2 requests psutil python-dotenv
# Optional GPU speed-up:
pip install xformers   # ignore if it fails on CPU

2.4 Clone Automatic1111 & place the model

git clone --depth 1 https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
mkdir -p stable-diffusion-webui/models/Stable-diffusion
curl -L -o stable-diffusion-webui/models/Stable-diffusion/v1-5-pruned.safetensors \
  https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors

2.5 Run & verify

python program_main/serve_local_sd.py --port 7860
python program_main/backend_main.py

If you see “WebUI is ready at http://127.0.0.1:7860” and an image opens, the install succeeded.

---

Troubleshooting

Symptom	Fix
404 /sdapi/v1/txt2img right after start	Normal — backend auto-retries; wait ~2s.
ModuleNotFoundError: numpy._core...	pip install --force-reinstall numpy
cannot import name '_c_internal_utils' from matplotlib	pip install --force-reinstall matplotlib kiwisolver
Torch “not compiled with CUDA” on a GPU box	Ensure the wheel matches CUDA (cu121/cu118); uninstall & reinstall torch/vision/audio.
Port 7860 stuck	lsof -i:7860 (Linux) / `netstat -ano ^

---

Environment variables (.env)

LOCAL_SD_HOST=http://127.0.0.1:7860
MODELSLAB_API_KEY=xxxxxxxxxxxxxxxx   # only needed for cloud SD

---

Minimal usage (backend)

from backend_main import start_local_server, generate_image_from_prompt, stop_local_server

start_local_server()  # idempotent, loads default SD-1.5
urls = generate_image_from_prompt("cat astronaut, anime style", model="local")
print(urls[0])
stop_local_server()