stable-diffusion

ThinkExponential AILab — Stable Diffusion WebUI

Runs AUTOMATIC1111 Stable Diffusion WebUI in a Docker container for local AI image generation. Compatible with Windows + WSL2 on both AMD and NVIDIA GPUs.

Supported models

This WebUI is powered by AUTOMATIC1111 Stable Diffusion WebUI. For the full list of supported model formats and architectures, refer to the AUTOMATIC1111 repository.

Place .safetensors or .ckpt model files in the sd_models Docker volume, or download them through the WebUI.

Pre-installed models (CPU / OpenVINO image)

The CPU Docker image ships with two SD 1.5-based checkpoints so you can start generating immediately:

Model	Style	Size
deliberate_v2	Photorealistic	~2 GB — high quality portraits and scenes
dreamshaper_8	Versatile	~2 GB — good at both artistic and photorealistic

Prerequisites

All platforms

• Docker with Compose v2 (docker compose)
• At least 8 GB RAM (16 GB+ recommended)
• At least 20 GB free disk space

NVIDIA (WSL2 on Windows)

1. Update WSL kernel: wsl --update
2. Install NVIDIA driver ≥525 on Windows from nvidia.com/drivers
3. Inside WSL2, run the included setup script to install the NVIDIA Container Toolkit and configure Docker:

   chmod +x scripts/setup-wsl-nvidia.sh
   ./scripts/setup-wsl-nvidia.sh

   (See ollama/README.md for the manual installation steps.)
4. Minimum 4 GB VRAM (6 GB+ recommended)

AMD (WSL2 on Windows)

1. Update WSL kernel to 5.15+: wsl --update
2. Install ROCm 6.x inside WSL2 using the included setup script:

   chmod +x scripts/setup-wsl-amd.sh
   ./scripts/setup-wsl-amd.sh

   Or follow the ROCm installation guide manually.
3. Add your user to render and video groups (the script does this automatically):

   sudo usermod -aG render,video $USER

4. Supported GPUs: RX 6000/7000 series (gfx1030+), Instinct MI series — check the AMD ROCm compatibility matrix

CPU only — Intel Silicon (OpenVINO)

No additional prerequisites beyond Docker. The compose file sets the OpenVINO-compatible flags automatically:

• --skip-torch-cuda-test — bypasses the GPU detection that would otherwise prevent startup on CPU-only hosts
• --precision full --no-half — disables half-precision arithmetic, required for CPU inference
• PYTORCH_TRACING_MODE=TORCHFX — enables OpenVINO acceleration via torch.compile
• --enable-insecure-extension-access — allows WebUI extensions broader access to internal APIs; this is enabled for compatibility with common extensions and assumes you are running on a trusted local network and not exposing the WebUI directly to the internet

Note: Image generation on CPU is very slow (minutes per image). This mode is intended for testing the interface or on Intel CPU-only machines. See the Intel Silicon installation guide for more details.

Quick start

NVIDIA GPU

docker compose -f docker-compose.nvidia.yml up -d

AMD GPU

docker compose -f docker-compose.amd.yml up -d

Intel CPU only (OpenVINO, slow)

docker compose -f docker-compose.cpu.yml up -d

Open http://localhost:7860 in your browser.

Adding models

Download model files from Hugging Face or CivitAI and copy them into the Docker volume:

# Find the volume mount path
docker volume inspect thinkexponential-ailab-stable-diffusion_sd_models

# Or copy a model file directly into the container
docker cp my-model.safetensors stable-diffusion-webui:/app/stable-diffusion-webui/models/Stable-diffusion/

After copying, click Refresh in the WebUI model selector.

Adding models to the CPU container

The sd_models Docker volume mounts to models/Stable-diffusion/ inside the container. You can add models without rebuilding:

# Download from Hugging Face (example: Realistic Vision v5.1)
docker exec stable-diffusion-webui curl -L -o models/Stable-diffusion/realisticVision_v51.safetensors \
  "https://huggingface.co/SG161222/Realistic_Vision_V5.1_noVAE/resolve/main/Realistic_Vision_V5.1_fp16-no-ema.safetensors"

# Or copy a local file into the running container
docker cp my-model.safetensors stable-diffusion-webui:/app/stable-diffusion-webui/models/Stable-diffusion/

After adding a model, click 🔄 Refresh next to the checkpoint dropdown in the WebUI (no restart needed).

Tips for better results on CPU (OpenVINO)

1. Select "Accelerate with OpenVINO" from the Script dropdown at the bottom of the txt2img / img2img tab
2. Set OpenVINO Device to CPU in the script options
3. Use a better model — select deliberate_v2 or dreamshaper_8 via the checkpoint dropdown (top-left)
4. Recommended settings:
   • Steps: 20–30 (more steps = better quality but slower)
   • Sampler: DPM++ 2M Karras or Euler a
   • CFG Scale: 7
   • Resolution: 512×512 (SD 1.5 models) or 768×768 max
5. First inference is slow (~2-5 min) because OpenVINO compiles the model. Subsequent runs with the same resolution/settings are much faster due to caching
6. Use descriptive prompts:

   photo of a golden retriever in a sunlit meadow, shallow depth of field,
   professional photography, 8k, highly detailed
   

7. Use negative prompts to avoid common artifacts:

   blurry, low quality, deformed, ugly, bad anatomy, watermark, text
   

API access

The WebUI exposes a REST API at http://localhost:7860/sdapi/v1/:

curl -X POST http://localhost:7860/sdapi/v1/txt2img \
  -H "Content-Type: application/json" \
  -d '{"prompt": "a beautiful sunset over mountains", "steps": 20}'

Stopping the stack

docker compose -f docker-compose.nvidia.yml down     # stop and remove containers
docker compose -f docker-compose.nvidia.yml down -v  # also remove volumes (model data)