Source: examples/stable_diffusion

Stable Diffusion Web UI on SkyPilot

Run the InvokeAI Stable Diffusion web UI on a cloud GPU with SkyPilot, using the official ghcr.io/invoke-ai/invokeai:v6.14.2-cuda image (pinned; newer tags are listed in the InvokeAI releases).

Steps

1. Install SkyPilot by following the instructions.

2. Run git clone https://github.com/skypilot-org/skypilot.git && cd skypilot/examples/stable_diffusion

3. Run sky launch -c stable-diffusion stable_diffusion_docker.yaml

4. Run ssh -L 9090:localhost:9090 stable-diffusion

5. Open http://localhost:9090 in your browser.

6. On first launch, open Model Manager in the UI and install a starter model (for example, a Stable Diffusion 1.5 or SDXL checkpoint). Models and generated images are stored under ~/invokeai on the cluster, so they survive container restarts.

7. Type in a prompt and click Invoke to generate images.

8. Once you are done, run sky stop stable-diffusion to stop the VM, or sky down stable-diffusion to delete it.

If a model you want to install is gated on Hugging Face, pass a token with sky launch --env HUGGING_FACE_HUB_TOKEN=<token> ....

Tips

• Prompts like A photo of an astronaut riding a horse, cinematic lighting, highly detailed work well; use the negative prompt field to exclude things you don’t want.

• The Canvas tab supports inpainting and outpainting; Workflows exposes the node editor for more advanced pipelines.

Running locally

docker-compose.yml starts the same container on a local machine with an NVIDIA GPU: docker compose up.

Included files

docker-compose.yml

version: '3.8'

services:
  invokeai:
    image: ghcr.io/invoke-ai/invokeai:v6.14.2-cuda
    container_name: invokeai
    restart: on-failure
    ports:
      - "9090:9090"
    volumes:
      - ~/invokeai:/invokeai
    environment:
      - HUGGING_FACE_HUB_TOKEN=${HUGGING_FACE_HUB_TOKEN:-}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

stable_diffusion_docker.yaml

# Runs the InvokeAI Stable Diffusion web UI (https://github.com/invoke-ai/InvokeAI)
# on a single GPU using the official ghcr.io/invoke-ai/invokeai:v6.14.2-cuda image.
#
# Usage:
#   sky launch -c stable-diffusion examples/stable_diffusion/stable_diffusion_docker.yaml
#   ssh -L 9090:localhost:9090 stable-diffusion
#   # then open http://localhost:9090
envs:
  HUGGING_FACE_HUB_TOKEN: ""  # optional, for gated Hugging Face models

resources:
  accelerators: L4:1
  ports: 9090

setup: |
  set -e
  docker pull ghcr.io/invoke-ai/invokeai:v6.14.2-cuda

run: |
  set -e
  # Models and generated images are kept under ~/invokeai on the host.
  mkdir -p ~/invokeai
  docker rm -f invokeai 2>/dev/null || true
  docker run -d --name invokeai --restart on-failure \
    --gpus all \
    -p 9090:9090 \
    -v ~/invokeai:/invokeai \
    -e HUGGING_FACE_HUB_TOKEN=${HUGGING_FACE_HUB_TOKEN:-} \
    ghcr.io/invoke-ai/invokeai:v6.14.2-cuda
  docker logs -f invokeai