🎨 Design & Creative — go live in one command

This stack is a creative studio:

                         ┌──────────────── Caddy (auto-HTTPS) ────────────────┐
   designers ──────────▶ │  https://studio.yourdomain.com                     │
   (browser / API)                            │                               │
                         └───────────────────┼───────────────────────────────┘
                                             ▼
                                    ComfyUI (node-graph generation)
                                     │            ▲
                        /prompt API  │            │  rendered assets
                                     ▼            │
                              NVIDIA GPU  ·  ./storage (models, outputs)

You get a programmable, API-driven image/video generation pipeline — behind real HTTPS on your domain, running on your own hardware.

Replaces: a production designer + stock-image spend + per-render SaaS credits
Time to live: ~30–60 minutes (longer if you download large models)
You need: a server with an NVIDIA GPU, a public IP, a domain, and Docker

A note on the image. ComfyUI ships no official Docker image. This stack uses yanwk/comfyui-boot — a well-maintained community image (~1.6k★, releases through 2025, active commits). A solid alternative is ghcr.io/ai-dock/comfyui; swap COMFYUI_IMAGE in .env if you prefer it. Both were verified before shipping; neither is published by the ComfyUI project itself.

Step 1 — Get a GPU server

ComfyUI is a diffusion-model engine: a GPU is strongly recommended. Any Ubuntu 22.04+ host with an NVIDIA GPU (8 GB+ VRAM) works — a GPU VPS from Lambda, RunPod, Vast.ai, Hetzner GEX, or a cloud g5/A10/RTX instance. SSH in, then install Docker:

curl -fsSL https://get.docker.com | sh

Open ports 80 and 443 in your provider's firewall.

CPU-only caveat. You can run without a GPU by setting COMFYUI_IMAGE=yanwk/comfyui-boot:cpu and commenting out the deploy: block in docker-compose.yml — but renders will be extremely slow (minutes to tens of minutes per image). Fine for kicking the tires; not for real work.

Step 2 — Install the NVIDIA container toolkit

Docker needs the NVIDIA runtime to hand the GPU to the container:

curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
  | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
  | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' \
  | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit
sudo nvidia-ctk runtime configure --runtime=docker && sudo systemctl restart docker

Verify the GPU is visible to Docker:

docker run --rm --gpus all nvidia/cuda:12.6.0-base-ubuntu22.04 nvidia-smi

You should see your GPU listed. (Skip this whole step for a CPU-only trial.)

Step 3 — Point your domain at it

Create a DNS A record for studio.yourdomain.com → your server's IP. Verify it resolves before continuing (HTTPS issuance depends on it):

dig +short studio.yourdomain.com   # should print your server IP

Step 4 — Get the stack

git clone https://github.com/arunrajiah/ai-workforce.git
cd ai-workforce/departments/design/deploy
cp .env.example .env

Step 5 — Fill in `.env`

In .env, set at minimum:

DOMAIN → studio.yourdomain.com
ACME_EMAIL → your email (for Let's Encrypt)
COMFYUI_IMAGE → leave the default GPU image, or pick the CPU tag for a trial

Step 6 — Launch

docker compose up -d

Caddy will fetch a TLS certificate automatically. The ComfyUI image is large (several GB) and pulls on first run, so give it a few minutes. Then open https://studio.yourdomain.com — the node-graph editor loads in your browser.

Watch it come up: docker compose logs -f comfyui caddy

Step 7 — Add models

ComfyUI needs at least one checkpoint to generate. The ./storage folder is mounted into the container, so drop model files on the host and they appear inside:

# from the deploy/ folder — put a checkpoint where ComfyUI expects it
mkdir -p storage/ComfyUI/models/checkpoints
cd storage/ComfyUI/models/checkpoints
# e.g. an SDXL / SD1.5 checkpoint from Hugging Face or Civitai
wget <URL-to-a-.safetensors-checkpoint>

Reload the ComfyUI tab — the checkpoint appears in the Load Checkpoint node. Drop LoRAs, VAEs, and ControlNets under the matching storage/ComfyUI/models/* folders the same way.

Step 8 — Render your first asset

In the browser: load the default workflow, pick your checkpoint in Load Checkpoint, type a prompt, and hit Queue Prompt. Rendered images land in storage/ComfyUI/output/ on the host as well as in the UI.

You're live. 🎉

Optional — AI prompt-expansion (brief → prompt → render)

ComfyUI exposes a /prompt HTTP API, so you can drive it from an automation tool. A simple, provider-neutral pattern:

From a small script or automation step, take a short creative brief.
Build a flow: a short brief → HTTP Request to an OpenAI-compatible endpoint (set LLM_BASE_URL / LLM_MODEL / LLM_API_KEY in that flow) that expands the brief into a rich, detailed prompt.
Inject the expanded text into a ComfyUI workflow JSON and POST it to http://comfyui:8188/prompt (inside the Docker network) or https://studio.yourdomain.com/prompt from outside.
Poll /history/<prompt_id> for the finished image.

Any OpenAI-compatible model works (a local Ollama server, a hosted gateway, etc.) — the studio stays vendor-neutral.

Operate it

Task	Command
View logs	`docker compose logs -f comfyui caddy`
Confirm the GPU is in use	`docker compose exec comfyui nvidia-smi`
Update ComfyUI	`docker compose pull && docker compose up -d`
Back up models & outputs	`tar czf studio-backup.tgz storage/`
Stop everything	`docker compose down` (your `storage/` folder persists)

Troubleshooting

No HTTPS certificate — DNS isn't pointing at the server yet, or ports 80/443 are blocked. Check docker compose logs caddy.
CUDA / GPU not found — Step 2 didn't finish. Re-run the nvidia-smi container test; if it fails, reinstall the toolkit and systemctl restart docker.
502 / blank page — the image is still pulling or ComfyUI is booting. Watch docker compose logs -f comfyui; first start can take several minutes.
"No checkpoints" in the UI — you haven't added a model yet. See Step 7; files go under storage/ComfyUI/models/checkpoints/.
Out-of-memory during a render — the checkpoint is too large for your VRAM. Use a smaller model (SD1.5), lower the resolution/batch size, or add --lowvram support via the image's env options.

Swap a component

Prefer a different generation engine or a lighter footprint? Every piece is replaceable — see the Design & Creative directory for alternatives (AUTOMATIC1111, InvokeAI, HivisionIDPhotos, StableStudio).