How to run ComfyUI on a cloud GPU: setup, models, cost per image

Why rent instead of buying a 4090

ComfyUI is bursty by nature: an evening of prompting, a batch of a thousand images, then nothing for a week. A rented card bills only the seconds the queue is running — $0.262/hr for an RTX 4090 on-demand, $0.131/hr interruptible — so a heavy month rarely reaches a fraction of the card's retail price, and you can jump to a 32 GB or 96 GB card the day a workflow needs it. No driver upgrades, no 450 W heater under the desk.

Pick the card: 24, 32 or 48 GB

Rule of thumb: Flux dev in FP8 wants ~17 GB, SDXL ~8 GB, current video models 24 GB with offloading and 48 GB without. The cheapest-GPU guide ranks the cards by images per dollar.

Deploy the template (30 seconds)

1. Open the console, filter on the card, pick a machine.
2. Choose ComfyUI in the template picker. It ships ComfyUI-Manager, exposes the UI on a TLS-terminated port and mounts /workspace/ComfyUI/models.
3. Set the disk (60 GB is plenty for the OS and cache) and attach a volume for models — next section.
4. Deploy. The instance is running in about 30 seconds; click the port link and the node graph is there.

powergpu launch --gpu rtx-4090 --template comfyui \
    --disk 60 --volume sdmodels:/workspace/ComfyUI/models
# ✓ instance i-3c91ab04 running (28.6s)
# ✓ https://i-3c91ab04.powergpu.io:8188  (ComfyUI)

Models on a volume, once

Checkpoints are 2–12 GB each and downloading them every session is the most common way to waste GPU-hours. Create a volume the first time, mount it on the models folder, fill it once — Hugging Face and Civitai downloads run at multi-Gbps from the datacenter — and every future instance in the region starts with the library present. A 120 GB library costs $9.60/month. Put custom_nodes on the same volume so installed nodes persist too.

Headless: workflows as an API

Enable dev mode in ComfyUI settings, then "Save (API format)" on any workflow. The JSON you get is a request body: POST it to /prompt, poll /history/<id>, fetch the images from /view. That loop is how batch pipelines run a thousand prompts overnight on interruptible capacity, and how a serverless endpoint serves the same workflow behind autoscaling workers.

curl -X POST https://i-3c91ab04.powergpu.io:8188/prompt \
  -H 'Content-Type: application/json' \
  -d @workflow_api.json
# {"prompt_id": "a1c2…", "number": 12}
curl https://i-3c91ab04.powergpu.io:8188/history/a1c2…

What an image costs, honestly

Roughly 6,246 Flux images per dollar on-demand, 12,491 interruptible. Speeds are typical for a tuned 4090; your workflow, resolution and step count move them.

Five traps that waste GPU-hours

• Downloading models every session — the volume above fixes it permanently.
• Leaving the instance running overnight — stop it; per-second billing means an idle GPU is pure waste. The disk survives stop/start.
• Running batches on-demand — queues are interruptible by nature: each prompt is a restartable item. Pay half.
• Picking a card by hourly price alone — a 5090 finishes Flux batches faster than its price gap over a 4090; images per dollar is the metric.
• Forgetting the outputs — write results to the volume or download them before destroy; the instance disk dies with the instance.