Skip to content
DreamSprite Logo DreamSprite

ComfyUI Setup

Midjourney is awesome for quick cloud ideation, but when you want full control, zero queues, no monthly fees, and complex workflows (like locking poses with DWPose + OpenPoseXL2), ComfyUI is where the real magic happens. It’s fully node-based, runs completely local on your GPU (or in the cloud), and scales beautifully whether you’re on a high-end card or renting power remotely.

As of March 2026 (v0.17.2), the portable Windows NVIDIA build is still the fastest, most beginner-friendly path - no system Python conflicts, everything bundled.

Local vs Cloud: What’s the Difference?

You can run ComfyUI locally (on your own PC) or in the cloud (renting remote GPUs via services). Here’s a quick breakdown:

AspectLocal (Your PC)Cloud (Rented GPU)
CostOne-time (hardware) + electricity; $0 ongoing for gensPay-per-hour/use (e.g., $0.35-$2+/hr); no upfront hardware
PerformanceDepends on your GPU; consistent speed once set upAccess to faster/newer GPUs (RTX 5090, H100, etc.); often 2-5x quicker
Setup5-30 min install; full control2-10 min (some one-click); but manage instance/shutdown
Privacy100% offline; data never leaves your machineImages/prompts sent to provider; choose privacy-focused ones
Unlimited GensYes, as long as PC is onYes, but pay for runtime; stop when idle
Power/HeatYour PC fans roar, higher electric billNo local heat/power draw
Best ForFrequent use, privacy, complex custom nodes, offlineNo strong GPU, testing, bursts of high-res/Flux/video, portability
DownsidesUpfront cost, VRAM limits, maintenanceInternet required, potential queues, recurring cost

Quick rule of thumb: If you generate a lot (daily/weekly) and have a decent GPU - go local (cheaper long-term). If your PC is weak/old/no GPU, or you want occasional high-power bursts - cloud is perfect (start cheap, scale up).

Running ComfyUI in the Cloud (If No Powerful PC)

No high-end GPU? No problem - rent one hourly (often cheaper than buying). Popular 2026 options for ComfyUI:

  • RunComfy (easiest/no-setup): https://www.runcomfy.com - Native ComfyUI in browser, fast GPUs (16-141GB VRAM), one-click workflows, API scaling. Great for beginners; pay-per-use or subscription. Zero install - just sign up and launch.
  • RunPod (flexible/value king): https://www.runpod.io - One-click ComfyUI templates (RTX 4090/A100/H100), $0.35-$1+/hr depending on GPU. Spin up a “Pod”, access via browser or SSH, shut down when done. Community templates make it fast.
  • Vast.ai (cheapest spot pricing): https://cloud.vast.ai - Peer-to-peer rentals (RTX 3090/4090 often $0.30-$0.60/hr). Search for “ComfyUI” Docker images, deploy, connect via browser/SSH. More manual but lowest cost.
  • ThinkDiffusion or RunDiffusion: https://www.thinkdiffusion.com / https://rundiffusion.com - Managed ComfyUI interfaces, preloaded models/nodes, pay-per-hour or credits. Good for zero-config.
  • Other solid picks: Lightning AI (free tier possible for light use), Spheron, Hyperstack (H100/A100), or Lambda Labs.

How to start (general steps for most):

  1. Sign up on the platform (credit card usually required).
  2. Choose a ComfyUI template/one-click deploy (or upload your workflow JSON).
  3. Select GPU (start with RTX 4090/A40 for balance of speed/cost).
  4. Launch instance - wait 2-5 min.
  5. Access via web UI (browser link provided) - same as local!
  6. Generate - download images - shut down instance to stop billing.

Tip: Use spot/interruptible instances for big savings (cheaper but can pause). Always shut down when idle - costs add up fast!

No need to install anything else first (like Automatic1111 or Stable Diffusion webui) - ComfyUI is completely standalone.

(We’ll explain what models are, how to choose them, and our favorites in the next section: Models - jump there when you’re ready!)

Local Installation (Run ComfyUI on Your Own PC)

Everything below is for installing and running ComfyUI directly on your machine. Skip this if you’re using a cloud provider above.

  1. Download the Latest Portable Head to: https://github.com/Comfy-Org/ComfyUI/releases/latest Download: ComfyUI_windows_portable_nvidia_cu130.7z (If you have an older card, grab the cu128 or cu126 variant instead.) Extracted size is ~10-15 GB - put it on a fast SSD if you can.

  2. Extract the Archive Use 7-Zip (free) or Windows built-in extractor. Extract to a folder like D:\AI\ComfyUI or C:\ComfyUI. Tip: If Windows blocks the file, right-click -> Properties -> General -> Unblock.

  3. Update NVIDIA Drivers Grab the latest Studio or Game Ready drivers from https://www.nvidia.com/Download/index.aspx. Restart your PC after installing.

  4. Place Your Models (Do This Before Launching) Put checkpoint files (.safetensors or .ckpt) in: ComfyUI\models\checkpoints\ VAEs -> models\vae\ ControlNet models -> models\controlnet\ (Subfolders are fine - ComfyUI scans them automatically.) Good starters: Nova Anime XL or Juggernaut XL - download from Civitai or Hugging Face.

  5. Launch ComfyUI Double-click run_nvidia_gpu.bat (or the cu128/cu126 version if you chose that). First run takes ~1-2 minutes (it sets up deps). Your browser should open http://127.0.0.1:8188 automatically. You’ll see “Starting server…” in the terminal - totally normal.

Option 2: Manual Install (Git Updates or Customization)

  1. Prerequisites

  2. Clone the Repo Open Command Prompt/PowerShell:

    Terminal window
    git clone https://github.com/Comfy-Org/ComfyUI.git
    cd ComfyUI

  3. Virtual Environment & PyTorch

    Terminal window
    python -m venv venv
    venv\Scripts\activate
    pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130

  4. Install Dependencies

    Terminal window
    pip install -r requirements.txt

  5. Launch

    Terminal window
    python main.py

    (Add --listen for network access, --lowvram if needed.)

Portable is way easier for most people - updates are just re-downloading and extracting.

After Installation (Cloud or Local)

These steps apply whether you installed locally or are using a cloud instance.

Install ComfyUI-Manager (Do This Right Away)

Manager is essential - it lets you install/update custom nodes (DWPose, ControlNet Aux, JoyCaption, etc.) with one click.

  • In portable: Look for Install-manager-for-portable-version.bat in the folder (run it).
  • Or manually: Download from https://github.com/ltdrdata/ComfyUI-Manager, drop into custom_nodes/, restart ComfyUI.
  • Once running: You’ll see a Manager button top-right (or enable with --enable-manager in the .bat file).
  • Click “Install Missing Custom Nodes” if prompted on first workflows.

First Test Workflow

  1. Load the default workflow (or drag a simple one from Civitai).
  2. Load Checkpoint -> pick your model.
  3. CLIP Text Encode -> add positive/negative prompts.
  4. KSampler -> 20-30 steps, CFG 7-8.
  5. VAE Decode -> Preview Image or Save Image.
  6. Hit Queue Prompt - your first local generation!

Common Gotchas & Tips

  • Slow first load? Normal - it scans models.
  • OOM errors? Try --lowvram or --disable-smart-memory flags.
  • Updates: Re-download portable (overwrite files, keep your models folder) or git pull + pip install -r requirements.txt --upgrade in manual.

You’re now set up for unlimited, private, fully customizable generations.