How to Use ComfyUI: A Complete Step-by-Step Tutorial

ComfyUI is free and open-source (GPL-3.0). Run it locally with zero subscription cost, or use the cloud version if you don't have a GPU. Two paths:

Local desktop install (step by step)

1. Go to comfy.org/download and grab the installer for your OS.
2. Extract the archive to a short path on an SSD — e.g. C:\ComfyUI\. Long paths or slow HDDs cause model-loading errors on Windows.
3. Run the launcher for your hardware: run_nvidia_gpu.bat for Nvidia GPU on Windows, run_cpu.bat for CPU only (very slow). On Mac: the Desktop app handles this automatically via Metal.
4. Your browser opens at http://127.0.0.1:8188 — this is your local server. Keep the terminal window open while using ComfyUI.

Manual install via GitHub (advanced)

# Clone the official repo
git clone https://github.com/comfy-org/comfyui
cd comfyui

# Install Python dependencies
pip install -r requirements.txt

# Launch with Nvidia GPU
python main.py --cuda-device 0

# Launch CPU-only
python main.py --cpu

# Launch with low VRAM optimizations
python main.py --lowvram

ComfyUI loads a dark canvas in your browser. This is your node graph workspace — a visual programming environment where AI processing steps are wired together like a circuit board.

A default text-to-image graph looks roughly like this on the canvas:

• Load Checkpoint (left) outputs MODEL, CLIP, and VAE connections
• CLIP Text Encode nodes (top middle) turn your positive and negative prompts into CONDITIONING
• KSampler (middle) takes MODEL, CONDITIONING, and a LATENT image and runs the diffusion steps (steps, cfg, seed, sampler, scheduler)
• VAE Decode (right) turns the sampled LATENT into a pixel IMAGE
• A Control Panel on the right holds Queue Prompt, Save, Load, and Load Default, plus a queue/running status readout

ComfyUI is the interface — the intelligence comes from a checkpoint model, a 2–12 GB file trained on billions of images. Without one, nothing generates. Here's exactly which model to start with and where to get it.

Which model to start with

Where to put the files

# Checkpoint models
ComfyUI/models/checkpoints/your-model.safetensors

# LoRA fine-tunes
ComfyUI/models/loras/your-lora.safetensors

# Upscale models (RealESRGAN etc.)
ComfyUI/models/upscale_models/RealESRGAN_x4plus.pth

# ControlNet models
ComfyUI/models/controlnet/control_v11p_sd15_canny.pth

# VAE (optional, some models need separate VAE)
ComfyUI/models/vae/vae-ft-mse-840000-ema-pruned.safetensors

After placing a file, click Refresh in the control panel. It will appear in the relevant node's dropdown instantly.

Press Ctrl+D to load the default Text-to-Image workflow. You'll get a pre-wired graph — model loader, text encoders, sampler, VAE decoder, image saver. Ready to run.

1. Click the Load Checkpoint node dropdown → pick your downloaded .safetensors file.
2. Positive prompt (top CLIP Text Encode): be specific. "a majestic snow leopard on a mountain peak at golden hour, ultra-detailed fur, photorealistic, dramatic lighting"
3. Negative prompt (bottom CLIP Text Encode): "blurry, low quality, deformed, watermark, text, extra limbs, bad anatomy"
4. Resolution in Empty Latent Image node: 1024×1024 for SDXL · 512×512 for SD 1.5 · 1024×1024 for FLUX.
5. Press Queue Prompt or Ctrl+Enter. A green border sweeps through each node as it processes. Your image appears in the Save Image node output.

This is the most important section. Once you understand data flow, you can build any workflow — not just copy them. Each wire carries one data type. Mismatch types and ComfyUI won't let you connect them — it's a built-in type system.

Standard text-to-image node flow

1. Load Checkpoint outputs MODEL, CLIP, and VAE.
2. CLIP feeds two CLIP Text Encode nodes — one for the positive prompt, one for the negative prompt — each producing CONDITIONING.
3. An Empty Latent Image node produces a blank LATENT at your chosen resolution.
4. KSampler takes MODEL, the positive and negative CONDITIONING, and the LATENT, and outputs a sampled LATENT.
5. VAE Decode takes the sampled LATENT plus VAE and outputs a pixel IMAGE.
6. Save Image takes the IMAGE and writes it to disk.

Your base generation might be 512px or 1024px. An upscale node takes that output and runs it through a dedicated super-resolution model — producing a 2048–4096px result with recovered fine detail that the base model couldn't render at small scale.

Add the upscaling nodes

1. Download RealESRGAN_x4plus.pth from xinntao/Real-ESRGAN releases. Place it in ComfyUI/models/upscale_models/.
2. Double-click canvas → search Load Upscale Model → add it. Select RealESRGAN_x4plus from its dropdown.
3. Double-click canvas → search Upscale Image (using Model) → add it. Wire: Load Upscale Model output → Upscale Image (UPSCALE_MODEL pin).
4. Rewire image flow: VAE Decode output → Upscale Image (IMAGE pin) → Save Image.
5. Queue Prompt. Output will be 4× your base resolution with recovered detail.

ControlNet types at a glance

How to set it up

1. Via ComfyUI Manager → Install Custom Nodes → search ComfyUI-ControlNet-Aux. Install it and restart.
2. Download a ControlNet model matching your base checkpoint. For SDXL: controlnet-canny-sdxl. For SD 1.5: control_v11p_sd15_openpose. Place in ComfyUI/models/controlnet/.
3. Add a Load Image node → feed your reference photo into a preprocessor node (e.g. DWPose Estimator for poses, Canny for edges).
4. Add Load ControlNet Model → select your downloaded model.
5. Add Apply ControlNet node. Wire: preprocessor image output → Apply ControlNet IMAGE pin. Wire: Load ControlNet → CONTROL_NET pin. Wire: your positive CONDITIONING → CONDITIONING pin.
6. The Apply ControlNet output CONDITIONING replaces your positive conditioning into the KSampler. Adjust strength (0.5–1.0) to control how strongly the structure is enforced.

FLUX vs. SD node differences

Set it up in 5 steps

1. Download FLUX.1 Dev (.safetensors) from Hugging Face / black-forest-labs. Place in ComfyUI/models/unet/ (not checkpoints — FLUX uses a separate unet folder).
2. Download the FLUX VAE (ae.safetensors) and both text encoders (t5xxl_fp8_e4m3fn.safetensors + clip_l.safetensors) from the same repo. Place VAE in models/vae/ and encoders in models/clip/.
3. In ComfyUI, press Ctrl+D → look for a FLUX template in the built-in library, or load one from the Comfy.org workflow browser.
4. In the Dual CLIP Loader node, select t5xxl_fp8_e4m3fn for clip_name1 and clip_l for clip_name2. Type: flux.
5. Write your prompt in the CLIP Text Encode node. FLUX handles long, natural-language prompts well — describe the scene in sentences. No negative prompt needed. Queue and run.

Save your workflow

Click Save → exports your full node graph as a .json file. Every node, every setting, every wire. Share this file and anyone can load your exact workflow.

The workflow-in-image trick

Every PNG ComfyUI outputs has the full workflow embedded in its metadata. Drag any ComfyUI-generated PNG onto the canvas and the complete node graph loads instantly — no JSON file needed. This is how the community shares workflows without documentation.

Built-in template library

Install ComfyUI Manager first — it's the package manager for everything else. Without it, installing nodes requires manually cloning GitHub repos.

# From inside your ComfyUI directory:
cd custom_nodes
git clone https://github.com/Comfy-Org/ComfyUI-Manager
# Restart ComfyUI — Manager button appears in UI

You've completed the foundation. Here's the learning ladder — each level unlocks new categories of output and control.

AI Diffusion Myths vs. Truths

Official links & community