ComfyUI beginner guide (2026): cloud, desktop & first workflow

ComfyUI is an open-source node-based application and inference engine for image, video, audio and other generative workflows. The model and workflow determine the hardware, nodes and files required. This guide was checked September 13, 2026 against official ComfyUI documentation; it does not prescribe a universal GPU or invent performance benchmarks.

Choose your starting route

1. Use Comfy Cloud for the least setup

Comfy Cloud is the official hosted version. It provides preinstalled models and custom nodes and uses server GPUs, so a beginner can choose a template, update the input or prompt, run it and download the result without installing Python or model files.

Choose it when the goal is understanding nodes and workflows before managing hardware. Cloud is a subscription service, and its available models and custom nodes are curated. Check the current account price and supported template before committing a larger project.

2. Use Desktop for a supported managed local install

ComfyUI’s system requirements list Desktop for Windows with NVIDIA hardware and macOS on Apple Silicon. Desktop manages its own Python environment and dependencies and is currently beta. The stable Desktop release can trail portable or manual builds.

Follow the exact Windows Desktop guide or macOS Desktop guide. Choose the recommended GPU option, install into a dedicated location with enough storage, and keep models and custom nodes inside the managed environment. Do not install node dependencies into the system Python by accident.

3. Use Portable for a self-contained Windows install

ComfyUI Portable bundles its own Python environment for Windows and can run with NVIDIA GPUs or CPU mode. Extract the package and launch the matching batch file. Portable tracks newer features more closely than the stable Desktop build.

Use CPU mode only to prove the application starts; most modern generation workflows are impractical there. Select the package that matches the GPU and driver environment, and use the included update scripts rather than mixing instructions from another install type.

4. Use manual or CLI installation for custom environments

comfy-cli can install, launch, update, snapshot and bisect local ComfyUI installations. Manual and CLI routes fit Linux, controlled Python environments, automation and teams that need to pin versions.

This route owns the most maintenance: Python, PyTorch, device drivers, system libraries, model files and node dependencies. Record the ComfyUI commit or release, Python and PyTorch versions, GPU/runtime, workflow JSON, model checkpoints and custom-node versions before calling a result reproducible.

Run your first workflow in seven steps

1. Start from an official template. Open the Templates panel and choose the simplest workflow matching the task and model family. Official templates can detect missing model files and prompt for supported downloads.

2. Run it before editing. A successful unchanged run proves the install, model path, node set and device can work together. Save that known-good state.

3. Read the graph left to right. Identify model loading, prompt or input conditioning, latent or media processing, sampling or inference, decoding and save/output nodes. Follow connections rather than memorizing node names.

4. Change one input. Edit the prompt or replace the permitted source image. Keep the model, seed and other settings fixed so the effect is interpretable.

5. Queue and inspect the complete output. Use Queue or Ctrl+Enter. Review the saved file at its intended size and duration, not only the node preview.

6. Save the workflow with the output. Retain the workflow JSON, exact model files, node versions, seed and input assets. PNG metadata can carry a workflow, but do not assume every export or sharing service preserves it.

7. Duplicate the known-good workflow before experimenting. Change one node or parameter, label the variant and keep the prior version available for comparison.

Understand the minimum graph

ComfyUI’s official text-to-image tutorial explains the basic checkpoint, text-encoding, sampling and save-image path. A checkpoint commonly bundles or connects a denoising model, text encoder and VAE. Different model families can use different loaders and conditioning nodes.

Do not copy recommended sampler, step or guidance values from an unrelated model. Start with the official template for the exact model family. A setting that is valid for SD1.5, FLUX, a video model or a distilled checkpoint may behave differently elsewhere.

Models, workflows and custom nodes are separate dependencies

• Workflow JSON describes nodes, connections and settings; it does not contain every model or node implementation.

• Model files can be large and may have separate licenses, required encoders, VAEs or memory needs.

• Core nodes ship with ComfyUI and follow its release lifecycle.

• Custom nodes execute third-party code and can add Python or system dependencies.

• Hosted partner nodes call remote services and can have accounts, prices, retention and terms separate from local inference.

ComfyUI’s custom-node installation guide tells users to understand a plugin and avoid unknown sources. Install the fewest nodes required by the workflow, record their versions and review their repositories and dependencies. A workflow from another person is not automatically safe or complete.

Troubleshoot from the first failing boundary

A model name is null or missing. Confirm that the exact required files exist in the documented model directory, then refresh or restart using the instructions for your install type. Do not rename arbitrary files until the workflow’s loader and model family are understood.

A node is red or missing. Identify whether it is a core or custom node. Install only the named trusted dependency, restart, and rerun the unchanged workflow before adding anything else.

Out of memory. Reduce batch size, resolution, duration or model size according to that model’s guide. Close other GPU work. Do not promise a universal VRAM minimum; node graph, weights, precision, offloading and media dimensions all matter.

The output is black, corrupt or unrelated. Return to the known-good template, verify model and VAE compatibility, check node connections, and change one variable at a time. More negative-prompt terms are not a general repair.

An update broke the graph. Follow the official update guide for the install type. Restore the recorded core and custom-node versions or use a saved snapshot before changing the workflow itself.

Make a complex workflow easier to use

ComfyUI APP mode can expose selected workflow inputs and outputs through a simplified interface while keeping the node graph underneath. Use it after the workflow is stable, not as a substitute for recording its dependencies and versions.

For a team handoff, package the workflow, model/source links and licenses, custom-node list, environment versions, test inputs, expected outputs and a short failure guide. A screenshot of the graph is useful context but cannot reproduce the pipeline.

Related workflows

For current downloadable video checkpoints, use the open-weight AI video model guide. For an API integration instead of local nodes, compare image and video generation APIs.