Learn ComfyUI
The node-based interface for Stable Diffusion, Flux, Wan and LTX. Build the pipeline as a graph, and get the whole workflow saved inside every PNG.
What ComfyUI is good at
- Full control of every step of generation
- Workflows saved as JSON and embedded in output images
- Open video models (Wan, LTX) on your own GPU
- Custom nodes for almost anything
Quickstart
- Load the default workflowCheckpoint, prompts, latent, KSampler, decode, save: learn these six nodes first.
- Change one node at a timeFix the seed in KSampler and change steps, cfg or sampler one by one.
- Save the workflowExport the JSON and keep it next to the images; any output PNG also carries it.
- Install the ManagerComfyUI Manager installs custom nodes and finds missing ones when you load someone's workflow.
Settings cheat-sheet
| Setting | What it does | Start with |
|---|---|---|
| Checkpoint | Load Checkpoint node | A model you know |
| Prompts | CLIP Text Encode positive / negative | Short negative |
| Latent size | Empty Latent Image | Native model size |
| KSampler | seed, steps, cfg, sampler, scheduler, denoise | Model defaults, seed fixed |
| Save Image | Output path and prefix | One folder per project |
| Manager | Custom node installer | Install first |
ComfyUI guides
Guides for ComfyUI are on their way. The quickstart and cheat-sheet above are kept current in the meantime.
The five most-asked questions
What is ComfyUI?
An open-source, node-based interface for diffusion models. You connect nodes (load model, encode prompt, sample, decode, save) into a workflow that runs on your GPU or a hosted machine.
Which models can I run?
Stable Diffusion families, Flux, Wan, LTX-Video and many others with the right nodes, subject to your GPU memory. Model cards state the needs.
Where is my prompt saved?
In the output PNG's metadata, together with the complete workflow. Drag the PNG back into ComfyUI and the workflow reloads with every setting.
What do the KSampler settings mean?
seed (repeatability), steps (how many denoising passes), cfg (prompt adherence), sampler and scheduler (the denoising method) and denoise (how much of the latent is changed). Start from the model's recommendations.
Is ComfyUI free?
Yes, it is open source. You pay only for hardware or a hosted GPU if you use one, and model licences apply to the images.