Advertising disclosure: We may earn a commission when you buy through links on this page. It does not change the price you pay or the order of the list.
ComfyUI: local image and video generation, step by step
ComfyUI is a free, open-source app for generating images and video on your own graphics card, built from connected nodes instead of a single prompt box. It runs FLUX, Stable Diffusion and newer video models; install the desktop or portable version, load a ready workflow, and check the widget below to see what your card can handle.
Which models fit
Runs on it
- DeepSeek-R1-Distill-Qwen 32B · Q6~40 tokens/s
- Qwen3 32B · Q6~40 tokens/s
- Qwen3 30B-A3B (MoE) · Q6~150+ tokens/s
- Gemma 3 27B · Q8~37 tokens/s
- Mistral Small 3.2 24B · Q8~42 tokens/s
- Magistral Small 1.2 24B · Q8~42 tokens/s
- Mistral Small 24B · Q8~43 tokens/s
- gpt-oss 20B · MXFP4~150+ tokens/s
Too big for it
- Llama 3.3 70BRent a GPU →
- DeepSeek-R1-Distill-Llama 70BRent a GPU →
- Qwen2.5 72BRent a GPU →
Best quality that fits, 8,192 tokens of context, one user. Speed is an estimate from memory bandwidth (generation reads every active weight once per token); real numbers depend on the program and settings.
On your computer or on a server?
On your own computer
Full privacy and no subscription; the model size is limited by your memory.
Which hardware →On a VPS
Reachable from anywhere and shareable with a team; small models run on CPU.
Ollama on a VPS →What ComfyUI is and how it works
Instead of a form with a prompt field, ComfyUI shows the whole generation pipeline as a graph: a node loads the model, another encodes your prompt, another samples the image, another saves it. You connect them with wires. It looks intimidating at first, but it means you can see and change every step, and add upscaling, ControlNet, inpainting or video stages exactly where you want them.
Everything runs in your browser against a local server, or inside the desktop app, and nothing is uploaded anywhere. New models usually get ComfyUI support first, which is a big reason it has become the standard tool for local image generation.
There are three ways to install it. ComfyUI Desktop is a regular installer for Windows and macOS that sets up Python and the Manager for you. The portable package for Windows is a folder you unzip and start, with versions for NVIDIA, AMD and Intel cards. A manual install from GitHub suits Linux users and anyone who wants full control over the Python environment.
ComfyUI Manager and ready-made workflows
You rarely build a graph from scratch. A ComfyUI workflow is saved as a JSON file, and images made in ComfyUI carry their workflow inside them, so dragging one onto the canvas rebuilds the whole setup. The built-in templates cover FLUX, Stable Diffusion and common video models.
ComfyUI Manager is the add-on that installs missing custom nodes and models a downloaded workflow needs. It comes ready in the desktop app; in portable and manual installs it is built into ComfyUI but has to be switched on, as the ComfyUI docs explain. Install custom nodes only from sources you trust, because they run code on your machine.
When your graphics card is too small
Image and especially video models are hungry for VRAM. An NVIDIA card is the smoothest path; AMD cards and Apple Silicon Macs work too, with more setup and slower results. If a model does not fit, use a smaller or quantized version, or rent a cloud GPU by the hour: many GPU clouds offer a ready ComfyUI template. Keep models on persistent storage and stop the instance as soon as you finish, because billing runs while it sits idle.
Install and run it in 5 steps
- Download ComfyUI Desktop from comfy.org, or the portable Windows package for your graphics card from the GitHub page.
- Run the installer or unzip the portable folder, then start ComfyUI and let it finish its first setup.
- Open the templates, pick a FLUX or Stable Diffusion workflow and let ComfyUI download the model it asks for.
- Type your prompt in the text node and press Run; the first image takes longest while the model loads.
- Use ComfyUI Manager to add missing nodes for workflows you download, and check the widget above for bigger models.
Questions people ask
What is ComfyUI used for?
ComfyUI generates images and video locally with models such as FLUX and Stable Diffusion. Its node-based workflows let you combine steps like upscaling, inpainting and ControlNet in one pipeline.
How do I install ComfyUI Manager?
The desktop app already includes it. In the portable or manual install it is built into ComfyUI but switched off: install its requirements and start ComfyUI with the option the ComfyUI docs give, and a Manager button appears.
Where can I find ComfyUI workflows?
Start with the built-in templates. Shared workflows come as JSON files or as images with the workflow embedded: drag either onto the canvas, then let the Manager install anything missing.
Can I run ComfyUI without a strong GPU?
Small models run on modest cards, and Macs work more slowly. For large FLUX or video models, rent a cloud GPU by the hour with a ComfyUI template and stop it when you are done.
More on local AI
Too slow? Rent a GPU from $0.33/hSee plan →