How to run ComfyUI without a good GPU

Most computers can't run current image and video models at a useful speed. You can shrink the model, wait a lot longer, or rent the graphics card. Here is what each choice costs you.

Your options

Current image models want a graphics card with 12 GB of memory or more, and video models want 24 GB or more (see ComfyUI GPU requirements). Without that, there are five ways to go:

Option	What you pay	Speed	Good for
Shrink the model to fit your card	Nothing extra	Slow to very slow	Pictures on a 6 to 12 GB card
Run on the processor only	Nothing extra	Minutes per small picture	Trying ComfyUI out, older small models
Rent a GPU computer by the hour	Every hour it's switched on, busy or not, plus storage	Fast	Long sessions, building your own workflows
A hosted ComfyUI website	Credits or a subscription	Fast, sometimes with a queue	Trying workflows without installing anything
Rent GPU time by the second	Only the seconds a job runs	Fast once started	Using models now and then, video

Make the model fit the card you have

If you have a graphics card, just a small one, you can often get a model running by giving something up:

Low-memory settings. Starting ComfyUI with --lowvram or --novram keeps most of the model in system memory and moves pieces to the card as needed. It runs, but much slower.
Compressed model files. fp8 files take half the memory of the originals, and GGUF files (loaded with an add-on) take a quarter to a half. You lose some detail and prompt accuracy.
Smaller settings. Smaller pictures, fewer video frames and lower video resolution all cut the working memory a job needs.
Smaller models. SDXL-class image models run on 8 GB, and some video models come in a small version (about 5 billion parameters) for cards like that.

This works for pictures. For video, the waits get long, often many minutes for a few seconds of low-resolution clip.

Run on the processor

ComfyUI can run with no graphics card at all if you start it with --cpu. Everything then happens on the processor, which is many times slower. Expect minutes for one small picture from an older model. Current image models take far longer, and video isn't realistic. It's a way to learn the interface, not to get work done.

Rent a GPU computer by the hour

Cloud GPU companies such as RunPod rent whole computers with a graphics card by the hour (RunPod calls them pods). Many come with ComfyUI already installed, and you open it in your browser. Prices run from well under a dollar an hour for a 24 GB card to a few dollars for an 80 GB one.

You pay for every hour the computer is switched on, whether it's working or waiting for you, and for the storage that keeps your models between sessions. You download the models and add-ons yourself, and you have to remember to switch it off. It's a good fit for long sessions and for building your own workflows.

A hosted ComfyUI website

Some websites run ComfyUI for you and show it in your browser. You pay with credits or a subscription and don't install anything. Your prompts and results live on their servers, and you get the models and add-ons they offer.

Rent GPU time by the second

Serverless GPUs start when a job arrives and switch off seconds after it finishes, and you pay only for the seconds in between. Nothing is charged while you're thinking about your next prompt. The catch is setup: you have to pack ComfyUI and the model into a worker image, create an endpoint for it and write the code that sends jobs and collects the results. The first job after a break also waits while the model loads.

That setup is what CozyGPU does for you. It creates the endpoints in your own RunPod account, one per model, and gives you a simple app to use them. With its settings, a picture from Z-Image Turbo costs about $0.0024 of GPU time and a 5-second MiniMax H3 video clip about $0.146. ComfyUI without the setup compares it with running ComfyUI yourself.

Which to pick

You have a 6 to 12 GB card and mostly want pictures: shrink the model first.
You want to build and share node workflows: an hourly GPU computer.
You want to try a few workflows without installing anything: a hosted website.
You want current image, video and audio models now and then, on a laptop, without paying while idle: rent by the second.

Common questions

Can I run ComfyUI without a GPU?

Yes. Start it with --cpu and it runs on the processor alone. It's fine for trying the interface, but a small, older image model takes minutes per picture, and current image and video models are too slow to be useful. To use those without a graphics card, rent one in the cloud.

Can I run ComfyUI on a laptop without a dedicated graphics card?

It will start, but built-in graphics (Intel UHD, Iris Xe, basic AMD Radeon) don't help much, so it runs on the processor and is very slow. A laptop is fine for using a cloud GPU, which does the heavy work elsewhere.

Can I run ComfyUI with 4 GB of VRAM?

With Stable Diffusion 1.5 at small sizes, yes. SDXL works with the --lowvram setting, slowly. Current large image models and video models don't fit in a useful way.

What is the cheapest way to run AI image and video models without a GPU?

If you use them now and then, renting GPU time by the second, because you pay nothing between jobs. With CozyGPU's settings a picture from Z-Image Turbo costs about $0.0024 of GPU time and a 5-second MiniMax H3 video clip about $0.146, plus CozyGPU's monthly price. If you keep a GPU busy for hours at a time, an hourly GPU computer can cost less per hour of work.

Do I need to install ComfyUI to use these models?

No. CozyGPU runs ComfyUI for you on a GPU it rents in your own RunPod account, and you use the models from a simple app on your Mac or PC. Nothing big is downloaded to your computer.

Is a GGUF model worse than the original?

A little. GGUF files store the model's numbers with fewer bits so the model fits in less memory. 8-bit versions are close to the original; 4-bit and smaller lose more detail and prompt accuracy.