ComfyUI GPU requirements: how much VRAM you need

The short answer

ComfyUI runs on almost any computer. What needs the memory is the model you load into it. The number that matters is VRAM, the graphics card's own memory, and the model has to fit in it, or mostly fit, to run at a useful speed. A rough guide for NVIDIA cards:

• 8 GB: SDXL-class image models. Compressed versions of bigger ones, slowly.
• 12 to 16 GB: most current image models as compressed files. Short, low-resolution video if you're patient.
• 24 to 32 GB: large image models, and 14-billion-parameter video models as compressed files at 480p to 720p.
• 48 GB and up: the biggest image models and video models at full quality. These are workstation and data center cards, and most people rent them.

What fits on common cards

These are the sizes people commonly report. The real limit moves with picture size, video length, the file format and your ComfyUI version, so treat each row as a starting point.

Why the model decides

A model is mostly a big file of numbers (parameters), and the graphics card has to hold them while it works. A quick way to estimate the memory a model needs:

• Full precision (bf16 or fp16 files): about 2 GB for every billion parameters. A 12-billion-parameter model needs about 24 GB.
• fp8 files: about 1 GB per billion. The same model needs about 12 GB, with a small loss in quality.
• 4-bit GGUF files: a little over half a gigabyte per billion, about 7 GB for that model. Quality drops more, and ComfyUI needs an add-on to load them.

On top of that come the text encoder (the part that reads your prompt, often 5 to 10 GB on its own), the decoder that turns the result into pixels, and working memory. Working memory grows with picture size and, for video, with every extra second and every step up in resolution. That's why video needs so much more than pictures.

What happens when the model doesn't fit

ComfyUI moves the parts that don't fit into your computer's system memory and swaps them in as needed. That works, but each swap costs time, and a job that takes seconds on a big card can take minutes or longer. If even that isn't enough, the job stops with an error like CUDA out of memory (torch.OutOfMemoryError).

What people do about it:

• Start ComfyUI with --lowvram or --novram, which keeps even more of the model in system memory. Slower still, but it may finish.
• Use a smaller picture size, fewer video frames or a lower video resolution.
• Switch to an fp8 or GGUF version of the model, or to a smaller model.
• Rent a bigger card. Running ComfyUI without a good GPU compares the options.

System memory and disk

When the graphics card is full, system memory (RAM) is where the rest of the model waits. 16 GB is a bare minimum, 32 GB is comfortable for large image models, and 64 GB helps a lot with video.

Model files are big and load faster from an SSD. For a sense of scale, the models CozyGPU runs, each packed with ComfyUI and everything it needs, range from 17 to 69 GB per model.

NVIDIA, AMD, Intel and Apple

• NVIDIA is what ComfyUI and nearly every add-on are built and tested on. If you're buying a card for this, it's the safe choice.
• AMD cards work through AMD's ROCm software, most smoothly on Linux. Support on Windows has improved, but more add-ons assume an NVIDIA card, and some won't run.
• Intel Arc cards work too, with fewer people testing them.
• Apple silicon Macs use their shared memory as graphics memory, so big models load, but they run several times slower. See ComfyUI on a Mac.
• Built-in graphics on most Windows laptops (Intel UHD, Iris Xe, basic AMD Radeon) are not practical for current models.

Laptops

A laptop graphics chip usually has less memory than the desktop card with the same name. The laptop RTX 4060 and 4070 have 8 GB, the laptop RTX 4080 has 12 GB, the laptop RTX 4090 has 16 GB and the laptop RTX 5090 has 24 GB. Laptops also slow down when they get hot, which long video jobs make them do.

The cards CozyGPU rents for each model

For comparison, these are the cloud GPUs CozyGPU picks for each model it runs. They're chosen to run every model at full quality with room to spare, so some models will also run on smaller cards at home as compressed files.

CozyGPU rents these in your own RunPod account by the second, so you don't need one at home. Running ComfyUI on a cloud GPU explains how that compares with renting a GPU computer by the hour.

Common questions

What is the minimum GPU for ComfyUI?

ComfyUI starts on almost anything, even with no graphics card at all. For current models, an NVIDIA card with 8 GB of memory is a practical floor for pictures, and 16 to 24 GB for video. Below that you need smaller or compressed models, a lot of patience, or a rented GPU.

Is 8 GB of VRAM enough for ComfyUI?

For SDXL-class image models, yes. Bigger image models need compressed files (GGUF or fp8) and run slower, because ComfyUI moves parts of the model into system memory. Video works only with small models at low resolution.

Is 12 GB of VRAM enough?

It's enough for most image models as fp8 or GGUF files. For video, expect short, low-resolution clips that take many minutes each.

How much VRAM do AI video models need?

The well-known open video models with about 14 billion parameters, like Wan 2.2, want about 24 GB to run as compressed files at 480p to 720p, and 48 to 80 GB at full quality. Smaller video models run on 8 to 12 GB at low resolution. CozyGPU rents 80 to 96 GB cards for its video models.

Does ComfyUI need an NVIDIA graphics card?

No, but NVIDIA cards are the best supported. AMD, Intel Arc and Apple silicon work with more setup, and some add-ons (custom nodes) only work on NVIDIA.

How much RAM does ComfyUI need?

16 GB of system memory is a bare minimum. 32 GB is comfortable for large image models, and 64 GB helps with video, because ComfyUI moves parts of a model into system memory when the graphics card is full.

How much disk space does ComfyUI need?

ComfyUI itself is small. The model files are not: an image model with its text encoder is often 10 to 30 GB, and a video model 20 to 60 GB. Plan for a fast SSD with a few hundred gigabytes free if you want to try several models.