ComfyUI has no single VRAM requirement. Its README says it can run even the biggest open-source models on as low as 4GB of VRAM plus 8GB of RAM by streaming weights, but more VRAM means models stay on the GPU and run faster.
What decides the number
- The model. The size of the checkpoint or diffusion model you load is the main factor. The model publisher's page is the place for its own requirements. Quantized checkpoints (FP8 and similar formats ComfyUI supports) are smaller.
- How ComfyUI loads it. By default models are unloaded to CPU memory after use;
--highvramkeeps them in GPU memory, and--lowvramand--novrampush the other way. See ComfyUI low VRAM mode. - What else is on the GPU. ComfyUI reserves some VRAM for your OS and other apps: see what
--reserve-vramdoes.
See your own numbers
At startup ComfyUI logs your total VRAM and RAM. With ComfyUI running, the /system_stats endpoint returns vram_total and vram_free per device:
curl http://127.0.0.1:8188/system_stats
If a run fails for lack of memory, go to the ComfyUI out of memory fixes. If you need a bigger GPU than you own: If you would rather not set the machine up yourself, the ComfyUI template on Aquanode provisions the box, installs ComfyUI and starts it when you deploy it; once the deployment is Ready, the Overview tab shows the ComfyUI URL.
Sources
- ComfyUI README. github.com
- ComfyUI cli_args.py (VRAM flags). github.com
- ComfyUI model_management.py (startup log). github.com
- ComfyUI server.py (/system_stats). github.com
- ComfyUI QUANTIZATION.md. github.com