An out-of-memory error means the model plus the working memory for the run did not fit in your GPU's VRAM. Fix it by making ComfyUI offload more, loading a smaller model, or moving to a GPU with more memory.
Try these in order
- Update ComfyUI. Newer releases may load models differently (see dynamic VRAM). How to update ComfyUI.
- Check what else is using the GPU. If other software holds VRAM, raise the amount ComfyUI leaves for it with
--reserve-vram. - Force offloading. Start with
--lowvram. The help text says it makes the text encoders run on the CPU, and that it does nothing when dynamic VRAM is enabled. If that is not enough,--novramis documented as "when lowvram isn't enough". - Be aggressive about RAM.
--disable-smart-memoryforces ComfyUI to offload to regular RAM instead of keeping models in VRAM. - Move the VAE to the CPU.
--cpu-vaeruns the VAE on the CPU. - Use a smaller model. A quantized checkpoint takes less memory; ComfyUI documents its quantized formats (FP8, NVFP4 and others) in
QUANTIZATION.md.
python main.py --lowvram --cpu-vae
If an error repeats after all of that, the model likely needs a bigger GPU than you have: see how much VRAM ComfyUI needs. To free memory between runs without restarting, see how to clear VRAM in ComfyUI.
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 cli_args.py (flag help text). github.com
- ComfyUI QUANTIZATION.md. github.com