ComfyUI troubleshooting, written as fixes
Every page here starts from a specific error message, explains why it happens, and ends with a change you can apply. No “reinstall everything” advice, no video you have to scrub through.
Fixes
-
How to Fix ComfyUI “CUDA Out of Memory” Error with FLUX.1 Models
FLUX.1-dev wants ~34GB of VRAM at full precision. The fp8, GGUF and launch-flag combination that fixes the CUDA OOM error on 8GB, 12GB and 16GB GPUs — plus how to tell fragmentation apart from a genuine shortage.
What this site covers
ComfyUI and the local image-generation stack it sits on — PyTorch, the NVIDIA driver, GGUF quantization, and the model files themselves. The failures worth documenting are the ones where the error message points somewhere other than the actual cause, and those are the pages you will find here.
Coming next: missing-node errors after a ComfyUI-Manager update, RuntimeError: expected scalar type mismatches, and safetensors load failures.
Each fix page lists the GPU and driver configuration it was verified on, so you can tell whether it applies to your machine before you spend an hour on it.