Unloads models back to system RAM between nodes instead of keeping them on the GPU, letting 6 GB cards run SDXL-sized workloads by trading speed. ComfyUI 4.x defaults to dynamic VRAM management, so this flag mainly helps on cards it misjudges. Expect slower iterations but fewer out-of-memory errors.
Looking for more? Search all 7,657 commands — works offline, in English or Spanish, and fixes typos.