curl -X POST http://127.0.0.1:8188/free -H 'Content-Type: application/json' -d '{"unload_models": true, "free_memory": true}'

Unloads loaded models from VRAM and frees cached memory, the recovery move after an out-of-memory error or before switching to a heavy workflow. The JSON body toggles model unloading and memory freeing independently. Retry your workflow after the VRAM drops.