An error occurred in the image: OutOfCacheError: CUDA out of memory Tried to allocate 256.00 MiB. GPU

When I use instant ID to change faces in the torch2.31+CUDA121 environment, I encounter OutOfCacheError: CUDA out of memory Tried to allocate 256.00 MiB GPU， The image size is 1024px * 1024px, and my graphics card is ntx1060 6G. This can run perfectly with torch 2.1+CUDA121. Isn't torch 2.1+CUDA121 more stable? TORCH2.31+CUDA121 can run with xFormers, but the speed will be slower.

Start by reproducing the Instant ID face-change workflow with the reported 1024px image, GTX 1060 6GB GPU, and the two stated PyTorch/CUDA environments. Compare memory behavior with and without xFormers, then document the confirmed compatibility issue or supported workaround; the issue names no source file or test to inspect.

Written by the indexing model from the issue text.