Problem at high resolutions using PuLID embeddings

[self-destruction]

I'm using a slightly modified version of the original PuLID. The change is to use a custom SDXL model, in my case Juggernaut-XL-v9. If you use a square resolution, everything works fine at low resolution. However, having tested the 2048x2048 resolution, the face literally blurred out of shape in the final image.
Unfortunately, I can't provide this image right now. But you can easily replicate it by taking my cli-version of PuLID and running it with the command python app.py --save_dir ‘{OUTPUT_DIR}/PuLID_output’ --save_file_name ‘{filename}’ --face_img ‘{face_img}’ --h 2048 --w 2048.
Also, if you want to examine this code, the first thing you should look at is this file: https://github.com/self-destruction/PuLID_cli/blob/main/pulid/pipeline.py. As you can see, I have temporarily commented out the using of hidiffusion there.
Thanks for your work!

[self-destruction]

I will also add that I use the following commands to install dependencies:

!pip install diffusers==0.27.0 
!pip install transformers==4.36.2
!pip install einops ftfy basicsr facexlib insightface onnxruntime-gpu accelerate timm apex opencv-python spaces==0.19.4
!pip install --pre xformers==0.0.20
!pip install torchvision==0.15.2+cu118 --index-url https://download.pytorch.org/whl/cu118
!pip install torch==2.0.1+cu118 --index-url https://download.pytorch.org/whl/cu118
!pip install --upgrade numpy
!pip install hidiffusion

Yes, it's a bit messy, but it works for me.