lldacing
Dependencies (requirements) details
The current requirements.txt says:
- numpy
- opencv-python
- timm
But this isn't complete, BiRefNet code imports:
- huggingface_hub
- functools
In addition, if you mention numpy (which is pulled by ComfyUI) you should also mention:
- safetensors
- einops
- kornia
- Pillow
I agree that mentioning torch and torchvision is dangerous, even when the Manager black lists them.
I also think that removing OpenCV and timm is desirable, a quick look at the code shows that CV2 can be avoided. You are using it in the modified image processor, which already claims to be adapted for PIL. So I think it should just use PIL instead of cv2.blur(). In the case of timm a quick search makes me think it just uses the DropPath layer implementation, so copying it (and its helpers) should be enough to get rid of it.
If I'm correct a few changes should allow to reduce the dependencies to just things pulled by ComfyUI. Are you interested on it?
