MFaceTech/InstantID

train problem

Closed this issue · 3 comments

I used 40,000 datasets for fine-tuning and trained for 8 epochs. I found that the face ID was not maintained.
Is this because there are not enough epochs or there is an error in the data?

image
image

Whether you choose to fine-tune or train from scratch, 40,000 images with 8 epochs should result in good ID preservation effects, but you must ensure that the image size to the model is 1024*1024.

Whether you choose to fine-tune or train from scratch, 40,000 images with 8 epochs should result in good ID preservation effects, but you must ensure that the image size to the model is 1024*1024.
Thank you for your attention. I am also wondering why this problem occurs, because my images are generated, so the size is 1024. Now I suspect that there is an error in my training data processing. I need to check.
Below is the training configuration and sample data:
image

During the data processing, is there anything that needs to be modified in get_face_info.py? For example, max_side in the resize_img function?