how to filter low quality iamge and improve my model quality

[nairoj]

I tried training myself to train controlnet, but my results were poor on portraits and good on non-human images.

type	from	to
non-human
portraits

The questions that often come up on portraits are

1. The resulting images have weird colors, for example, like
2. The picture quality is low

I have used 400,000 image-text data to create 400,000 depth data and depth_leres data. I trained the model for 4 epochs with a batch size of 256 (using gradient accumulation) to obtain the results mentioned above.

I would like to have some empirical guidance on what may have gone wrong with my training. Do I need to add more data to improve the quality of the results? Is there any way to filter out low-quality images (like image with JPEG artifacts) in the dataset?