You don't actually need (or want) all the random quality tags to get a face accurate in training. Only one you need is (in this case) the style "3D" so the trainer knows what it's looking at (most models do terrible if fed 3D images without being told its 3D, but its not a problem if they're told). You want varying angles and a lot of pictures with those angles tagged more than anything. Getting the face right is also really doesn't need (or want) you to train only the face, just have all the other frameings (at least from the front) not be so far zoomed out that the face is an indistinguishable blob.

If you want an idea of what a character dataset should look like for a character model, you can download the training data for my public domain Sherlock Holmes model (its under "optional files"). The picture quality is lower than ideal due to the constraint of the dataset, but knowing how low you can go for training is still helpful.