A lot of things have improved in the 23rd version
In images where the face is still not properly visible, the habit of generating fake faces remains.
I think there is a big tendency to create a new picture without recognizing the face of the picture properly or if the head is too small (such as the front). Same goes for the eyes of the picture
To improve the recognition of relatively small faces in the overall image, you can try changing the parameter shown in Screenshot 1 from 384 to 512 or higher. I attempted to create a node for this (Screenshot 2), but I’m too lazy to polish the code to a state where it wouldn’t be embarrassing to publish on GitHub. Maybe someday I’ll get around to it. For now, you can simply open the file comfy_extras/nodes_qwen.py, for example with a text editor, find the line total = int(384 * 384) and change it to 512 * 512. This isn’t a cure-all, but it will significantly reduce failures. And yes, if you have limited RAM, it’s better not to do this.
You always want open-source projects to achieve everything you expect in one go, but that's not currently feasible.