neuraltextualinversion / neti Goto Github PK

In NetiMaper.get_output(), here, the embedding variable is L2-normalized.

From reading the paper, the intent is to normalize the word embedding that is input to the text encoder. But if the output_bypass is enabled, then embedding is actually the concatenation of the word embedding and the output bypass. By doing normalization as implemented, the word embedding will be scaled down too much.

Also, great project, thanks for your work!

In the paper "An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion" that text inversion can do style transfer. I think that NeTI can also do this. But there are no related experiments in NeTI paper.
So is it possible to make NeTI do style transfer?

Hello, in your project, I trained the characters(a girl) with 3 pictures and 10 pictures respectively. I found that in the results of training with 10 pictures, the faces of the characters in the pictures were processed worse. Why？

Hi! Thanks for the great paper!
Does the current code support generating images with 2+ personalized tokens? I can't seem to find a way to do it.

test teddybear model from 'https://drive.google.com/drive/folders/1fRLRpl-IRoKbHqQt6hQ7yFx9laUIa_-A'
File "scripts/inference.py", line 97, in main
for truncation_idx in infer_cfg.truncation_idxs:
TypeError: 'NoneType' object is not iterable