Retrained from scratch after pruning a few images. I think the results are more accurate than in the previous version, also a bit more consistent, but let's see what you all think. It definitely feels that the embedding shines at close range, not so much at wider shots. That may be a consequence of the dataset I chose, so maybe I'll try to address that in the future.

The TI seems already a bit overtrained at 400 steps. As usual, in case you're getting saturated images with a given prompt, try reducing the embedding weight (0.99 should already solve the issue in many cases). And if you need help building a working prompt, just follow the examples I'm posting or let me know ih the comments!