Textual inversion training produces random images

[HappyCatKitten]

this is only trained on 200 steps, but not getting even slight results.

[vladmandic]

wild guess, you have xformers enabled while training.

[quasiblob]

IIRC I got similar results when I once accidentally pointed "dataset directory" of "Train" tab to my source images folder, instead of using a folder that had images with captions, (or maybe it was empty directory can't remember exactly). I don't think having xformers on would have that big of an effect, I think I've trained embeddings with xformers on and off.

[MacroPolo]

TI didn't work properly for me and it was down some bad combination of dependency versions. Updating xformers to 0.0.17.dev446 fixed the problem. There has been a few newer versions of xformers since I updated, but I haven't tested with those.

The issue and solution is described in more detail on #7264 (comment)