Modelpublic

Anzhc/SDXL-Text-Encoder-Longer-CLIP-L

sourceHugging Faceopenrailupdated 1y agoView on Hugging Face

7likes34downloads

Model Card

An experiment, with CLIP L trained with up to 770 tokens with ~10k anime dataset, without adjusting arch. Concatenation is used to accummulate features.

output(4)

output(5)

Token-adjusted, with images removed if they ca'nt meet token criteria:

output(6)

output(7)