/vt/ - Virtual Youtubers

>>45119284
Yeah I can get that because it will fundamentally change the output of the model that you are used to because it is straight up ignoring certain tokens that arnt aligned with CLIP hence part of the "not listening" or "ignoring my prompts" feel some models have.
However I think its less obvious if its only a single token missing here or there. See >>45114739 Vs the 8 first tokens in a row in Elysium V3
I dont think AI Sugoi can be attributed to incorrect CLIP positions.
Objectively speaking it is a broken model and should not be the default expected behaviour from them.
Fixing the CLIP positions doesnt change much when its only a few or so missing but you overhall improve the quality of the model for the better by making it perform as expected.