>>105633163
isnt that more a text encoder thing? as in the text encoder isnt moving "throws up" in the "vomit" direction as well as "vomits" is due to the extra semantic meaning in "throws up"?

because its not the video model understanding the text, its the text encoder encoding the text in a way the video model can understand, right?

>>105633183
no torch compile for you (i think) but 16gb of vram is fine (king of the vramlets, if you will.) even a 3060 12gb is reasonable enough for 480p sloppa now