https://files.catbox.moe/r0a4o8.webma recent update has allowed us to experiment more with individual block weights for the motion model, I've been experimenting for some better settings but each block has 2 attn's to weigh so we're talking like, 16 variables squared (half for effect, half for scale) so finding more promising settings will take time, but there could be more stability to be found. additionally, 2 new repos in regards to video generation models seem to imply even better cross context consistency, the brain behind animatediff in comfy is porting one to work, and theres a chance it could be a upgrade from what we have now.
