Atrocious model - object permanence/consistency is nonexistent
#38
by kabachuha - opened
Hello!
I'd like to like your model, the competition and stuff, but... the model is simply a disaster.
Censorship/copyright aside, I don't know how you are going to sell it as a "world model" - there is simply NO object permanence. Objects appear, disappear, morph, transform, the image2video doesn't hold the consistency. I'm using a dev model with diffusion decoder for "higher quality" and still all the bad things persist.
The current architecture / data approach is a dead end, sorry. You really need to rethink how the model works or a "world model" is not going to happen.
See the attached video.
Please, please fix it!
After the text encoder change there is some improvement in permanence / prompt-following!