5 weaknesses in the video generation model “Sora”


Although it is a model with too high performance, it also publishes some failure outputs, so publish them.⬇️

It seems that the chair dug up is not considered a hard substance.
The physics of the chair is also problematic, and it will eventually float in the air.🪑
From:@SuguruKun_ai

2ˇ When containing a large number of objects, animals and people often appear suddenly.
The prompt specified “5 gray wolves”, but they started with 4 and eventually increased to about 10.

3ˇ They don’t seem to be good at recreating “complex interactions” between multiple objects and characters.ㅤ
The video uses the long prompt at the top of the demonstration video, but even though it says “Blow out the candle to put it out” and “The light of the candle goes out” twice, the fire does not go out. It has not disappeared yet.

4ˇ Physical modeling is inaccurate It looks good up to the middle (the explosion itself is specified by the prompt), but the second ball misses the goal.

5ˇ Objects are suspended. This is not an official weakness, but I personally think it is worth releasing.
About ten seconds later, despite no specific instructions, the cloud giant stopped. This has also occurred in models such as Gen2, but there also seems to be a tendency for subjects to suddenly and unnaturally stop moving.

Original X post:
https://twitter.com/SuguruKun_ai/status/1758305222043312254

Scroll to Top