"Sora 2 does seem to be a big step forward in resolving some of the technical problems associated with AI video, producing clips with much more realistic physics and body mechanics. It's not going to replace camera operators yet, but OpenAI is now out to take on the likes of Instagram and TikTok using the model for its own Sora social media app, which features 100% AI-generated video and allows you to deepfake your friends."
"Example footage shows Sora 2 can generate video of things like gymnastics routines, which used to be horrific with previous AI models. OpenAI says one problem is that video models are overoptimistic: they will morph objects and deform reality to successfully execute a text prompt. For example, if a basketball player missed a shot, the ball might teleport to the hoop. It says Sora 2 is much better at avoiding this and respecting the laws of physics."
Sora 2 improves AI video generation by producing more physically accurate, controllable clips with realistic body mechanics and synchronized dialogue and sound effects. The model persists world state across multiple shots and can insert real-world elements from reference footage, including humans, animals, and objects. Training focused on advanced world simulation and large-scale video pre- and post-training. The model reduces overoptimistic morphing and deformations that previously broke realism. A Sora social app uses 100% AI-generated video and supports deepfakes, raising concerns as some genuine filmed footage has been mistaken for AI-generated content.
Read at Creative Bloq
Unable to calculate read time
Collection
[
|
...
]