news

is sora out of date? meta's "strongest video model" does not use dit, but uses llama to work miracles

2024-10-05

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

author|wang zhaoyang

at this juncture when the main technical leader of openai sora went to google and multiple reports pointed out that openai sora was having difficulties with internal quality issues, meta unceremoniously released its video model "movie gen" and directly used a complete the evaluation system declared that it had defeated sora.

what's even more cruel is that meta is also "killing people". although this model is not open to the public like sora, it has made the 95-page technical report of the new model (not open source, but contains a lot of details) public and told everyone :

this model not only defeated sora in effect, but also used a new technical route - which proved that sora's technical route is no longer the most advanced today.

dear vincent video players, please don’t “copy” sora.

1

"media pedestal model"

to be precise, meta released a series of models, a combination created to achieve "ai-generated media content". this is also what the title of this technical paper means: movie gen: a cast of media foundation models

this set includes:

the largest basic gen video generation model, movie gen video, has 30 billion parameters.

the largest basic video generation audio model, movie gen audio, has 13 billion parameters.

the personalized movie gen video obtained by further post-training the movie gen video model is used to generate personalized videos based on individual faces. and a new post-training process that generates movie gen edit for precise video editing.