OpenAI introduces Sora, its text-to-video AI model

catculation@lemmy.zip to Technology@lemmy.world – 424 points –
OpenAI introduces Sora, its text-to-video AI model
theverge.com
141

You are viewing a single comment

The quality is really superior to what was shown with Lumiere. Even if this is cherry picking it seems miles above the competiton

I can't understand how the shadows and reflections are so accurate (not perfect, but convincing) like here or here.

The second one is easy as you don't need coherence between reflected and non-reflected stuff: Only the reflection is visible. The second one has lots of inconsistencies: I works kinda well if the reflected thing and reflection are close together in the image, it does tend to copy over uniformly-coloured tall lights, but OTOH it also invents completely new things.

Do people notice? Well, it depends. People do notice screen-space reflections being off in traditional rendering pipelines, not always, but it happens and those AI reflections are the same kind of "mostly there in most situations but let's cheap out to make it computationally feasible" type of deal: Ultimately processing information, tracking influence of one piece of data throughout the whole scene, comes with a minimum amount of required computational complexity and neither AI nor SSR do it.