> Still, I am mystified by how rapidly Sora went from launch to shutdown. Does anyone have any guess what happened there?

My guess is they over committed server/energy resources, since they were generating ~30 images per frame of 1 second of video for results that may be discarded and then tried again.

Now that energy costs are increasingly less predictable because of the war, they're prioritizing what is sustainable. Willing to blow up the $1 billion Disney deal for Sora, because that's a popular IP that would have increased discarded server time.

I'm also curious if Sora has been used by Iran to generate those Lego propaganda videos critical of the President. Given how close Sam Altman is with the current administration, I wouldn't be surprised if Sora is now reserved for U.S. government propaganda only.

Might be why the latest Iran propaganda video could be created in PowerPoint: https://bsky.app/profile/rachelbitecofer.bsky.social/post/3m...

Are there known tells that could be used to determine which model the video came from?

(This sort of question, and the Grok sexual abuse, is why I'd like to see mandatory invisible watermarks on generated images/video)

I don't think so. There are tons of self hosted models for video (they are smaller and easier to run).

Most people serious about this stuff usually have their own pipelines.

I'm not sure, but you could be right. Sora is/was the top-of-the-line platform for video generation, and the Lego IP videos were polished. Makes sense to outsource when your own energy grid is being destroyed. Anyone with an account and VPN could utilize the platform.

I'd like to know what self hosted models they've been using, if any, and who provided them, trained on Lego IP.

Since you seem to be better informed, I'm also interested in what self hosted models for video you recommend for creating my own Lego movie clips now that Sora is no longer an option for a paid service. There's tons, right?

Look up Wan and Hunyan for starters.

These are open weight models, so you can fine tune them on Lego content… But presumably they already have enough training data since they were made by Chinese companies who don’t give a shit about Western IP rights.