AI video generator "accidentally" included Disney character in demo reel

[Carl]

(Luna came up in the VIG SIG meeting last night) 

Oops

https://boingboing.net/2024/06/21/ai-video-generator-accidentally-included-disney-character-in-demo-reel.html

An AI startup called Luma recently launched a new AI-generated  video product called "Dream Machine." On the website, the company pitches the tool as:

…An AI model that makes high quality, realistic videos fast from text and images. It is a highly scalable and efficient transformer model trained directly on videos making it capable of generating physically accurate, consistent and eventful shots. Dream Machine is our first step towards building a universal imagination engine and it is available to everyone now!

If that sounds like a bunch of AI-generated marketing copy gibberish — well, it is! But that's not all, folks. As a proof-of-concept, the company released a trailer (above) for a fully AI-generated movie called Monster Camp. On the surface, it mostly just looks like an uncanny valley ripoff of the Monsters, Inc. franchise, complete with all the weird hallucinations.

But look a little closer and … wait, is that Mike Wazowski from Monsters, Inc?!

[Walter Martinez]

That's funny! I love  Mike Wazowski!

--
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/rssc-list/558c260d-0979-4a69-95c5-547977b3c0f6n%40googlegroups.com.

[Chris Albertson]

I think the term is “overtraining”.  Counter-intuitively it is the result of not using enough training data.

In the ideal case, you have 20 tons of training data and a model that can hold 2 tons for parameters and the training process tries to find the best way to cram the data into a place where it will not fit.   As it turns out, the ONLY way is to reduce the data to rules and generalizations that let the model later emit something that was derived from the data but not the data itself because that was compressed away.

But with too little data, some of it can be stored literally.

This is going to be a big issue if people use copyrighted data to train their models.  Then you can be sued if your commercially released model spits out verbatim copies of the copywritten material.

To view this discussion on the web visit https://groups.google.com/d/msgid/rssc-list/CAPjT4OdQ8LMa3BpqxZc8QqkmeTd7vifQA5N6c_CbXigKSwyZow%40mail.gmail.com.