AI image replacement - What sorcery is this?

[daqq]

[helius]

It's an application of TensorFlow or something similar, trained on a dataset of internet images. The "Partial Convolutions" are some type of visual frequency data extracted from the object image that maps onto a portion of the dataset allowing it to be matched.
For example. the convolution of the image data around a missing eye is matched with a representation of eyes in the training data.

[chickenHeadKnob]

I am more amazed that Ernest Borgnine is still alive.

[BrianHG]

I like how often my posts are duplicated...
https://www.eevblog.com/forum/chat/nvidia-ai-reconstructs-missing-areas-in-photos-with-amazing-results/msg1490787/#msg1490787

[daqq]

Whoops, my apologies!

[BrianHG]

Whoops, my apologies!

No problem, I actually think it's cool.

[TerraHertz]

Google "deepfakes"

Caution: You may see porn in the search results.
Anyway... an AI-assisted method of creating completely fake video of people doing stuff, using a mix of bodies from one video and faces from another.

Btw, utilities to do this with voice (AI learning the characteristics of someone's voice and speech mannerisms, then constructing 'that voice' saying whatever you script it to) have existed for a few years now.
You can't rely on audio to be the truth, and now not even video.

[BrianHG]

Btw, utilities to do this with voice (AI learning the characteristics of someone's voice and speech mannerisms, then constructing 'that voice' saying whatever you script it to) have existed for a few years now.

Not at this new AI based text to speech quality level:
https://www.eevblog.com/forum/chat/latest-google-text-reader-ai-is-unbelivably-near-perfect/