LARRUZ uses ChatGPT to write a sitcom script with Joe Biden and Donald Trump, then generates their voices with ElevenLabs. The fictional product: Hair Force One, a hair-care line for world domination.
Hair Force One
Reddit user put together a sitcom featuring Joe Biden, Donald Trump, and it's 100% AI generated. Watch this.
Hey Don, what do you want to do tonight?
The same thing we do every night, Joe. Try to take over the world.
And how are we going to do that, Don?
We'll launch a line of hair care products called Hair Force One. We'll make billions and use the profits to control the world.
But Don, isn't your hair already a national treasure?
Why thank you, Joe. With your smile and my hair, we'll be unstoppable.
Okay, so in this case, the user used this prompt. We generate the script for the sitcom using ChatGPT.
How do you feel about adding some ice cream to the mix, Joe?
Ice cream doesn't belong in hair products. But since it's your favorite, we can have some as a treat after we finish our batch.
Ice cream. I'll scoop to that, Knorf.
Okay. And for Weiss's, he used 11 labs. So I have a video, I'll put a link to it. You can actually use 11 labs to create realistic voices. So here is like one that's available on the website.
The most realistic and versatile AI speech software ever.
Eleven brings the most, but you can fine tune it to any voice you want.
Ladies and gentlemen, we present to you Hair Force 1, the ultimate hair care solution for world domination.
It'll make your hair so soft you'll feel like a fluffy kitten.
We're making money. Hand over comb, Joe. World dominate.
Okay. It's actually pretty well done because it's not only the script but the prompt that we used to generate these videos as well. And the videos were generated using another open source software. So this is a model called a text to video synthesis model Scoop. I have a video of how to create videos using the model Scoop. So I'm going to put a link to that as well. This is just a glimpse of what personal entertainment could look like in the future. Now you might think that the video quality is pretty bad, but later in the video I'm going to show you another open source model which is able to generate pretty consistent and high quality videos. Now here's another example of somebody put together a text video for Iron man flying to meet his fans. And it's actually pretty well done as well. And the way it's done is they put together just like three seconds clips and and then I made a 28 second video out of it. However, the group from Microsoft Research just published this paper at diffusion or diffusion for extremely long video generation. And here they showed that you can generate much longer videos. So they actually give an example of a Flintstone episode which is 11 minutes long. And here is the episode. So the weight versus for different segments. They also included these prompts that were provided. So for example, a cartoon title card for the Flintstones, right? That's like the first one and then Fred is saying something in the room and it generated the video. So they are simply combining those together to get much longer videos. And the quality of these videos are much better. There is a lot more temporal coherence in this video compared to what we just saw. It's simply amazing to look at the progress because the first text to video open source softwares was released just last week. And now you have this new software with the help of which you can generate much longer videos. And they look much better. You don't really see any deformities at all. So it's amazing how much progress has been made in such a short period of time. So here's a quick overview of this paper from Microsoft. So initially they had these 16 different prompts and based on each one of them they generated 16 frames. So here are the frames that are generated. Then they simply extrapolate these frames and extended those to 26 frames. Then like even further extending them to over 3,000 frames. And then they simply put together to get a full video out of it. In closing of the video, I wanted to share this other insane new model which helps you edit 3D scenes with simple text. So a few weeks ago I made this video how to edit images just with text prompts. But this paper takes that work to a next level. This paper is called Instruct Nerf to Nerf. Essentially what they're doing is they have these 3D scenes and just by using text instructions they are able to modify these scenes. So for example, put a mustache, change it to a bronze statue, or can change him to Albert Einstein. Now how does it work? So essentially what you do is you take your object and you take 360 images around the object, right? And then for each of the image, you simply apply Instruct pix to pix to modify that image. So for example, here an example of all the key point frames. And then the Instruct Pixtopix model is applied to it to modify them and. Right. And as a result you get some pretty cool results. So let's say this is the original scene that was created from 360 frames, right? And then just by the instruction, so let's say make it autumn, it really changed the whole scene, right? And the results are pretty consistent. So here's another one. Namibian desert. And it looks pretty cool. Here's another one. You can change it to midnight. Change the weather completely. This is pretty nice. Or put a strong over there. And it doesn't look like that. It's not a real video. So. So that's the amazing. Okay, so let's look at another example. So for example, we have a bear, right? So you want to change it to a grizzly bear. Change it to panda bear, polar bear. And this looks pretty great. This video was a little different than what I usually create. So there is not a tutorial. But I wanted to share the state of the art approaches that are happening and nobody else is really covering these because these are not mature products yet. If you like this video and would like to see similar state of the art approaches, please comment and I'll make more of these. Don't forget to like and share the video with your friends. This is definitely an iPhone moment for AI and what a time to be alive. Thanks for watching. See you in the next one.
