• Bot Eat Brain
  • Posts
  • Generate a music video using AI 🤖😋🧠

Generate a music video using AI 🤖😋🧠

Hey! Ho! What do you know? It's the Bot Eat Brain newsletter, back again with another sampler platter of delicious ai morsels for you to chew on.

Here's what's on the menu:

  • Create an AI Music Video 🎵

  • Protein folding for great good 🧬

  • Byte-sized bonus treats 😋

  • Today's featured AI Art 🎨

Today's vibe: we're gettin' jiggy with it 🕺

A robot with a fishbowl for a head containing a gold fish.

Time to cut a rug.

Create an AI Music Video 🎵

Do you see that shit? Well, it would be even better to listen to it. That's an AI-generated music video, each image is generated by stable diffusion, and each scene change is timed to the thump of the base.

Thanks to Nathan Raw's detailed tutorial and example code your next career as a music video producer is just a few clicks away.

How does it work?

Notice that this isn't just a clean cut between still images or even a clean cut between different AI-generated scenes.

We are smoothly transitioning between scenes and the rate of the transition is timed to the beat of the song. This creates a much more interesting effect than hard cuts, which could be achieved without AI.

The smooth transition between images generated by AI is called interpolation.

Consider this montage of blueberry spaghetti:

This image is an interpolation between two distinct concepts: blueberry and spaghetti.

If you imagine all images which you might label "spaghetti" and all images you might label "blueberries", these images sit right in the middle. If we generate several examples of blueberry spaghetti we can smoothly transition between them by asking the AI to generate intermediary images.

The AI can transition from pure blueberries to pure spaghetti and vice versa, or to any other scene thematically relevant to your music video.

Best of all, we can control the rate of change, this lets us achieve the effect of a smoothly transforming scene that suddenly "jumps" on the beat.

Not a coder? No problem. You can try the demo out in your browser using Google Collab. Hover your mouse over each section of code and click the play button that appears.

You don't need to understand the code to run it, but you will need to create a Hugging Face Account.

Our Predictions 🔮

  • New TikTok Trends as soon as creators get their hands on this. 💃

  • Music videos that blend reality and imagination. Not all images need to be AI-generated for the technique to work. 🤖

  • We'll see many creative combinations of real footage and AI-generated content 🎨

Protein folding for great good 🧬

Proteins are the chemical building blocks of life. In each human cell, there may be anywhere from 10,000 to 20,000 different proteins.

The function of these proteins is largely determined by their shape. For example, the proteins which make up DNA are folded into that famous double-helix structure enabling 6-feet of DNA to neatly fit inside every one of your cells.

Protein folding is exactly what it sounds like, it's how a protein molecule folds on itself to take some unique and potentially useful (or harmful, or neutral) shape.

Knowing the shape a protein takes can help us discover new drugs, understand how diseases function, and unlock the mysteries of our bodies.

The challenge of protein folding is to figure out how a protein will fold up in 3D space based only on the chemical description of the protein.

You probably remember symbols on the right from your Chemistry class (however long ago that might have been). If not, maybe from TV. But just because scientists know what atoms make up a molecule, does not mean they know the shape of that molecule in 3D.

Proteins can consist of millions of sub-structures and individual atoms which makes predicting their shape extremely difficult and computationally intense.

Using stable-diffusion to fold proteins 🤞

Remember how we went from blueberries to spaghetti to generate cool images?

Well, the same concept can be applied to proteins, iteratively moving from unfolded mess to a slightly less unfolded mess in a process that mirrors biology.

Imagine a spring. If you pull the spring taught it will snap back into shape. As long as you don't pull the spring past its breaking point it will pull itself back to its original form.

null

If an unfolded protein existed in nature it would do the same thing. Pulling itself back into its preferred shape. The forces between atoms in the protein molecule act like springs and force the protein into a particular shape.

foldingdiff mimics this natural process, smoothly transforming from unfolded model to the stable shape of a folded protein.

Best of all? There's a live demo. Because it's 2022 and of course you can generate your very own proteins at the click of a button.

What to take away from this? 🤔

  • Advancements in protein folding directly lead to new drugs. 💊

  • This is a cool new technique that can speed up this process. 💨

  • AI is now being applied to some of the hardest questions in science. 🧑‍🔬

Byte-sized bonus treats 😋

Until next time ✌️

🤖😋🧠

P.S. Want to write for Bot Eat Brain? Know someone who'd be perfect to join the writing team? Click here to join our squad.

Today's Featured AI Artist

"determined warrior" by @eye_for_ai