Texti Newsletter #11

First month of the year is almost out! How's your new year resolution progress going on? Still on the starting point? Well my advice is – stop procrastinating, just do it.

The highlights of the week:

Google Lumiere
Sam Altman to join the AI Hardware race
Tips and tricks how to make the right image prompt

Google Lumiere

TLDR

AI Video generator that can:

Text-2-Video - generates videos based on the input text
Image-2-Video - upload a pic, write a prompt, voila video. Live-photos the way you wanted them to actually be.
Video-2-Video - can modify existing style based on instructions
Fill-the-gap (Video in-painting) - can replace your empty feelings with an virtual girlfriend. Surely not the main intent.

Extended Version

As promised at the beginning of the year, 24 is all about videos.

You all know Google ain't gonna sit and watch how everybody's kicking their ass. They have the best fucking engineers in the world!
The big brother just, launched Lumiere – a project that is about to blow your mind – if it was launched 2 months ago. Videos are slowly becoming the norm. I personally kinda get used to all these visual augmentation that we're seeing on a daily basis.
In December, I would've been praising google as the gods of AI, the impossible is nothing! Today, however I feel like this is what just should've been done already, bruh, what took you so long?

So what was actually announced:

Text-2-Video

Exactly as it sounds, write a prompt and generate a video out of it. Just because I've previously tried Pika, and well to say the least it was something really well experimental rather than actually useful, my hopes aren't really high on this one.

Image-2-Video

This one is new to me and I didn't see this one yet. Upload a pic and make a short video out of it. This is literally live-photos on your iPhone or Motion Photos on your Pixel. I do keep my live photos on at all times, but it's only because it can capture hilarious moments, and add context. AI can't do that – so to me I'm a bit unsure about the potential of this feature.

Video-2-Video

This one I'm most stoked about to be honest with you. I've seen it in practice in a different startup, DomoAI.
The concept is basic again, you upload a video, add a prompt and it'll generate the same video in the same style. Why would this make a world difference? Simple, my main complain about Pika was it wasn't generating enough action or motion in a video. With a reference video this isn't a problem any longer because the motion is natural, AI just adapts the style.

Source

Origami Style

For example this is how Domoai is converting the videos today, and you can surely try it out as well.

Source

Anime Styled

Is it perfect? Not by any means, but hey it's already looking nice. And it managed to preserve all the shapes and things I actually wanted to see.

Let's see how Google's team will show-off their take on this.

Fill-the-gap

This one might be able to help you out and fill in your old pictures and vids where you want your ex removed from the photo - because he fucking sucks and must burn in hell❤️‍🔥!

Here's a little less innocent example provided by google. We can see here how the AI managed to fill in the gap and maintain the motion. Now this - is more of an interesting and useful purposeful usage of AI. Also isn't this dog a cutie 🥰 🥺

Sam Altman to join the AI Hardware race

TLDR:

OpenAI are looking for more hardware for AI, and are considering TSMC (Taiwan Semiconductor Manufacturing) to produce their own branded chips to stay competent in the current market. Follow their stocks it might hit the sky soon.

Not so Extended Version

There is no doubt that the current gold rush is AI, AI is also powered by hardware and it order to make this parrot speak all you need to do is get those powerful chips.

Any big tech company either buys the chips commercially form Nvidia, making em a big pile of cash, either build themselves like Apple and Google does. OpenAI, noticed they have a gap in their business plan, if they continue to expand they're going to run out of free money eventually, but by the time they need to be established on the market. The next logical move is to build their own product, to become fully self sustained and independent and keep pushing the prices they want.

“

To build a great company, you have to be a monopoly and dominate the market.

— Zero to One, Peter Thiel

Which is exactly what OpenAI is slowly crippling towards. Building a monopoly, so that they could control all the aspects of this industry! Well done guys!

Tips and tricks for image prompts

Today's tips and tricks come from Chase Lean, the author My Logo Creator GPT, that I was praising in the last letter, the ice-cream generator.

He wrote a nice twitter thread about this. OpenAI's dall-e works a little more different than the standard prompt generators out there like midjourney.

Avoid negative prompts.
Normally you'd write things like, "avoid fingers", "no blood" to avoid some stuff in your image, OpenAI needs a little different instructions, like make it positive with "family friendly" or "perfect anatomy" etc.
Multiple images
To generate multiple images, all you need to ask Dall-e is "generate one after each other" – this will make gpt work without waiting for extra confirmation
Image size
If you ask chat gpt to generate images in a specific orientation it will also use different image resolutions.
- Square: 1024x1024
- Wide: 1792x1024
- Tall: 1024x1792
Be specific
This is the classical advice, AI's are mimicking people, so exactly like a human, if you say you want food, but don't specify that you want sliced fish, mixed of avocado paste and seeds grown over the duration of months, filtered, cleaned and then boiled at a specific temperature, which will be eventually combine into a dish, your served food will likely not look like sushi. As in real life, want something? Say it!
Avoid imagine and create
Apparently these 2 keywords make the Dall-e confused 🤨. I also didn't know that, woah!
Dall-e modifies your prompt
It seems that your prompt is automagically adjusted by the interpreter, and to avoid that you need to append to your prompt magical instructions:
Use this prompt EXACTLY. DO NOT change or add anything.
Add an image style
Greater results are yielded if you specify which image style you want, Artistic, Vibrant, Minimalistic, etc.
Words in the beginning of a sentence have higher priority than those at the end.
The order matters. Roses are red and violets are blue is not ≠ equal to Violets are blue, and roses are red.
See the example bellow, on the left you have the first prompt, and you a heavy domination of roses, and on the right it feels like violets are dominating the picture, although it very specific.