This is the video transcript. Read the original article with all the details →
Good morning, welcome back to the ValorosoIT channel, the channel dedicated to vintage computers and electronics. But this time it's not really vintage! We are at Varese RetroComputing 2024 together with Michele Petino.
Nice to meet you, good morning!
And he won't tell us about vintage computers, but he will tell us about something more modern. Let's hear it!
Yes, we are here with a station that is a little different from those you can find at the other stations, as they specifically invited us to show what is the most recent evolution of information technology and technologies related to 3D modeling, rendering and the new frontiers of artificial intelligence.
So, this is your station, right?
Yes. For the occasion we brought a latest generation workstation: it is a 13th generation Intel i9, with a 4090 RTX, GeForce 4090 video card. We chose this card mainly for both its power and the amount of video memory. Having 24 GB of video memory allows us to work better with these algorithms for generating images from text, because they require a lot of space, in fact.
So we're talking about generative artificial intelligence, which is able to create an image starting from a text.
Exactly. For example, in this case we are seeing ComfyUI, with the new engine installed that came out practically in August (2024), which is called Flux. In this case we are using Flux Dev, the FLUX.1 [dev].
You see, this is quite a process... let's say, it's quite a complex flow. In this case we have turned on an LLM Prompt Generator as a starting point.
These flows here, who sets them?
These here can be generated as you want, you can modify them as you want, you see?
Do you edit it, though?
Certainly. In this case I used one made by a Civitai user, which is a site where you can find this type of... both models and...
So we start from a base…
If desired, yes, or even from scratch. After that, this is quite sophisticated because you can set several sources, let's say, of the text to be processed, and then various subsequent functions, such as adding noise to zoom in with more detail, or zooming in, enhancing the face, etc.
In this case we are using a Prompt Generator which is connected to ChatGPT.
Ah, at this point we tell him: Create a prompt in English for generating an image of...
I mean... you don't even want to bother creating a prompt?
Exact. In this case the generation is quite... eh... complex, because he goes to do something complicated as a prompt. And we get this.
When we launch the generation, here we can see a preview of what he is calculating.
In this case, in fact, a crochet animal. When it finishes calculating it - we see it from this green bar at the top that reaches the bottom - we will have the image next to it.
An image that is generated... here it is.
This image clearly doesn't exist, he didn't take it from a database: it was created by the artificial intelligence algorithm, creating it practically from nothing, based on the prompt we entered into it.
Okay. But then, where did the prompt that ChatGPT created go?
The prompt he created is the one he used to do this job, it's this one here, up here. In this case he created it... quite easy, absolutely, exactly.
He didn't bother too much, ChatGPT. Other times, however, it creates much more complex prompts.
Let's now try to create a slightly more complex image. We write in English, this way ChatGPT better understands what it has to do in this case. So you bang a little more than before.
Exact. Then, create a prompt to generate a detailed image of a wizard in the forest, with strange nature, glowing mushrooms, in a fantastic atmosphere. In the background, a detailed castle.
Let's try to see what comes out with this type of description.
Ok, I ran the command.
He is generating.
He created the text, you can tell. That's the whole prompt, so very complex.
Yes, much more than the line he had done before.
Here now it will start to show me the preview of this image.
So, despite the RTX 4090, it takes a little while, I see.
Mostly yes, but it's not so much the video card in this case as the loading of all the necessary parameters and models into the video card's memory.
Ah, well, we're starting to see something...
And here is our image. As you can see, we have glowing mushrooms. The glowing mushrooms! We have our wizard, the castle in the background, and I have to say, overall he did a nice job. We have the mist, the mushrooms are very credible, even if a little unreal...
You asked for bright ones, but let's say if you put them on normal ones... they weren't bright.
Exactly, in fact. Handsome!
And you can create these kind of fantasy images with this or even… Anything. It can be a portrait, a person... he is very good at understanding even the details.
So we can show some images that we have generated previously, for example for this event, in which - as you can see - Varese Retrocomputing 2024 also manages to reproduce the writing well.
He makes mistakes every now and then, but rarely. And let's say that this Flux, compared to Stable Diffusion, for example, has an excellent ability to reproduce writing, but also hands, feet and forks. Exactly, they were the Achilles' heels.
Beautiful images!
Yes, indeed! That many artificial intelligence tools, in fact, when you ask him to generate an image, for him the writing is a little blurred, let's say... it's not clear what he would like to write.
Here is the... the O for retrocomputing, which seems like a strange symbol... but it almost seems intentional.
Oh, indeed. But he probably did it for... who knows what reason.
Ah, only he knows.
Only he knows, exactly.
Well, one of the key points of artificial intelligence is that, all in all, you lose a minimum of control over what you do.
Absolutely yes. I was very impressed when they interviewed the managers of OpenAI, who produce ChatGPT. He was asked: But what exactly happens inside ChatGPT when it generates a response?
And their response was: Honestly, we don't know.
Exact. Because in any case it is self... it self-creates, it self-generates.
These beautiful photos here.
We created these yesterday, for today, to bring something. You also see, Flux's ability to generate images with reflections and very detailed, credible lights is remarkable. Look at the reflections on the metal as they are consistent with, say, the colors and brightness of the scene.
Yes, very beautiful.
Here on display, in addition to these image generation engines with artificial intelligence, a little more linked to what my work is, namely 3D modeling and rendering, I also brought SketchUp with this D5 Render, which is a very powerful rendering engine and which in recent months has been enriched with new features that allow the creation of renders and their improvement with artificial intelligence, to have an even more realistic visual effect.
But for example, now, the spinning blades are rendered in real time?
In this case yes. Everything you're seeing I'm moving in real time. This scene is very, very simple, but with this software - with an adequate machine - you can create very, very complex scenes, even of 4, 5, 6 billion polygons, without too many problems.
Thank you very much then for this presentation.
You're welcome, it was a pleasure.
Why are you laughing?
Did it come out bad? Shit! So let's stop and do it again.
Thank you then for this presentation.
It was a pleasure.
It was a pleasure for me too.
Please, subscribe to the @ValorosoIT channel, activate the notification bell and we'll see you in the next video. HI!