AI showcase: Generating images – as fast as an arrow on a laptop

This AI showcase demonstrates that amazing and fast results can be achieved with open source AI models with minimal effort on an affordable laptop. A picture is worth a thousand words. That's why the first AI showcase comes from the field of image generation.

Introduction

The series of AI showcases here on the Dr. GDPR blog begins with image generation. The following will be shown:

• Very good results in shortest computing time on own hardware
• Very short programming time, thanks to extremely good open-source frameworks
• Own laptop and not a desktop PC, a server or even a commercial AI model
• Full data control (everything works without internet connection)
• Usage of Open-Source AI Models (here Image Generation, soon: Text Applications, Audio Applications etc.)

Why a laptop? What works very well on a laptop works even better on a server (especially the calculation runs faster then).

Forget your prejudices against Open-Source, which may be justified outside of AI. When it comes to AI, there is no better market than Open-Source. The freely available tools are extremely powerful and overshadow everything that has come before.

The Showcase

To generate images based on a prompt (text input), many know this from DALL-E, Midjourney or also generally from Stable Diffusion.

The image generator is created from scratch. it runs on a local laptop with OS Ubuntu (everything without Microsoft, Apple, Google or Meta). Ubuntu is a popular OS based on Linux. The licensing fee is zero.

The application runs completely without an internet connection after all downloads for setting up the application have been completed. Of course, no service from OpenAI, Microsoft, Google etc. is used. Everything runs locally.

The aim of this showcase

It should be demonstrated that image generation is possible in high quality with high speed on own hardware and entirely without third-party services.

Similarly, it should be shown that programming is possible within minutes. Previously, software projects of this kind took months, if they were even feasible at all. I speak from experience over the past 30 years.

Furthermore, the Showcase should illustrate that hardware below “who can afford it?” can produce results in lightning speed.

For this contribution, a small demo was programmed. It generates 20 images and measures the time for image creation. The prompt “Albert Einstein” was used, plus one of four randomly set styles (no preset, black and white, etc.). Other prompts were also tried.

Time required for the demo

The total time required was less than it took to write this article. Yes, this text was written manually. A future AI showcase will demonstrate that an application for text creation can be programmed in minutes and run on your own hardware. Too bad for Microsoft, Google and co. who will then no longer receive any data.

The time required in detail:

• Download open-source AI model: just a few minutes
• Programming demo application: 5 minutes

That's it. It should be noted that a model like this has several gigabytes of data. When buying hardware, large hard drives are therefore recommended. These hard drives do not cost much. A terabyte is 1024 gigabytes. The aforementioned laptop has a 1.5 Terabyte hard drive. That's big enough to experiment for some time without having to delete unnecessary AI models again.

Results

The quickly programmed demo application generated 20 images in less than 29 seconds. And that was on a laptop. It would have been twice as fast on a desktop PC of comparable quality.

Each picture therefore took approx. 1.45 seconds.

Each image has a resolution of 1024 × 1024 pixels. Even people who are not graphic designers know that this is above “small”.

Here are the results in tile view:

One of these pictures in detail:

Most of the pictures are usable. A few have small errors, such as a pair of glasses that are interrupted. This could be fixed, but was not done for the demo.

The demo also only uses half the possible quality for the generation. The quality is also responsible for the image composition. A higher quality would slightly increase the computing time per image, but would also reduce the above-mentioned spectacle problem.

A still higher generation quality and even higher image resolutions can be programmed just as quickly. The generation speed per image then jumps up to around 7 to 10 seconds (on my laptop, depending on the image resolution).

I personally achieve very high picture resolutions by upscaling. Upscaling means upscaling the resolution. This also takes place with an AI model, which of course runs on my AI laptop. Upscaling can quadruple or even eightfold a resolution, better than any top-class image editing program can achieve in the conventional way.

Advantages

The advantages in brief:

• Any flexibility
• Maximum optimizability
• Own asset instead of sourcing from someone else
• Full data control
• 24/7 operation = same costs
• Opportunity to always benefit from the latest technologies

Alle Bilder in diesem Beitrag wurden von einem Computer-Programm erzeugt. Verwendet wurde das selbst entwickelte KI-System von Dr. DSGVO, ähnlich zu Midjourney. Die Bilder dürfen auf Webseiten frei verwendet werden, mit der Bitte, eine Verlinkung auf diesen Blog zu setzen.