Artificial intelligence researchers all over the globe are, at this very moment, beginning to work on systems that can draw pictures by observing the world rather than solely by being programmed to recreate images in their memory from scratch. OpenAI image generator is one of these.
OpenAI and Google
The AI company, OpenAI, isn’t the only organization working on algorithms that create art just by looking at something. Google researchers have published a paper describing a machine learning model that can draw and paint human faces and imagine new faces within a limited set of guidelines.
Imagen is a new image-to-image generator built on large text models that, hang on, that’s too fast. Let’s start with a short explanation. Text-to-image models allow you to take a sentence like “a dog on a bike” and get its corresponding image, something that’s been done for years but recently has seen a massive jump in quality and accessibility.
Scientists have developed techniques to make artificial neural networks that are more advanced than previous ones by using something known as diffusion. It’s a method of refining an image, and top-to-bottom generators. It then determines their accuracy based on whether the picture generated is horrible.)
The other part is improved language understanding through large language models using the motor approach. The specialized aspects of which aren’t the main concern here, but along with many other recent advances, have led to satisfying language models like GPT- 3 and others.
How the OpenAI Image Generator Works
It begins with Imagen generating small images of about 64×64 pixels. It then creates two “super resolution” passes to enlarge it to 1024×1024. This isn’t what we usually know as upscaling. The AI super-resolution produces new details that are harmonious with the smaller image.
This isn’t a new concept, and in fact, artists working with AI models used this fashion formerly to produce much larger pieces than what the AI can handle in one go. However, you end up with a much larger commodity and further intricately detailed. You can indeed do it over and over again, but only if you split it into several pieces.
The advances Google’s experimenters claim with Imagen are plenty. It was announced that textbook models could be used for the encoding part and that their quality is more crucial. That makes sense since a detailed picture of gibberish is undoubtedly worse than a slightly less complicated picture of precisely what you wanted.
Google vs OpenAI Image Generator
Google tested this, and the results show Imagen coming out ahead in tests of the human evaluation, fidelity, and accuracy-wise. This is relative, but to be at par with the perceived quality of DALL-E 2 (a combination of WALL-E and Dali) is remarkable.
DALL-E 2 is a private beta that has users, so it’s more than a research paper. This takes OpenAI ahead of what Google is doing. This is clearly the choice DALL- E 2’s experimenters made. It’s to craft the training dataset earlier and eliminate any content that can violate their own guidelines.
An AI model will only be as good as the data is put on. And not everyone can spend the time and trouble it might take to delete the awful stuff the scrapers pick up as they build multi-million-images or multi-billion-word datasets.
Biases are expected to show up and expose the system processes, which can result in loose testing grounds. But while forgetting systemic discrimination is a lifelong design for many of us, an AI has it easier. Its generators can remove the content that caused it to work poorly in the first place.
And for other tech stories and news, read more here at Owner’s Mag!
