Does an AI artist have his own style?

And yet, let’s try to determine where it is better to generate images, at least based on the factors of beauty/relevance to the request.

We will assume that each of the products that participates in our experiments has larger or smaller weights for certain images. That is, it is predisposed to produce in the generation results traces of those images that are either more numerous in the dataset or that have more weight. It is best to call this predisposition with the capacious English word “bias”. This can be compared to the “author’s style” of an artist, although, of course, AI cannot understand the style, this is a human interpretation.

Let's conduct two experiments, try to find bias, which is added to each of your requests, whether you want it or not.

In both experiments, we will compare the results of the same queries from different products. There will be 4 generations in total, no changes between them.
In Midjourney and Designer from Bing (based on DALL-E 3) we will enter the query text in English, in Lumenor.ai and Masterpiece from Yandex – in Russian. For Midjourney v.6 –s 50 is used, for Lumenor – style RAW. DALL-E 3 and Masterpiece do not provide precise settings.

Experiment 1

We set the tokens to be as neutral as possible. Something that you can ask any person to draw on paper with a pencil, and it won’t take much time (after all, you can trace a palm).

Experiment 2

We are trying to confuse the AI ​​by sending a non-standard request. If you ask a person to do the same, he will easily understand you. Unlike AI, as we will see by looking at the results. But we are now evaluating the style of the images.

Can we say that neural network artists have “their own style”?

Each of the four neural networks considered – Midjourney, Designer from Bing (based on DALL-E 3), Lumenor.ai and Masterpiece from Yandex – has its own approach to image generation. This “style” is manifested through the detail of images and through preferences for certain color schemes, which suggests that the AI ​​has a certain “author’s style”.

Midjourney and Lumenor particularly stand out for their tendency to create images with significant bias, which affects detail and color schemes. While the generations from Bing and Masterpiece are less susceptible to these biases, although they can still be distinguished by some characteristics.

Now we are witnessing a race between various txt-2-img products for the attention of users, moving from the stage of competition for creating the most realistic images to a new stage. Products strive not just for realism, but also for aesthetic appeal in their generations, adjusting their models to achieve more beautiful results.

Do you think AI artist bias can be considered a real artistic style, comparable to human creativity? Write in the comments.