Midjourney and its wealth of styles, both internal and random, is undeniable. Since the appearance of coded styles (sref option) and personal styles (p option), the stylistic explosion has become more than evident. The current trend is to discover styles and use much simpler prompts instead of moving away from this trend and continuing to advance in developing highly detailed descriptive prompts.
The truth is that something is happening within Midjourney, I am referring to its interpretive engine, as straightforward prompts like the one we will see below vary greatly in the results depending on whether a style is applied or not.
This is the case:
Good Morning in the ghetto --s 750A very basic statement of principles. Good morning in the ghetto. One might expect to see perhaps a cup of coffee on a small table in the street, a decaying street, with people wandering around, etc. Anything similar. The atmosphere of a crowded neighborhood. A ghetto, what else, right? The first result generated is as follows:
The four scenes are very similar in terms of perspective. The absence of characters stands out, just a few cars, no people. Are they still sleeping? Maybe. In image 3, that is, the bottom left, we see an abandoned alley, full of graffiti. In summary, despite not seeing anyone on the streets, they could pass as scenes from a ghetto.
What happens when we add style? For example, the author’s current personal style:
Good Morning in the ghetto --s 750 --p 82546naThe change is spectacular:
What happened? The images bear no resemblance to the previous set. In three of them, African American characters appear. Why? Are all the ghettos in the world inhabited by people of color? Image 3 (curiously, it often stands out) reminds us of the overcrowding in cities like Kowloon, true ghettos in themselves. The realism is striking, as is the sharp contrast of light and shadows. By simply applying the personal style, characters, and a much more realistic, even radical, atmosphere emerge out of nowhere.
One might think that the personal style has been trained with images of people of color, beautiful women, etc. But that is not the case. The author’s style is always based on a clear pattern: realistic image versus illustration, well-defined illustration versus low resolution, that is the general pattern. Logos, tasteless images, etc. are usually ignored, making the previous result doubly surprising.
Apart from this consideration, we can see that applying coded styles has its importance. One aspect to keep in mind is that just as we seek character consistency between scenes for developing a comic or story, we also need consistency in style. If we find a personal or coded style (sref) that satisfies us 100%, we can use it in all our generations, giving them our personal touch as image creators (I refrain from saying “as artists”).
Next, we will see the 9 archetypes that we usually use to illustrate style guides, applying the previous personal style:
Food Truck
Pet
Mountain
Muscle car
Tea pot
Jewelry
Portrait
Beach
City
Any of these scenes would be much simpler without the applied style. For example:
Beach
The images are good, but compared to the previous generation with the personal style, they lack strength and drama. We could think that Midjourney’s base style tends toward simplicity, with a mix of pictorial and realistic results.
One way to track the evolution of our style is to generate the same sequence of images every time we finish a training session. That is, after evaluating several hundred images in the corresponding categories—currently three (ranking by aesthetics, prompt accuracy, and niji)—we generate the same prompt using a fixed seed value, allowing us to observe the evolution of the style. Sometimes the change isn’t immediately noticeable, but gradually it shapes what will become our style. Thanks to the personal code system, which is generated every time we evaluate an image, we have the option to use previous personal styles.
Personal style from June 13, 2024:
The style as of today, September 1, 2024:
The prompt is the same; the only thing that changes is the –p code.
The evolution to a more pictorial style with more contrast can be seen. Along the way, many different “styles” have emerged due to a biased evaluation of images. For example, after evaluating solely based on image lighting and not content, we had this style on July 31, 2024:
Quite different, indeed.
The training system is not the same as a LORA in Stable Diffusion. As we see in the news from the Midjourney channel on Discord, a system is being prepared to have different personal styles specifically trained.
