I am very skeptical about neural networks: they are still incapable of producing a number of interconnected independent works. Neural networks are primarily inspired by works like this one that have spread across the web:
And objectively this picture looks very good, but if you look at the team (prompt) that the author used, it will become clear that the neural network did not particularly adhere to his requirements:
The person specified several times that the girl should be depicted in full height, that she should have cat ears and yellow hair, but in the end midjourney gave him an abstract image “on the theme”. And this image is beautiful, but it does not meet the requirements of the “customer”.
This behavior of the neural network makes the work very difficult when you need to generate many characters in the same style, which is a necessary task when developing a game. The idea of creating a crowd of heroes and standing out by their number against the background of other short stories pays off immediately: firstly, it is very difficult to get exactly the desired picture through a neural network, even if you pile up three-story requests and you can always rely only on luck. In addition, the neural network rewards characters with its favorite artifacts.
The eyes suffer the most. Look at this:
I was happy with this portrait, but I couldn’t get decent eyes. My goal was a pleasant-looking result; I didn’t want to leave artifacts and glitches that can be forgiven for a neural network, as we forgive them for children’s drawings: they say, “so what, this green daub doesn’t look like anything at all.”? My son drew it, I’ll hang it on the refrigerator.”.
I really liked this picture without jokes, but now I understand that this crooked daub does not carry any artistic value, and I absolutely don’t care who drew it – be it a person, a neural network, or a cat. She looks bad.
I didn’t want the graphics in the game to have value only because they were drawn by a neural network, so I began to shamelessly edit all the drawings. I understood that the user ultimately doesn’t care who drew this picture: what matters to him is how it looks.
This is an example of the most severe retouching I have done on a character. Most of the other heroes didn’t go through such severe edits, I really got carried away here, but I’m happy with the result.
Another problem was emotions – what www.quacksino.co.uk kind of visual novel is it when characters show the same facial expression?? But any attempts to feed the neural networks of already drawn characters and sign something like “angry” to the promt produced completely absurd results:
This screaming face is very cool, but it doesn’t look at all like the result I wanted. As a result, I again had to try to draw something on top of the result obtained from the neural network:
By the way, I’m not that great of an artist, and the style chosen for Mechahearts is too realistic for my skills. With realism, I wanted to enhance the “uncanny valley” effect and deprive myself of the opportunity to deviate from the intended concept and draw the entire game myself. I’m used to a more cartoonish style and hard brushes; I don’t like realism and stylization like "oil paintings".
During this whole experiment of creating Mechahearts, I came across another almost impossible task: drawing a character from a different angle, and when I needed to draw some characters with their backs, I spent a lot of time in a dismoralizing and confusing search of requests.
In the end, after completely mindlessly burning through traffic and resources on the Midjourney servers in an attempt to depict my pre-retouched maid from behind, I ended up with an image that I could work with:
Neural networks are very inaccurate in details, and as a result, sleeves, cap and hangers disappeared. I had to either draw it all again or collage it from other results:
In the end, it turned out tolerable, in my opinion, but this drawing is still much worse than the one that a person could create.
But the limitations imposed by the neural network do not end there, and sometimes its impenetrable madness forces me to give up and just not even try to generate the scene that I need for the plot. It will be easier for me to rewrite the plot.
I needed a beautiful redhead girl lying on the floor. Along the way, the idea had to be slightly reworked, because neural networks have even worse problems with complex poses than with eyes or hands. A complex pose without broken anatomy can only be achieved by chance, and it is not a fact that this particular pose will be necessary.
Working with a neural network is like shooting fast in the blind – it’s fun, you’re inundated with pictures, but 99% of them are junk that simply can’t be used in the current project. I’m used to developing games alone, without limiting my ideas to other people’s capabilities, but now I have to adapt to the peculiar skills of a neural network. I’m adjusting the plot for her sake. Inventing a story based on possibilities is a normal process – this is how movie scripts are written and games are made, it’s just that this approach is unusual for me personally.
Does using a neural network ultimately make life easier for a developer?? It all depends on his own wishes and skills. My characters would be creepy cross-eyed freaks with three extra fingers if I didn’t have the skill to edit them, and I have this skill because I’ve been drawing for a long time myself, without the help of neural networks.
On the other hand, adjusting a portrait or background is still much faster than drawing it from scratch, so I can devote more time to the plot component. Of course, I could choose a style that is more similar to mine and then adjusting the drawings would not be so difficult, but then I would generally stop understanding why I spend time drawing up three-story queries in midjourney, if I can draw like that myself? The funny thing is to get a result that is as far from my style as possible.
There are no difficulties when working with a neural network. There are inconveniences when working with it.
You can talk a lot about whether neural networks simplify development from a technical point of view or, on the contrary, only confuse you, but you need to put the moral point of view above all, because without the proper passion and love for your project, no matter how technically cool and advanced it is, no game will be completed.
You need to enjoy development so that your brain perceives working on the project as a reward. Otherwise, he will always try to force you to do something else, and creating games will become painful work.
When I draw, even if I do it on the table, I feel that I am developing as an artist. Without those hundreds of drawings that I crumpled up and threw away, I would never have gotten the image I needed. And in this feeling of development lies pleasure: my brain knows that in order to achieve what I want, I must change and improve, and rewards me with dopamine when I do this.
In the case of working with a neural network, I do not develop at all. The wording “prompt-artist” is, IMHO, nonsense, which should create the illusion that you are gaining skills and developing as a specialist while you go through different requests or even repeat the same one 50 times.
But it is the neural network that learns and develops, not the promt-artist.
Selecting a query is a very boring and illogical puzzle, the solution of which does not make you better or smarter, but solves a one-time problem. The experience of this solution is unlikely to be useful next time. Generating pictures is easy physically, but very exhausting mentally, because the feeling that I’m wasting my life and not gaining skills is very vile. This is probably why I take such pleasure in editing these pictures later, because that’s where creativity begins.
And on this good-natured note I end the story about how a short story is made using a neural network. What is the game itself about and what distinguishes it from other novels? I’ll write next time.