Tablet menu

How to generate watercolor architectural sketches in a neural network?


Choosing a neural network and compiling a query

I used Blue Willow - a free analogue of Midjourney, but at the initial stage of development, it works in the same Discord. If you look in the "daily theme" section, you can find many interesting results with a given word, every day is something new. It is especially interesting to watch how, day after day, people stick together several past topics into one new one. The "Royal blue" color theme was good, it turned out that it is great for conveying shadows in watercolor. The key in the images I received was: concept art, architecture, anime style, watercolor painting, retro futurism. In the case of the house in the evening light, the golden hour also adds warmth and magic.


I am satisfied with some results, and they are issued quite quickly. Anime girls, all sorts of psychedelics, cats, concept sketches, and abstract photography with smudges of color and smoke are all wonderful. A decent level of architectural rendering is by (Midjourney is much better with this). The very cool style of "isometric diorama" toy images from Midjourney is also missing here (yet?). And this, of course, is a monstrous time waster - there are a lot of crooked options.

If I had not been sitting with a small child, I would not have started to mess with this in my life. Only watercolor sketches left no regrets for the time spent, they can really be used for design inspiration without the nasty feeling that you are plagiarizing. The stream of generated pictures on the chosen topic shakes the imagination a little, it is more effective than Pinterest. The main thing is not to get carried away.

Difficulties with using a neural network

There were no problems, the entire structure and names of chats were copied from there, then everything changed a little, chats in different languages, contests, informative icons in the chat list were added. It is inconvenient that there is no site with a common gallery yet. To evaluate the capabilities of a neural network, you either need to spend a lot of your time on endless variations of requests, or spend the same time scrolling through heaps of chats with ready-made image options in order to copy something from there.

I liked picking up the good parts of the queries from the chats on the topics of the day. In each topic, winners are chosen daily, you can limit yourself to just viewing their work, but this is not very convenient, you need to jump on the links. I am glad that the network is developing, something new often appears there.

Last modified on
back to top