6 Tips on How to Query Neural Networks for Great Results
Miscellaneous / / July 14, 2023
Neural networks are capable of creating unique texts and unusual images. The main thing is to ask them correctly!
1. Add Details
Imagine that the neural network is a creative specialist. She, like designers or authors, needs a clear and detailed technical task for high-quality work. The more details you specify, the easier it will be to work, and the result will be closer to what you wanted.
In a request for text, in addition to the basic “write” command, you can add a writing style, the role of the author, and the format of the material. It is also better to open the topic in a few words. For example, the request “Write a post on Telegram on behalf of a travel blogger in an informal style about the most popular sights of Tver" will work more efficiently than "Write a post about sights of Tver.
To expand a request for an image, indicate not only the object, but also its characteristics, the situation in which it is located, the action that it performs, and the preferred color scheme. Let's say, not just a "scientist", but "a young scientist in a white coat in a scientific laboratory works with test tubes."
2. Avoid abstract concepts and denials
It is difficult to predict how the neural network will represent “friendship”, “beauty” or “life in 1,000 years”. Of course, if you're just experimenting and want to test an artificial intelligence fantasy, queries like this will do. But if you are waiting for a picture or text for practical use, it is better to use the most accurate concepts. So, to get an image with a young couple, instead of "love" you can write "a guy and a girl on a date in an amusement park."
Requests with negatives are also unpredictable. Although the neural network is smart, it may not take into account “not”, “without” or “except” and give the opposite result. If the task allows, try to select replacements. For example, instead of "not sitting," use "standing" or "leaning on a table."
3. Let's get some inspiration
Neural networks know many artistic trends, the style features of different writers, artists, photographers and more. If you add the name of a famous cultural figure to the query, you can get an interesting and unusual result. For example, to see how Alexander Pushkin would run his Telegram channel or how Claude Monet would draw Zaryadye Park. This also works with the titles of books, films and paintings.
An important point: if you are preparing an image or text for commercial purposes, be careful - such an experiment can lead to a copyright infringement lawsuit. In the US, artists with similar applied to court. There have not yet been such precedents in Russia, and there is still no theoretical basis for considering such cases, but lawyers suggestthat the situation may change soon.
4. Use English whenever possible
For creating texts, such advice will not be useful, but it will help in working with illustrations. Neural networks are trained in English: even if they later learn other languages, the first one is easier for them to understand. With an English-language query, the result will be more accurate.
If you know English only at a basic level, contact translators with machine learning - they better understand the context and are able to choose the right words. For example, you can try Yandex translate, DeepL or Google Translate. And to find synonyms for adjectives, use the tool loseevery: it will help to replace constructions with the word very (“very”) with others with the same but more precise meaning. The service will offer several options to choose from. For example, to the request very ugly (“very scary”) - unsightly (“ugly”), hideous (“disgusting”) or horrid (“terrifying”).
5. Set frames
If you know exactly what project parameters you need, tell the neural network about it. For texts, specify the character limit, for illustrations - the level of stylization, aspect ratio, resolution or weight (significance) of specific words from the query. The last parameter will help to place accents. Let's say that in a simple query "fat cat" the words are equal in importance - the neural network can give out a slightly chubby cat. And in "fat:: 6 cat:: 5" the adjective has more weight - the animal will turn out to be fatter.
6. Show examples
If you already have texts or illustrations that are suitable in style and format, show them to the neural network. So she will better understand what is asked of her, and immediately do it beautifully.
You don't need to insert the whole object - just share the link. Just check in advance that access to the content is open. In some neural networks that create images, you can stop at this step, in others, you should indicate in the request that the content by link should be used as an example.