FAQ: Photo Booth and Photo Zone Rentals by Neural Solutions

Our neural network is based on prompts and reference images.

A prompt consists of tokens. Tokens are the basic building blocks of textual input that the model uses to analyze and interpret a prompt. A token can be a word, part of a word, a short combination of words, or even an entire sentence, as well as a punctuation mark or even a space. Tokens at the beginning of a prompt are usually given more attention because the model is trained to recognize the beginning of the text as the most important piece of information. This means that the style, setting, and key objects specified at the beginning can set the tone and direction for the entire image. Initial tokens help set the context for the entire prompt. For example, if you start a query with the phrase "impressionist style," the model will focus on that style during image generation, which will influence the interpretation of all subsequent details.

The placement of theme and style elements at the beginning of the prompt can also influence the composition of the image. The model will tend to place key elements so that they meet the expectations initially set. Placing important attributes and features at the beginning of a prompt can increase the likelihood that they will be correctly perceived and visualized because the model will be better able to recognize and interpret the information presented at the beginning of the text. The tokens in the middle and the last tokens may have less impact on the result, but they are still important to complete and clarify the description. The first character (token) in a prompt line has a weight of 1.0, the further away from the beginning the weight decreases until it reaches 0. This means that at a certain point the text is no longer considered. Therefore, the shorter the prompt, the more accurately the model will render the image.

In addition to placing the token at the beginning, it can also be reinforced by enclosing it in parentheses, for example: (((token))), or by specifying the weight in decimal values, for example: ((token:1.1)), where 1.1 = () - one parenthesis. However, with square brackets, for example: [[[token]]], or by decreasing the coefficient, the strength of the token's weight is similarly reduced, where 1.1=[].

Example:
(token) is an increase of 1.1 times the weight of the token.
((token)) - increase the weight of the token by 1.21 times (= 1.1 * 1.1)
[token] - decrease the weight of the token by 1.1 times
(token:1.5) - increase token weight by 1.5 times
(token:0.25) - reduce the weight of the token by a factor of 4 (= 1 / 0.25) \(token\) - slash is needed so that the enhanced token can be read by the model in conjunction with the previous token.

The reference image in the ID3 method is also a token with a constant weight of 0.7 times. If you add a text token with a higher weight, the text will override the image value. This gives you the flexibility to customize the style by combining text and image. The image should be used with care, because the model does not interpret the objects on it in the textual equivalent: it does not "think" in words, it sees a set of images, which it uses as one big token. Therefore, generations may show unwanted objects from the referent, which will be difficult to remove with words via prompt.