Methods
- Semantic Gradient Prompting(SGP)
- Weighted Additive Prompting (WAP)
- Weighted Subtractive Prompting (WSP)
SGP is the simplest of all prompting methods. It's helpful to gain confidence and control over the generation process and gradually modify one feature of the generated image.
Term A, defines the subject of the generated image. The second, named Term B, defines one property of the subject defined by Term A and varies following a semantic gradient of concepts. (for example,Microscopic, Minuscule, Tiny, Small, Medium, Large, Huge, Enormous, Gigantic, Colossal, Immense, Massive, Gargantuan, Titanic, and Monumental). The “seed number” S is fixed to ensure that it does not affect variations in the generated images. The resulting prompt has the following structure:
Creature, Microscopic --seed 99
Creature, Minuscule --seed 99
...
Creature, Immense --seed 99
Creature, Massive --seed 99
...
Creature, Monumental --seed 99
Term A, the black circles represent the various values assigned to Term B, and the black lines connecting them spatially represent the prompts. The purple squares between Term A and Term B represent the generated images (GI). The white line connecting the squares visualises possible interpretation of the type of movement the SGP method allows in the latent space.
WAP is a slightly more complex but potent prompting method that involves the use of weights to influence the operations conducted by the model's transformers. It is helpful to move among concepts in a very controlled manner. The introduction of numerical values for the weights offers several advantages. This method can be used to navigate between any pair of concepts, regardless of their characteristics. In this case, it doesn't matter that the first term is a substantive and the second is an adjective. It also opens the possibility of defining infinite intermediate prompts between two given prompts to explore the transition between two adjacent images carefully.
Term C and Term D, that remain constant for all prompts in each series. Weights are introduced using MidJourney’s semantics ::W where W indicates the term’s importance as a numeric value. In particular, for each prompt, Weight C specifies the importance of Term C and Weight D for Term D. By keeping one of the two Weights constant (Weight C=100) and varying the other (e.g. Weight D=[0, 10, 20, ..., 100]) one can move in a very controlled manner from Term C towards Term D. When Weight D's value reaches 100, it remains fixed and Weight C starts descreasing (e.g. Weight C=[100, 90, 80, ..., 0]). Within each series, the seed value S remains constant. The prompts will be structured as follows:
Term C, the black circle on the right represents Term D, and the black line connecting them spatially represents the prompts regardless of weight values. The purple squares between Term C and Term D represent the generated images (GI). The white line connecting the squares visualises the movement that the WAP technique allowed in the latent space.
WSP is the most unpredictable yet spectacular prompting technique. For some combinations of weight values, it produces images that would be otherwise difficult (or impossible) to conceive and create. The use of numerical values for the weights offers the same advantages acquired in WAP. In addition, it allows the exploration of concepts and ideas that are difficult or impossible to express in natural language, opening the possibility of unveiling unusual visuals and unexplored areas of the latent space.
WSP builds on the controlled gain with the WAP's prompt structure and enhances its flexibility by changing the use of the Weight values. The new prompts consist of the same two terms (Term C and Term D) and their respective weights, Weight C and Weight D. In this case, one weight remains fixed and positive while the other starts at zero and decreases into negative values.
If Weight C - Weight D ≤ 0, MidJourney returns an error and does not proceed with the image generation process. Therefore, it is required for the designed prompts to have the sum between weight values that is higher than zero. To bypass this limitation, prompts will initially have, for example, Weight C=100, while Weight D gradually decreases from 0 into negative values (In this example, Weight D=[-10, -20, -30, ..., -99.99]).
This image schematises the series shown where the WSP method was applied. The empty circle at the left represents Term C, the black circle on the left represents Term D, and the black line connecting the two represents the prompts regardless of weight values. The purple squares on the right side of Term C represent the generated images (GI). The white line connecting the squares visualises the movement that the WSP technique allowed in the latent space. Compared to the movement allowed by the WAP method, this time, the movement proceeds from Term C towards the opposite direction of Term D. The red line on the right represents the limit given by Midjourney v6.0, imposing that the sum of weights values in a prompt must be positive.
