Methodology

Given the unpredictable nature of image generation, evaluating the reliability of various techniques is crucial. To assess the effectiveness of our prompts, we will follow a testing method recommended by Clarinet, which you can explore further in this discussion.

An effective prompt should minimize GPU usage while maximizing output quality. We will execute each prompt four times, categorizing outcomes as follows: - Robust: At least two usable images per grid. - Reliable: At least one usable image per grid. - Marginal: 1 or 2 usable images across four grids. - Inert: No usable images produced.

Experiment 1: Positioning Multiple Texts

For our first experiment, we will utilize an Art Deco-themed source image and aim to place text in each of the four corners. The first attempt will focus on generating an image that retains its aesthetic while successfully placing the text in the specified locations.

Assuming we’ve generated an initial source image, we’ll craft a prompt that specifies the desired text locations: /imagine prompt: <Link to Source Image> Art Deco poster illustration of a lady preparing for an evening event, French, modern. Text "A" at the top right corner, "B" at the top left, "C" at the bottom left, "D" at the bottom right. --ar 2:3 --iw 3 --v 6.0

After evaluating the results of this initial attempt, we found that while some images contained text, none achieved the desired placement in all corners, categorizing the prompt efficiency as Inert.

In our second attempt, we will modify the prompt to guide the AI in replicating an image where the text is already positioned. This adjustment significantly improves our success rate, yielding usable images in 75% of the trials.

Experiment 2: Crafting Longer Text

In our second experiment, we aim to replace single characters with words in the corners of the image. By using longer phrases such as "Apple," "Banana," "Cat," and "Dog," we observe how the prompt changes the generated images.

Our prompt will now read: /imagine prompt: <Link to Image 1-E> Art Deco poster illustration of a lady getting ready for an evening event, French, modern. Text "Apple" at the top right corner, "Banana" at the top left, "Cat" at the bottom left, "Dog" at the bottom right. --ar 2:3 --iw 3 --v 6.0

Despite many images displaying the correct words, none were deemed usable, leading us to conclude that aligning the reference image with the prompt text is essential for success.

Experiment 3: Text’s Influence on Image Composition

In our final experiment, we investigate how different placements and types of text can influence the overall composition of an image. We create a new reference image with the word "Hello" positioned near the subject.

The prompts for this experiment will vary the placement of the text: 1. Near the lady's hand. 2. Inside a mirror reflection.

The results reveal that the text's position indeed affects the image's demeanor, illustrating the power of text in shaping visual narratives.

Key Takeaways

• Image prompting and weight can effectively preserve the original style of a reference image.
• Pre-placing text in specific areas increases the likelihood of successful text placement in new images.
• Matching prompt text to the reference text enhances the probability of generating longer, accurate text strings.
• Text can influence color, elements, and the overall aesthetic of the image, with a notable effect when prompt and reference align.

By understanding these techniques, Midjourney users can enhance their text generation capabilities, unlocking new creative possibilities.