Definitive Guide to Creating an Effective Prompt in Firefly 3 [2024]

Generative image models have become indispensable tools for graphic artists and designers seeking to reduce manual work and optimize processes through these technologies. There are various generative tools, such as DALL-E, Stable Diffusion, Midjourney, among others. This time, we will explore Firefly, Adobe’s generative tool. We will use version 3, the latest available as of the date (06/02/2024).

Before we begin, it’s important to consider certain principles when drafting a good prompt, as this will determine how closely the generated image aligns with our imagination. It’s crucial that the tool correctly interprets each word we provide.

Therefore, we will establish a basic criterion for structuring the parameters that the tool will use. It’s worth noting that this is the minimum the AI might need; however, this can vary depending on the generative tool we’re using. For example, in Stable Diffusion, it’s possible to add a greater number of parameters, giving the user more control over the tool compared to Firefly. However, Stable Diffusion can be challenging for beginners, as an incorrect parameter can yield undesired results.

Keep in mind that the more complex and lengthy the prompt, the greater the accuracy, but it also increases the risk of errors and confusion for the machine due to the complexity of the provided text.

“The more complex the prompt, the better the results and accuracy will be, but it will also increase the risk of errors in interpreting the context of the prompt.”

A basic template for structuring a good prompt is all we need to start in the right direction

My recommendations for a good prompt are as follows:

• Keep the text structure simple and organized.
• Avoid the use of complex vocabulary that may lead to ambiguity.
• The prompt should provide a clear context of what you want. Establish an order in the writing, prioritizing context, then features, and finally styles.
• Consider it as an interactive process: start from the basics, generating an initial image, and then add more context to refine the result.
• Use English as the provided text, as the tools are optimized for this language.

Taking into account the above, the following image shows a detail of the minimum structure that a good prompt should contain.

Considering these 5 points, you can now create your own prompt with a simple and easy-to-understand structure for the machine. It’s essential that the first point is specified clearly. The context of the image is the frame of reference for the AI to generate it. It can be a place, a person, an object, or a combination of them, according to your needs. Remember that the context will be the main element of the image. For example, if you want an image of a dog walking on the moon, the context would be ‘a dog walking on the moon’. This clearly establishes a specific element in the image: a dog, just one, and where it is, on the moon, which is a specific place. With this, you’ll already have a general idea of what the AI will deliver.

Then, add more details and characteristics that provide further information to the context. For example, ‘A small Chihuahua dog walking on a cheese moon.’ Here, we’ve added more features to the elements of the image, such as the type and size of the dog, as well as the type of moon we need.

Subsequently, you can add information about the graphic style that the tool will use as a reference. For example, ‘Make it an oil painting, digital art, hyperrealism, etc.’ This will depend exclusively on your needs.

Additionally, you can add features such as the tone or color of the image, the camera perspective, or the lighting. It’s important to consider keywords at this point to provide specific details about the desired visual aspect.

Finally, as an optional step, you can indicate the intended use of the image. For example, ‘a book cover, a movie poster, wallpaper, etc.’ This part is optional, as it will not have a significant impact on the final result of the image. You can play with certain parameters and try different options to achieve the desired result.

Below, you’ll find the summarized structure of the prompt containing the elements mentioned above.

This prompt resembles how a human would ask another person to perform this action. However, some words are irrelevant to the machine and only complicate the text, which could lead to unintended errors. It’s crucial to keep it simple and direct.

As you can see in the previous image, we have removed the connectors as they do not provide relevant information for the creation of our image. We have left only the essential, which are the 5 points mentioned, separated by a period (.) to have a clear and direct reading order.

Now that we have understood all the theory, it’s time to take action and start creating our first AI-generated images.

Consider the following image as a starting point.

The image above shows how the elements needed to understand the prompt structure are established. With this information clear, let’s create our prompt, which would be:

"A mountainous landscape at sunset with a serene lake in the foreground. The mountains in the background are snow and the sky displays warm and vibrant colors. Impressionist style, visible brush strokes. For a desktop wallpaper"

The prompt is simple, with a clear structure and steps that the AI considers. We gave relevance to the context and the elements of the image, then added additional information to add detail and the desired styles. With this, let’s go to Firefly.

If we go to the following path: https://firefly.adobe.com/inspire/images we will see the following.

In this section of the application, we will find a text input field where we can insert our first prompt, along with reference images created by others. Let’s add the prompt we created earlier and click on “generate.”

Once Firefly processes the prompt, it will show us the result with 4 image variations, as shown above. By default, Firefly will use the latest processing engine, Firefly Image 3, which has improvements compared to version 2.

Below this, you will find the aspect ratio of the image, defaulting to Square (1:1), but you can change it to landscape (4:3), portrait (3:4), or widescreen (16:9).

After this, there’s the content type, which can vary between Art and Photo, or simply leave this option on automatic for the AI to decide the best option. Note that this option affects the style of the image.

By deselecting the automatic option in Firefly, it automatically selects the “Art” option, which is reflected as a parameter below the prompt. All information handled as parameters by Firefly will be added as keywords.

It’s important to note that these parameters take precedence over the prompt. For example, if in our prompt we specify that the image should have a ‘photo’ style, but the Firefly parameter is set to “art”, the latter will be more relevant and override the “photo” option.

After the content type, we find the “structure” option. This option forces the image to “maintain the structure of the reference elements.” As shown in the example, I added an image from the style gallery containing geometric elements. Additionally, I applied the maximum possible force to compel the AI to incorporate this structure. As you can see in the image, certain geometric elements such as the cube, triangle, and sphere appear.

This option is very useful if we have a reference image and want the AI to reproduce it in a similar manner, obtaining similar results. However, it’s important to consider the context because no matter how much we force the AI, if the prompt deviates too much from the reference, undesired effects may appear.

In the image above, I have inserted another reference to illustrate what I mentioned earlier. The structural reference comes from a photo of a building with a well-defined angle. However, our prompt provides something closer to nature, a landscape of mountains. This can create a conflict between the two models, which could lead to an unexpected interpretation of the prompt.

While these situations can be considered as errors, it’s also important to remember that in such cases, the user’s judgment and creativity come into play to adapt to the situation and achieve the desired result

The next option is styles. Similar to “structure”, styles will force the AI to follow the visual mode of our reference image. In this option, we can adjust the visual intensity and strength.

As seen in the example, I have added a cyberpunk-style image with intense colors for the AI to follow these styles in my final image. As mentioned earlier, please note that these options take precedence over the prompt. Therefore, if you have keywords like “oil style” or “drawing style” in your prompt, but your visual reference style deviates significantly from these keywords, they will be ignored. In that case, you’ll need to refine the prompt and determine if you really need the reference styles from Firefly or if you prefer to include them as keywords within the prompt.

Finally, as the last option, we have effects. Here we can add specific qualities to our image, such as the graphic style of the elements, colors and tones, as well as lights and camera angle.

In the example, I have added a paper appearance to the image format, cool colors, and a specific type of light and camera angle. Additionally, I removed the final part of the prompt to avoid conflicts with these options.

With all this information, you are ready to create your first images in FireFly and start experimenting with the prompt and its structure. Now, let’s go back to our main example and add some variations.

I have slightly varied the prompt to offer a variation while maintaining its context.

“A mountainous landscape at twilight with a tranquil pond in the foreground. The mountains in the background are shrouded in mist, and the sky showcases muted yet colorful tones. Realistic style, emphasizing dramatic lighting. For a desktop wallpaper”

As shown in the example, we now have a completely different image while maintaining its context of mountains. Now the image has a realistic appearance with cooler colors, emphasizing the pond in the center.

It’s all about experimenting with the different options and iterating on the image to achieve the desired result, as this process involves trial and error but you can always have a certain level of control over the final outcome if you know how to handle the prompt correctly. This not only applies to Firefly but you can experiment with other tools in the same way.

Finally, if you like the image, simply click on one of the 4 images to download it.

Additionally, within the options, you can generate variations based on the selected image, apply a structure or style to the image, or use the generative fill, which allows you to add a specific element to the image. As in the example where I added a moon.

Thank you for taking the time to read! If you enjoyed the content, I would greatly appreciate your support with a ‘like’ or by sharing it with others. Your support means a lot to me!

You can also see some of my work in my portfolio here: zanozano.com