Exploring Adobe Firefly as a Tool For Academic Use: Home

Adobe Firefly’s most powerful feature is text-to-image generation: given a text-based prompt, the program outputs a set of images that best portray what the prompt describes.

 

 

With a collection of millions of Adobe stock images and other public domain content, Firefly has a large variety of output to offer. Thus, the prompt is incredibly crucial in determining the quality of the output. There are many facets to writing prompts, and understanding the features of an AI-generated image can help users leverage prompts to produce the best-quality results and maximize the usage of the prompt.  

Consider the simple image of man and dog walking through forest. There are some prompt parameters that can change resulting images (Oppenlaender), such as:

1. Subject terms, such as “forest” or “man and dog in forest.” These compose the base case of the resulting image, arguable essential to any prompt. Firefly responds best to common nouns and will not recognize certain proper nouns (e.g. Mickey Mouse) due to copyright issues. However, places and generic proper nouns such as names will still be included,  but increasing the specificity of proper nouns can diminish the cohesiveness and appearance of the results, with obvious errors more likely.

Man and dog walking through a forest

2. Style modifiers: these can determine the appearance and texture of the image. Common style modifiers are art-related, such as “impressionist” or “oil painting” or “comic book art”

Man and dog walking through a forest oil painting style

3. Repeated words/phrases: these can strengthen the impact of the word/phrase and produce more precise results and emphasize the influence of repeated words/phrases, thus ascribing some degree of weight to them.

While in this case, the effect might not be as evident, notice the clearer definition of the man and dog in the below top result, given the prompt: “a masterpiece artwork of a man and a dog walking through a forest oil painting style with beach. The man and dog are on the beach.” Notice that the beach is more prominent than the forest.

An artwork of a man and dog walking through a forest oil painting style with beach

 

4. Quality prompts/images – These modifiers provide attributes/adjectives such as accomplishment or credibility; an adjective like “award-winning” or “bestselling” to a prompt, could strengthen the results by providing a semblance of a tethering to the real world to the image. The image below, given the prompt “a masterpiece artwork of a man and a dog walking through a forest oil painting style” is more detailed in its rendering of the man and the dog, and the art style of the trees are more evocative of impressionist paintings and provides an artist’s signature in the bottom right corner:

A masterpiece artwork of a man and dog walking through a forest oil painting style 

5. “magic” terms are unexpected phrases/words that add an element of randomness/surprise to a prompt and can affect the resulting images in surprising, possibly more vivid ways. Oppenlaender defines magic terms in two ways: “semantically distant to the main subject of the prompt” or related to non-visual phenomena. Given the prompt “an artwork of a man and a dog walking through a forest oil painting style with beach. The man and dog are on the beach during a sunset. there is a tornado of spirits” with a vague, non-sensory, random modifier “tornado of spirits” yields the following:

an artwork of a man and a dog walking through a forest oil painting style with beach. The man and dog are on the beach during a sunset. there is a tornado of spirits

Consider the second prompt “an artwork of a man and a dog walking through a forest oil painting style with beach. The man and dog are on the beach during a sunset with the flamingos feeling loud and raucous.” Adding the random image of “flamingo” deemphasizes the dog. Other results omitted the dog entirely. In addition, the sensory verb “feeling” ascribes a quality typical of flamingos to the entire image: here, the color pink yields bright pink flowers in addition to the flamingos, which were not mentioned in the prompt.

an artwork of a man and a dog walking through a forest oil painting style with beach. The man and dog are on the beach during a sunset with the flamingos feeling loud and raucous

6. Weights: with some other stable diffusion models, each attribute of a prompt can be assigned a weight. Adobe Firefly has analogous methods that are a little less straightforward but can still tweak the outcome significantly. Methods such as [raw-style=True] which coerces the model to stray from the prompt, [guidance ⊆ {0, 25}], which assigns a numerical value to the prompt, and [avoid={quality}], which causes the model to pay less attention to a specific attribute, are all ways to assign weight to elements in the prompt.

Looking at the previous prompt “an artwork of a man and a dog walking through a forest oil painting style with beach. The man and dog are on the beach during a sunset with the flamingos feeling loud and raucous” with the added element “[avoid=pink]” deemphasizes the pink quality and captures the “loud and raucous” quality by adding more flamingos. The dog is more detailed as well, and more results include it. Avoiding the color pink allows the model to pay attention to other elements in the prompt.

an artwork of a man and a dog walking through a forest oil painting style with beach. The man and dog are on the beach during a sunset with the flamingos feeling loud and raucous [avoid=pink]

 

7. Supplementary elements: Adding and removing supplementary elements (the former with a separate prompt) can be done using Adobe’s generative fill and remove feature. These elements are separate from the qualities of the original prompt unless specified otherwise Consider the same image as above (the flamingos), where another person is added. The generative fill prompt used is “girl in a silk dress fishing oil painting style”

an artwork of a man and a dog walking through a forest oil painting style with beach. The man and dog are on the beach during a sunset with the flamingos feeling loud and raucous [avoid=pink] + girl in a silk dress fishing oil painting style

Adobe’s stockpile of AI tools is growing and continuously updated. With the current set, there are a few ways that Firefly can best be used. 

Currently, the best ways to use Firefly in academic contexts involve templates for single-page advertising media best created through Adobe Express, such as flyers, social media posts, and posters. For example, this could relevant  when creating flyers advertising classes — this can drastically cut the time spent creating posters and advertisements and avoids the issue of using copyrighted images without proper citation. Below are examples of a class poster for an English creative writing class with the background image generated using text-to-template as part of Adobe Express:

 

Here's an example of a social media post advertising a biology course with the prompt “template for a post biology course preview.”

Adobe also uses AI with its Text Effects features, where users can describe what they want certain text to look like. With a prompt such as “furry gold text,” or “moody medieval mossy castle,” users can customize the appearance of text in a variety of ways:

Text prompt: furry gold

Text prompt: moody medieval mossy castle

There are several ways in which templates can be visualized creatively using the text-to-template feature. However, it works best assisting with visual elements where it utilizes Adobe Stock images and other elements to design the results. Templates that feature information and specific written details such as infographics and written reports will inevitably be modified by the user as it does not display accurate information. Currently, text-to-template does not support multi-page media.

 

Adobe Illustrator

Graphics can also be generated with Adobe’s text-to-vector feature, which is available as a part of Adobe Illustrator. A supplemental AI feature is generative recolor, which can change or enhance the color palettes of vector images.

Vector images are composed of vector graphics, made from geometric shapes based on mathematical formulas. A vector graphic file is stored as a series of points that should be connected. The primary advantage of vector graphics is that they are scalable, which means they do not lose resolution when they are resized. 

Once you create a project, you can generate vectors using prompts similar to Firefly’s text-to-image feature. There are four categories of vectors you can generate: subject, scene, icon, and pattern. Below are examples for each of the following.

1. Subject prompt: "Fox wearing a spacesuit patterned with muffins." This limits the boundaries of the image to the subject itself.  

 

As a side note, generative recolor is a feature that can be used to shift the color palette of vector images. Using a sample Adobe prompt “Trippy disco lights,” here is one of the resulting subject images:

Given a custom prompt “thunder gray,” here is one of the results:

 

2. Scene with prompt: "fox in space with planets shaped like muffins in the background." Note that background is included here

3. Icon prompt: "fox in space"

4. Pattern with prompt: "fox in space." This resulting pattern can subsequently fill any shape. 

 

 

Again, this feature can be advantageous in designing graphics that could supplement advertising and presentation in academic environments. It can drastically reduce the time spent on design where more time can be spent on content itself. A trivial example is a history presentation that could benefit from the use of clip art images such as a soldier to attract attention to the contents themselves. As Adobe's training data grows, its results will likely reflect factually correct information in the future.

In summary, vector art, being simpler and more structured than regular graphics, has a more limited range of results but generative AI errors are arguably less evident due to the geometric, “flattened” nature of vector graphics. Using vector images can be a simple and often elegant way to elevate graphic design and can be leveraged in academic contexts to produce clearer, more eye-catching content.

 

With the rise of image-generating AI tools, the issue of copyright remains relevant. With Firefly, those concerns can be dispelled since Adobe claims to train on open source and Adobe stock images and is still able to generate products of the same or higher quality compared with other image-generating AI tools in addition to several (continuously expanding) features that provide greater flexibility. Consequently, this means that Firefly’s output remains limited, since it omits certain proper nouns from prompts, but the importance of this shortcoming is arguably outweighed by its other benefits. Of course, the ethics of even using Adobe Stock to train the model has garnered some controversy, but this cannot refute the fact that it is commercially safer than its peer AI platforms. 

As with any technological tool, it is important to read the terms of use and understand its limitations. For example, using Firefly tools still in beta testing for commercial use is not permitted. Other restrictions include not using Firefly output to train another AI model or generating offensive content, which can include hateful content, graphic violence, or illegal activities. 

https://www.adobe.com/legal/licenses-terms/adobe-gen-ai-user-guidelines.html

Firefly contains a variety of templates for different media. To reiterate an earlier point, Firefly is best used for generating graphics. Firefly cannot populate results with accurate information and thus must be fact-checked and modified by users before results may be disseminated. AI, specifically large-language-models, has been responsible for spreading misinformation. Firefly is predominantly an image-generating tool, but it is important to make sure issues with content are kept to a minimum. Adobe states, “Generative AI features are not intended for professional advice,” which sums it up well. Be sure to read the user guidelines and rules before using Firefly.