Banner depicting an example AI-generated artwork

AI Apps: Image Generation

Artificial intelligence can create, edit, and manipulate images using machine learning algorithms to generate new images based on a provided prompt. Tasks which historically required knowledge of Adobe Photoshop or other editing tools to complete can now be completed by AI, though results may vary and will likely not be as polished as when completed manually. Like AI chatbots, these tools have strengths and limitations to consider.

Using various AI tools, you can:

  • Create Images: Produce a new image from a prompt (i.e. Create an image of two students sitting in a classroom collaborating on a project).
  • Use Generative Fill: 
    • Inpainting: Add something to a specific part of an existing image (i.e. add clouds to a blue sky).
    • Outpainting: Extend an image (i.e. add a new environment to an existing image).
    • Editing: remove part of an image (i.e. remove someone walking behind you in a photo).

Generate AI versions of a person: After being briefly trained on a handful of photos, artificial intelligence apps can create alternate versions of a person in different settings, wearing different outfits, etc.

An image of a college campus in Spanish Missionary style, produced by Adobe Firefly
An image of a college campus in Spanish Missionary style, produced by Adobe Firefly

When asked to detail the shortcomings of AI Image Generation, ChatGPT (whose parent company also created DALL-E) shared the following:

  1. Limited Control Over Output: While AI can generate images from textual descriptions, users don’t have a significant level of control over the exact outcome. The generated images may not always align with the user’s expectations or needs.
  2. Unpredictability: AI can sometimes produce unexpected or irrelevant results. This unpredictability could be problematic in professional or precise contexts.
  3. Ethical and Legal Concerns: The potential misuse of this technology for creating deepfakes, spreading misinformation, or infringing on copyright or intellectual property rights raises significant ethical and legal concerns.
  4. Lack of Transparency: As with many AI models, the decision-making process behind image generation is not fully transparent, which can make it hard to predict or understand why it produces certain results.
  5. Potential Bias: AI is only as good as the data it was trained on. If the training data contained biases, those biases could be reflected in the images it generates.

Explore the table below to compare a few of the most popular AI image generators/editors available:

Product Subscription Model Distinct Features
DALL-E Limited free credits granted monthly and available for purchase. Created by the same company which created ChatGPT.
Canva AI features available with Canva Pro subscription (or through 1 month free trial) Magic Edit feature can add to, replace aspects, and modify photos as prompted.
Adobe Firefly Available to Montclair State faculty, staff, and students via Adobe Creative Cloud. Trained exclusively on ethical and legal sources.
Adobe Photoshop Available to Montclair State faculty, staff, and students via Adobe Creative Cloud. The Generative Fill tool allows you to select an area of your image and describe what you want added. Can also provide prompts to alter your image’s color scheme, mood, or enlarge images without sacrificing clarity.
Midjourney Free trial with subscription plans available. Most realistic generation of human subjects.

Pedagogical Applications

Consider ways which AI Image Generation could be incorporated to help achieve learning outcomes in your course. Some ideas include:

  • Visual creativity and inspiration: For graphic design and art students, AI can serve as inspiration to help students get started on a project and inspire creativity. It can also serve to brainstorm ideas, visualize a concept or idea before committing to its design, and more.  
  • Analyze trends and biases: Since artificial intelligence is trained on human-produced works and data, it will often contain the same biases present in its dataset. For example, prompting DALL-E to produce an image of “college professors meeting in a classroom” produced an image of five older, mostly Caucasian men. Consider ways you could analyze and learn from these biases in your course.
  • Descriptive writing: Prompt AI to generate similar images within a particular theme or idea, and have students describe and analyze the output to draw conclusions on the decision-making process the AI utilized to determine its output. For composition and creative writing, consider emphasizing literary devices when developing descriptive prompts and critique the visual output with students.