☰
×

Dynamic AI
Co-Creation

A Human-Centered Approach.

Paragraph description....

publisher logo

Chapter 4: AI Images

TREE FORMATION - Image Box

1. Generative Adversarial Networks

latent space

Generative Adversarial Networks (GANs) have emerged as a groundbreaking AI technology providing a window into uncharted realms of visual creativity. At their core lies the concept of latent space - a vast high-dimensional constellation where every point corresponds to a unique potential image waiting to be manifested.

In this latent space, the GAN's generator neural network acts as an artist, interpreting the numerical coordinates and rendering them into visual form through its understanding of the training data it has ingested. Meanwhile, the discriminator network plays the role of an art critic, providing feedback that shapes the generator towards producing outputs that appear increasingly realistic and coherent.

What emerges from this computational call-and-response is a form of synthetic imagination - new visual artifacts that mimic the styles, compositions, and contents present in the GAN's training data while exhibiting novel permutations that could only arise from the mind-bending calculations occurring in the model's internal geometry.

TREE FORMATION - Image Box

Simply perturbing the latent coordinates by tiny amounts can spawn wildly divergent visual results, almost like brushstrokes of an abstract artist. Navigating and interpreting these high-dimensional manifolds reveals an uncanny synthesis of order and chaos, logic and creativity.

However, this powerful generative capability remains constrained by the present limitations of GAN architectures. Phenomena like mode collapse, where the generator fails to fully map the diversity of the data distribution, and training instability stemming from the inherent adversarial min-max optimization can produce artifacts or derail model convergence. BETTER

2. AI Image

GAN's introduces into human image creation a tool that uses language to constrain a dataset that will be a resource for what is a random process. When AI models like DALL-E 2 or Stable Diffusion are used to synthesize visual works based on text prompts, they are effectively remixing the copyrighted training data ingested during the model creation process. In response to prompt the specifies "a Van Gogh landscape painting", the GANs will train on images of Van Gogh paintings available on the web. The image generation will draw on aspects of the dataset, but its built in randomness, in the laten space and in the the give and take of the two neural networks building an image, that it will not likely lead to direct copyright infringment.

As AI image generation tools go mainstream, they have catalyzed numerous legal and ethical quandaries around copyright ownership and the very nature of what constitutes protectable intellectual property in the age of machine intelligence. This has sparked a heated debate around whether such generated outputs constitute infringement of the original image rights holders included in that training data.

There are arguments on both sides - some legal scholars contend that the substantive transformative process of an AI generating a wholly new image from semantic prompts constitutes fair use, much like a search engine displaying thumbnails. Others argue that the commercial interests driving the AI companies necessitate licensing agreements with rights holders for any training data usage.

Beyond copyright, there are thornier philosophical questions around whether AI-generated images can even be granted copyright protections themselves, as they are fundamentally computational and recombinative outputs lacking a clear human author. If responsible AI practices call for transparency in disclosing an image's artificial origins, would that undermine its perceived value or legal status?

Compounding the rights issues further are emerging concerns about generative AI's potential for enabling new forms of misinformation and manipulation. Deepfake images, videos, and multimedia could erode societal trust and truth if this technology is abused and visual information can no longer be treated as objective evidence.

Clearly, rapid evolution is needed around legal frameworks to keep pace as generative AI becomes ever more ubiquitous across creative industries. The decisions made will shape the incentives and liabilities for both human artists and AI developers in the coming decades.

3. Creative Strategies

BOX - first AI work auction

While GANs and diffusion models represent powerful visual creation tools, many artists stress the importance of using AI as a creative assistant rather than treating it as a simple autonomous image generator. There are a variety of key strategies that image creators employ:

  • Image-to-Image Translation: Using an initial image like a rough sketch, painting, or photo as the seed for an AI model can help guide and constrain the generative process into a desired visual style and direction. The human artist maintains control through their original input image.
  • Textual Inversion: This technique allows artists to give semantic meaning to their specific training images, enabling AI models to better understand and recreate the unique contents, styles, and attributes they upload as learning examples.
  • Prompt Engineering: By carefully crafting the text descriptions and instructions provided to an AI image generator, artists can leverage the latent knowledge encoded in the model to synthesize visual concepts grounded in rich contextual cues from language.
  • Iterative Curation: Rather than accepting an AI's first output, artists iterate by feeding the generated image back into the model as a new input, guiding it through a series of refinements and transformations to gradually shape it towards their desired visual goals.

Across all these methods, a key theme is maintaining the human's central creative role in alternately priming, curating, and combining the AI's generative capabilities with their own personal aesthetics and visions. AI becomes a collaborator and tutor that dramatically expands the toolset for ideation and realization while respecting the artist's labors as the orchestrator of the ultimate final artifacts.

Da-le
Midjourneye

4. Tools

The past few years have witnessed an explosive proliferation of platforms and software tools enabling AI-assisted image creation and manipulation:

The rapid pace of development in this space means that incredibly capable new tools are emerging constantly. With many offered through accessible web interfaces or applications, AI-powered visual creation is quickly becoming open to everyone, not just skilled artists and developers.

5. AI Artists

While the AI art tools themselves are impressive technological marvels, equally essential are the pioneering human artists pushing the boundaries of how this tech can expand modes of creative expression and communication. Notable AI artists include:

  • Anna Ridler: Works like "Mosaic Virus" and "Resiliences" critically investigate AI by having it generate images based on datasets containing political, social, and historical concepts and narratives.
  • Mario Klingemann: Coined the term "neurography" to describe his distinct visual style combining machine learning, computer vision, and curated training data to produce haunting, melancholic visuals.
  • Refik Anadol: Creates immersive AI-powered art installations that transform data into mesmerizing media sculptures, almost like machinic hallucinations.
  • Sofia Crespo: Uses AI diffusion models to reflect on biology, art, and the blurring of lines between natural and artificial in works like "Arachnid Manifestations."

Researchers like Ahmed Elgammal are also doing pioneering work using AI techniques to quantify and compare creative influence in artworks across different styles, cultures, and time periods.

As this vanguard of human pioneers continues exploring novel AI techniques, training methodologies, and multidisciplinary use cases, they are cultivating an expansive new paradigm for art that is inseparable from the maturing of artificial intelligence itself.

6. Unit Exercise

To provide firsthand experience with AI's creative potential in visual arts, this unit includes an exercise leveraging widely available tools like DALL-E 2, Midjourney, or Stable Diffusion:

  1. Visual Prompt Exploration: Start by using basic text prompts to generate initial images exploring various artistic styles (realism, surrealism, concept art, etc.). Analyze the outputs, make notes on standout visual elements, unique compositions, stylistic qualities, etc. Refine the prompts iteratively to home in on successfully evoked styles and themes.
  2. Image Transformation: Take the most compelling image(s) generated in Part 1 and use tools to apply transformations - inpainting to extend visual scenes, upscaling, stylization using different mediums, and beyond. Iterate and combine different AI effects to take the image in unexpected and original directions.
  3. AI-Human Iterative Enhancement: Now take the transformed AI output from Part 2 and engage in a process of iterative enhancement. Use your own artistic insights to make adjustments and modifications to the image, then reintroduce these changes into the AI toolset to further refine and develop the piece. This step should involve multiple cycles of human input and AI reprocessing to push the boundaries of what the image can become.
  4. Creative Reflection: Finally, analyze the completed Human + AI collaborative artwork through the lens of your own artistic goals and perspectives. How did the AI's capabilities expand or constrain your creative process? Where did the divergence between machine generation and human artistry become most pronounced? How might you use these tools and techniques differently for future AI-augmented artistic pursuits?

Through this exercise's phases of AI image generation, transformation, combination, and personal embellishment, participants will gain a structured journey through the current AI visual arts toolset. More importantly, they will develop keen insights into the collaborative AI + human dynamics that leading artists are pioneering to produce visionary new creative outputs at the intersection of technology and art.

7. Discussion Questions

As AI image generation becomes increasingly accessible and powerful, these technologies raise profound philosophical, ethical, and cultural questions that will shape the future of visual arts:

8. Bibliography