AI Art App: Create Pictures from Words

The Rapid Evolution of AI-Generated Visuals

The capabilities of artificial intelligence have been consistently expanding over the last ten years. We’ve witnessed the development of 3D dioramas powered by computer vision, innovative style-transfer techniques, and the proliferation of photorealistic selfie enhancements, face-swaps, and, notably, deepfakes.

Attention-Grabbing Visual Remixing

AI-driven visual remixing has repeatedly demonstrated its ability to capture attention. Maintaining that attention, however, can prove challenging once the initial novelty of an AI-generated effect diminishes. Applications focused on selfie retouching don’t face this issue, as there’s a consistent demand for machine learning as a means of enhancing reality.

Increased Processing Speed

A significant development in AI-enabled synthetic media has been the dramatic increase in the speed of these visual effects, facilitated by increasingly powerful mobile processing hardware.

Completion times for a finished result are now often virtually instantaneous – a pivotal change for productizing, and potentially monetizing, the creativity and power of neural networks and Generative Adversarial Networks (GANs).

From Retouching to Generation

While much of the app-based visual remixing of the past decade has centered on retouching, restyling, or augmenting existing images, rather than pure AI-powered image generation, this is also undergoing a transformation.

Introducing Dream by Wombo

Wombo, a Canadian startup initially known for its AI-powered lipsyncing video app, recently launched Dream (available on iOS and Android). This application leverages AI to create original “artworks” based on text prompts.

Simple Creation Process

The process is remarkably straightforward: users describe the desired image – for example, “A terrifying tree” or “The worst sandwich in history” – select a style from options like Mystical, Baroque, or Fantasy Art, or choose “no style,” and then initiate the creation process.

Within seconds – often less than 20 – the app displays the finished “artwork”.

Witnessing the AI at Work

The creation process itself is visible, offering a glimpse into the AI’s operation. The app showcases the rapid evolution of the image, from initial sketches to the final composition.

Variable Quality of Output

The quality of the generated artworks varies. Some are impressively rendered, while others are less successful.

However, because no two prompts yield the same image, users can repeatedly request new images from the same prompt until they achieve a desired result.

Implications for Artists

In essence, traditional Christmas card artists and pulp fiction illustrators may need to reconsider their careers. Everyone now has the potential to be an “artist”.

However, human artists likely have less to fear. Art created by a human mind and body will likely increase in value as “machine art” becomes more prevalent. (Similar to how each new NFT diminishes the meaning of “digital art”.)

Prompt Complexity and Style

The quality of Dream’s “art” is not consistent. Longer, more complex prompts can be challenging for the AI. The output often leans towards the abstract and distorted, rather than the specific and precise.

It’s generally more adept at depicting fantastical subjects than realistic ones. (A request for a “Madonna and child” resulted in an image resembling a notorious Spanish church restoration failure, rather than a work in the style of Botticelli.)

Speed and Monetization

The speed of production is both impressive and potentially concerning.

Immediately after generating an artwork, the app offers the option to “buy print,” linking to a web shop. This provides a potential revenue stream by transforming a visual trick into a tangible product. “Custom Wombo Dream Print[s]” are available starting at $20 for a matte poster or $45 for a framed print.

The ability to generate $20+ of revenue from approximately 20 seconds of processing time represents a potentially lucrative pipeline.

User Engagement

By the end of last month, users had already generated over 10 million images. (The Google Play app has also surpassed 1 million downloads shortly after launch.)

Virtual vs. Physical Art

Given that most people have limited wall space, many of these AI-generated creations will likely remain virtual. (“AI art” could be ideal for NFTs, however.)

The Value of AI Art

The position of “AI art” within the cultural and fashion value systems remains an open question.

It surpasses clip art and stock photos, and the Dream app’s output can be more engaging than typical art prints found in stores like Ikea. However, the results can also be unsettling, derivative, or simply odd.

What Constitutes Art?

But is it truly art? Or merely a visual output of a mathematical process? Is it an abstraction of human creativity lacking genuine emotion, identity, or soul, as code itself possesses none of these qualities? It simply executes instructions.

And would one genuinely want to display a coded abstraction on their walls?

Perhaps, if it’s particularly aesthetically pleasing. But is that art, or simply wallpaper? Wombo might find greater success selling Dream AI wallpaper, printed mouse mats, or t-shirts (merchandise) rather than “art” prints.

Future Outlook

There are many aspects to consider.

It’s clear that AI-generated art is incredibly enjoyable and engaging – a visual stimulant. It’s also undoubtedly a permanent fixture in the technological landscape. AI models will continue to improve – depending on how “better” is defined in the context of such a subjective field as art.

Generative art models may achieve more successful results by more fully integrating the user into the creative process, providing tools to customize and refine machine outputs to align with their vision, or to create outputs that feel more personally unique and meaningful. In other words, a more hybrid creation process may yield more powerful and moving art-like outputs.

We will also see a proliferation of these artistic AIs, each producing distinct “flavors” and “characters” of visual output, derived from their respective training data. Or, perhaps, art AIs with specialized “styles”. (“Specialisms” may be a more accurate term.)

Numerous other GAN-based image generation AI tools exist, and Pixray’s system is particularly noteworthy (its pixel art outputs are especially charming), although its processing speeds are considerably slower. However, Wombo appears to have been the first to successfully appify and monetize this technology.

The next decade of reality-bending machine learning promises to be a remarkable journey.

https://twitter.com/sacca/status/1473687958075195404