Generative Adversarial Networks (GANs) are a type of artificial intelligence (AI) that can be used to create unique and realistic images. For training, GANs are composed of two components: a generative network and a discriminative network. The generative network creates new images, while the discriminative network evaluates them with respect to a desired reference. The discriminative network trains to recognize authentic images from a mix of generated and authentic images. Since the first generated images are random, this is initially very easy. In the following phase, the discriminative network is frozen, and the generative network learns to mislead it. Once a certain level of successes is achieved, the discriminative network is further trained to better discriminate originals from copies. The whole process is repeated until the generative network outputs images, that are an acceptable substitute for the originals to human supervisors. At this point, any random input to the generative network results in images looking like the images in the training set.
Other techniques are coupled with this basic GAN, to organize the latent space (inputs to the GAN) so that similar images get mapped close to each other. Small input variations therefore result in small changes in the image. In two dimensions, a location has only a few neighbors. The high dimensionality of the latent space means that a very large number of ways to leave a given location coexist, and each one corresponds to a different way of changing the image. Reach out if you want to know more about the software we are using.
Cityscape 3D - click to expand
We designed Cityscape 3D as a banner for our LinkedIn newsletter about business uses of AI. We started with an image created while working on another project. We usually discard generated images, but this one was too good to throw away. It showed a city as seen by an AI, with no natural sense of color. The colors show what's happening in each location. Main streets are covered in vehicle tracks and mist rises from a busy street nearby.
The original image was a square. We used AI to crop it, then AI-driven in-painting filled the empty side bands when we changed its shape to a horizontal banner. If you use free AI image generation websites, you might be disappointed by your results. Seemingly simple jobs like Cityscape 3D usually require more than one try because the algorithm uses random inputs and has many other settings.
If you are interested in seeing more from the Cityscape series, just drop us an email at contact@calcool.ai.
The Power of Mathematics - click to expand
"The Power of Mathematics" is a figurative representation of the creation of the company. It represents a complex, mathematically generated shell sitting in the open in an arid, wintertime californian landscape. I have never set a foot in California, but it is the birth place of all the famous big techs. The exploration started with a list of keywords, often abstract, describing what the company strives to become. Image candidates were generated at half scale using Stable Diffusion, and AI upscaling was used to get full resolution. 27 pictures in total were made in a number of batch jobs, with successive refinements of the prompts to eliminate unwanted characteristics, while keeping the desired ones. In the end, the selection was based on AI-assisted analysis of images, to check whether core values were identified, and on aesthetics.
The following image is another potential option, depicting a business district that is an ideal setting for a tech startup. It brings to mind an incubator, with delicate colorful shells that give rise to unusual plants. Unfortunately, there is an artifact on the floor beneath the closest shell, and the image does not portray some of the puzzles associated with the company's name. The colors are also duller and the image would have required a significant amount of manual work to reach a suitable quality.
"Reboot" is an abstract representation of my mood when Estelle and I decided to sell the flat in Suresnes, and I moved to an old, cold and dated location in Saint Cloud, with a great view on Paris.
Reboot - click to expand
Contact us for custom generative art jobs, and stay tuned if you want to use or buy one of our superfast computers dedicated to graphics and AI.
Affiliate Links:
When you click on an affiliate link to reach the page of a product we recommend on Amazon, we may receive a small commission, and you pay the exact same price.
En tant que Partenaire Amazon, je réalise un bénéfice sur les achats remplissant les conditions requises.