An introduction to my diagramming practice
- Dean Kenning
Inside the black box of text-to-image AI
Above are our group's reflective notes and understanding of how AI works.
I will sumarise the thoughts presented on my diagram below, expanding on some detail where necessary.
Generative image AI works by training a neural network to generate new images. Here's a simplified explanation of how these generative image synthesis, work:
Generator Network:
The generator is a neural network that takes random noise as input and generates images.
It starts with random noise and learns to map it to the space of realistic images during the training process.
There is an additional 'Discriminator network, which is trained to tell 'real' from generated images.
(Image of generated noise)
Training Process:
The generator and discriminator are trained simultaneously in a competitive manner.
The generator tries to generate images that are indistinguishable from real images, while the discriminator tries to get better at telling real from fake.
Loss Function:
The generator is rewarded when the discriminator is fooled and penalized when it fails to generate realistic images.
The discriminator is trained to correctly classify real and fake images
Feedback Loop:
This process creates a feedback loop where the generator gets better at generating realistic images, and the discriminator becomes more adept at distinguishing real from fake.
This feedback loop does not require human input, eventually resulting in extremely fast renders.
Convergence:
Ideally, this process continues until the generator generates images that are so realistic that the discriminator can't tell the difference.
Generated Images:
The trained generator can then be used to generate new, unseen images by providing it with random noise. However it is 'trained' using a database of images. the software cannot 'generate' new imagery, but only find patterns of arranging noise to fool a trained 'discriminator' network.
It's important to note that the generated images are not truly "creative" in the human sense; the model learns patterns and distributions from the training data and generates new samples based on that knowledge.
Diagramming
allows you or the student to focus in and concentrate on a topic
(Making Sense, 2021. Dean Kenning.)
The Lost diagrams for Walter Benjamin, 2018
Extending ideas from thoughts, to explored conceptual analysis.
Diagramming as an intensive process, connecting neurons and ideas in a new and different way.
Using generative AI to create art.
Task:
Starting with an artwork, Can you create prompts for an AI software to recreate the original image?
Below I will show my progress on this task, through my own use of the software. The task was enjoyable and could be set as a homework task within a school setting, but it would be important to acknowledge that it is AI and not personally developed and also that some of these services require users to pay subscription costs and invest in virtual currencies.
The sleep of reason, produces monsters, Francisco Goya, 1799.
Progress:
Stable diffusion generated images: https://stablediffusionweb.com/#ai-image-generator
(slower response time, but higher quality output.)
Input:
Black and white etching of man sleeping on table while owls tear at his clothes and bats swarm aggressively. The cat watches from the floor as the owls spread their haunting wings above the figure, drowning in distant dreams. Diagonal composition. 1800s style antique art.
This is by far my most successful attempt at recreating the scene. The Stable Diffusion network is much more controllable, however it takes much longer to produce results.
Comentários