Artificial intelligence (AI) image generators are a rapidly advancing technology that allows for the creation of highly realistic images with little human input. These AI-powered tools can generate images of anything from photorealistic landscapes to fantastical creatures and are being used in a wide range of industries, including gaming, film, and advertising. In this article, we will explore the different AI image generators, their uses, and the potential implications of this technology. We also tested some AI-generated images and discuss the future of this technology.
The AI image generators I tested were: DALL-E 2, Midjourney, Shutterstock, Canva, Dream AI, Deep AI, Fotor and Dezgo. There are many more!
How do text-to-image AI art generators work?
Text-to-image AI generators typically use a combination of natural language processing (NLP) and computer vision techniques to generate images from text descriptions. The process typically begins with a user inputting a text description, such as “a plumber fixing a leaking sink.” The AI model then uses NLP to understand the meaning of the text and maps it to a corresponding image.
One of the key techniques used by text-to-image generators is deep learning, which involves training a model on a large dataset of images and their associated text descriptions. This allows the model to learn the relationship between text and image, and to generate new images based on text input.
The models are trained to generate images in two steps, first, a text encoder which will take a textual description and encode it into a feature vector. Next, the image generator takes the feature vector and generates an image.
The generated image can be further improved by an image refinement network that takes the generated image and the text description as input and fine-tunes the image to make it more realistic and coherent with the description.
AI generators are not perfect and the generated images may not be perfect but as the technology advances and more data is used to train the models, the generated images will become more realistic and coherent with the input text.
The time it takes for a text-to-image AI generator to create an image can vary depending on the complexity of the image and the capabilities of the specific model being used.
In general, the time taken will depend on the number of layers in the model, the size of the dataset it was trained on and the computational resources available.
For simple images, the process can take just a few seconds. But for more complex images, it can take several minutes or longer. Additionally, if the AI model is running on a personal computer, it might take a longer time than if it’s running on a powerful server with multiple GPUs.
The images below were created in response to the text: “Captain Cook’s ship Endeavour arriving in New Zealand and being greeted by Māori warriors”. You can see that the AI’s understanding of what “Māori” means is somewhat limited. Some of them entirely avoided any attempt at Māori warriors and instead focussed on better-understood European clothing or South American tribal regalia.
It’s worth noting that the first image generated might not be the final one, depending on the quality of the input text and the model, the user might want to fine-tune the generated image, and this process can take additional time.
Image generation is likely to improve as technology advances and models are trained on larger and more diverse datasets.
How does the AI know what a Dachshund looks like?
One of the tests that I ran was for the text: “A 10-year-old girl with a backpack and a ponytail walking down a post-apocalyptic street with a dachshund”. The results were pretty impressive for this test, although Canva seems to have a rosy picture of a post-apocalyptic world!
The AI model “knows” what a dachshund looks like through the process of training. During training, the AI is exposed to a large dataset of images, along with associated text labels or captions. These captions include information about the objects, animals, and scenes present in the images.
The AI model uses this training data to learn about the visual characteristics of different objects, animals, and scenes, and to develop a representation of what each of these things looks like. In the case of a dachshund, the model would be exposed to many images of dachshunds, along with captions that indicate that the image contains a dachshund. Through this process, the AI learns to recognize the visual characteristics of a dachshund, such as its long body, short legs, and specific facial features.
The more data the model is exposed to, the more accurate and diverse the images it can generate. The AI model is also able to learn from different variations and perspectives of the same object, allowing it to generate images with a high degree of realism.
The AI’s understanding of what a dachshund looks like is based on the data it has been trained on. If the training data is limited, the AI’s understanding will also be limited, and it may not be able to generate images of dachshunds that are highly realistic or diverse.
Is it stealing from existing art?
Text-to-image AI generators do not “steal” existing art in the traditional sense. The images generated by these models are not exact copies of existing art, but rather, they are new images created based on the model’s understanding of the text description provided. In a way, this is exactly what a human artist does when asked to create an image. If they didn’t know who Donal Trump was or what a Unicorn looked like, how could they create an image of “Donald Trump riding a Unicorn”? (Note: This image request was blocked by a couple of the AIs!)
The AI model is trained on a dataset of images, which it uses to learn the relationship between text descriptions and images. When a text description is provided, the model uses this learned relationship to generate a new image that is based on, but not identical to, the images in its training dataset.
It’s also important to note that text-to-image AI generators are not able to reproduce images that are protected by copyright laws. They can only generate new images based on the input text.
It’s possible that the generated images might have a resemblance to existing images, but that’s due to the model’s understanding of the text and its ability to generate images based on the understanding of the text which might be similar to the existing images.
That being said, copyright laws and their application to AI-generated images can be complex and it’s important to be aware of them when using these tools.
Will AI art generators replace human artists?
AI art generators have the potential to automate certain aspects of the art-making process, but it’s unlikely that they will completely replace human artists. While text-to-image generators can generate new images based on text descriptions, they still lack the ability to create art that is truly original and expressive in the way that human artists can.
AI art generators are based on machine learning models that were trained on a large dataset of images and text descriptions, this means that the models can only generate images within the scope of the data they were trained on. Human artists, on the other hand, can draw inspiration from a wide range of sources, including their own experiences and emotions, which allows them to create art that is truly unique and expressive.
Moreover, AI art generators are not able to replicate the personal touch and the creative process that human artist puts into their work. The art generated by AI is based on the understanding of the text and the data it was trained on, and it lacks the human touch, the emotions and the story behind the art that makes it unique and valuable.
In summary, AI art generators can assist human artists in creating art, but they are unlikely to replace them entirely. They can be used as a tool to generate new ideas, speed up certain processes, or generate variations on an existing theme. But the final piece of art will always have a human touch, interpretation, and interpretation that can’t be replicated by a machine.
Additionally, it’s important to note that AI art generators are not able to replace the creative process, the experimentation and the exploration of new techniques and styles that are integral to the art-making process. Human artists are always looking for new ways to express themselves and to push the boundaries of what is possible, which is something that AI art generators cannot do.
In the end, AI art generators can be seen as a tool for artists to use, but it’s not a replacement for human creativity, imagination and the ability to tell stories through art. They can make the process more efficient, but the final product will still depend on the human touch, interpretation and artistic vision.
How will we know which images are AI-generated?
Not quite convincing (Midjourney)
Determining whether an image is generated by AI can be challenging, but there are a few ways to identify whether an image is likely to be AI-generated.
One way to determine if an image is AI-generated is to examine the image for signs of manipulation or unrealistic elements. AI-generated images may contain inconsistencies such as unnatural lighting, unrealistic shadows, or objects that have been added or removed. Additionally, AI-generated images may show a lack of realism in some aspects, such as the textures, lighting, or the overall style of the image.
Another approach is to use reverse image search, which allows you to search for an image using an image instead of keywords. This can help you find the original source of an image and can help you determine if an image has been manipulated or if it’s an AI-generated image.
Another way to check if an image is AI-generated is to look for the watermark or the logo of the company or the website that generated the image if it’s available. Many AI image generators have a watermark on their generated images.
It’s important to note that these methods are not foolproof, and it’s still possible for AI-generated images to be indistinguishable from real images. Therefore, it’s important to approach images with a healthy dose of skepticism, especially if the image is being used to support a claim or to spread information.
As AI technology advances, it’s likely that new methods for identifying AI-generated images will be developed, and it’s important to stay informed about the latest techniques and tools for detecting AI-generated images.