Early last year, artificial intelligence company OpenAI unveiled software with the surprising ability to create accurate images from text captions 鈥� even obscure inventions such as 鈥渁n armchair in the shape of an avocado鈥�. The company has now released a new AI model that is聽smaller but聽capable of producing even better results.
Last year鈥檚 program 鈥� called DALL-E 鈥� was a large, 12-billion-parameter AI model that was trained on a huge set of images with associated captions. In recent years, most progress in AI has come from this sort of approach; training them with ever more data on ever larger computers.
However, this makes the AIs expensive, unwieldy and hungry for resources. at De Montfort University in the UK says this approach has been akin to 鈥渂urning cash for research鈥� in recent years.
Advertisement
The new GLIDE model has just 3.5 billion parameters and uses an approach that has come of age in the past year or so, called a diffusion model. The network is still trained using a large set of images, but these images are then gradually and deliberately destroyed by adding noise.
A pristine image has a layer of noise added that degrades it slightly, and then more noise is added, and so on, until the image is pure chaos. The AI, known as a neural network, watches this process and consequently learns how to reverse it. It can then begin with an input that is nothing but noise and slowly work towards a photorealistic image that matches the text description 鈥� effectively un-destroying a new image into existence.
In a test of the software鈥檚 performance, human judges preferred GLIDE鈥檚 images over those from DALL-E 87 per cent of the time in terms of photorealism, and 69 per cent of the time on how well they matched the text input.
at the Georgia Institute of Technology believes that AI and diffusion models such as GLIDE will have a large impact on commercial photo editing. 鈥淧hotoshop will become neural,鈥� he says.
鈥淭hey take a whole bunch of pictures and they progressively add noise, which is just a fancy computer science term for mucking up the pixels,鈥� he says. 鈥淵ou take an image that鈥檚 pristine and clear and you take it all the way down to the point where it鈥檚 completely unrecognisable; [the AI] is in fact learning the opposite, which is taking something that鈥檚 completely unrecognisable and 鈥榬estoring鈥� it back to pristine condition.鈥�
Although each image still takes 15 seconds to create on an A100 graphics processing unit (GPU) that costs upwards of 拢10,000, the work represents an important step forward, says Malekmohamadi. 鈥淚鈥檓 glad to see that this kind of research direction is leading toward a smaller model that could be trained on less powerful GPUs,鈥� he says. 鈥淭hat鈥檚 the ideal situation for the research community and even from an industrial perspective.鈥�
The OpenAI researchers, who weren鈥檛 available for interview, say in their paper that GLIDE can have difficulty producing realistic images for the most complex prompts. In a bid to solve this they added the capability to edit the images the system spits out. Users can ask GLIDE to create an image of 鈥渁 cosy living room鈥� and then select a region of the resultant image and ask for more details to be added, such as 鈥渁 painting of a corgi on the wall above the couch鈥�. Riedl believes that this sort of intuitive creative process will eventually make its way into commercial software.
Reference:
