|
Listen to this story
|
Stock image platform Shutterstock has now created its own generative AI service trained on its database of high-quality stock images. The new AI generative tool launched by Shutterstock uses LG AI Research’s EXAONE Atelier—a product emerging from the companies’ partnership since November last year. This algorithm has now been integrated into the platform and can provide high-quality images from simple text prompts.
The AI agent is built on a multi-model architecture for creative applications called EXAONE. The platform has also stated that the agent does not require long and complex prompts to provide good results. In contrast, learning how to prompt other generative models like Stable Diffusion and Midjourney is an art in itself, as seen by the rise of prompt engineering. In addition to this, the platform also has many easy-to-use filters to denote the required style of the output.
Using the relatively simple prompt “Fleet of battleships on a wide ocean”, we were able to obtain the following results from the model.
The tool is built on a massive “proprietary dataset”, which seems to be made up of the 424 million images on Shutterstock’s platform. Contributors to the website can choose to opt out and not allow their images to be included in the database, but those who do not opt out have a chance to be paid for their work. Paul Hennessy, Chief Executive Officer at Shutterstock, stated, “Our tools are built on an ethical approach and on a library of assets that represents the diverse world we live in, and we ensure that the artists whose works contributed to the development of these models are recognized and rewarded.”
As part of their ethical AI initiative, the stock image platform has also promised that contributing artists will be paid for their part in training the model. The Shutterstock Contributor Fund was set up in January last year to ensure that creators whose work was used in Shutterstock’s datasets would be adequately compensated. This approach to responsible AI was lauded by many, especially when Stable Diffusion and Midjourney created controversy for using the LAION-5B dataset which provided no such compensation for artists.
However, the model still has visual anomalies similar to those seen in other AI algorithms, particularly in the case of human features like hands and teeth, albeit to a lesser degree than others. Here’s an example shown below.
In addition to this limitation, free members of the platform can only generate images up to 5 times per day. After this limit, they are prompted to sign up for one of Shutterstock’s premium plans. However, one must also consider that the platform is actively supporting the artists that have contributed their work for the algorithm’s datasets.
Shutterstock have also partnered with AI think tanks like Meta AI, Open AI and LG AI Research in the past. Sam Altman, the CEO of OpenAI, has stated in the past that Shutterstock’s datasets were ‘critical’ to the training of DALL-E. Yann LeCun, VP and Chief AI Scientist at Meta, has also emphasised the importance of Shutterstock’s datasets for training their algorithms.
Their datasets have also been instrumental in training LG’s EXAONE. As part of this partnership, Shutterstock has also used EXAONE for AI-based image tagging, as the algorithm could provide detailed descriptions for the stock images in its library.
This move is representative of the growing dispute regarding the nature of AI-generated artwork, especially when it comes to companies like Midjourney and Stability AI. Recently, a class-action lawsuit has been filed against these companies, seeking compensation for the misuse of artists’ works to train generative AI algorithms. While other platforms, like Getty Images, have taken a hardline stance against generative AI, Shutterstock’s responsible approach to this divisive issue might set a precedent for the future of generative AI.
