Canva’s text-to-image tool is the latest example of the technology hitting mainstream apps. | Image: Canva
Australia design app Canva is the latest creative platform to launch a text-to-image AI tool. The company began testing the feature in September and is now rolling it out to the app’s more than 100 million users.
The feature is an implementation of open-source text-to-image model Stable Diffusion, with a few extra safety filters and a custom UI to help steer Canva’s users to get the results they want. Canva, which is available as a free app as well as a paid version with extra features, will give all its users the ability to generate 100 images a day with the tool.
Load up Canva’s text-to-image feature, and you’ll be prompted to “describe the image you want to see,” with a few sample prompts shown as inspiration (e.g., “a light watercolor painting of koi fish in a pond”). You can then choose from various styles (“photo,” “drawing,” “3D,” “painting,” “pattern,” and “concept art”) and the tool will generate a grid of four images to choose from and add to your design canvas. There’s also an option to report images that contain violence, nudity, hate speech, and “biased and/or stereotypical” content.
Image: Canva
The tool generates a grid of four images users can drag and drop into their design.
Image: Canva
Users can select particular styles for the image, from photo or drawing to concept art.
Image: Canva
Users can also report images for containing NSFW material, including nudity, gore, and hate speech.
“We’re treating this as very much a learning experience for our community,” Canva’s co-founder and chief product officer Cameron Adams told The Verge in an interview. “We’re keen to get this technology in front of them because it’s an emergent field, and the exact way it works and how customers will interact with it is still being developed.”
Adams says he’s already seen the tool used for a range of applications. “One of my favorites is students using it to visualize their stories, so they’ll write up a story in English class and use text-to-image to generate an image that matches that story. We’ve also seen it used for images in presentations, on flyers, and on T-shirts, which they can print through Canva.”
The feature is only the latest example of text-to-image AI tools reaching ever larger audiences. The launch of Stable Diffusion in particular has accelerated access to these systems, as its open-source implementation allows companies to integrate it into their own products for free. Text-to-image is quickly becoming a staple in creative platforms, and just last month, Microsoft launched its own text-to-image tool, Microsoft Designer, (powered by OpenAI’s DALL-E system) as part of its Office suite.
The rise of these systems has also sparked some controversy, particularly their use of copyrighted imagery for training data. Many artists have found that their work has been used without their consent to create these commercial products, though the companies and researchers responsible say using this data is covered by provisions like the US fair use doctrine.
When asked about these issues, Adams said, “I think there are legitimate questions about the extent to which AI productions can be considered fair use, and that’s going to differ around the world. We’re keeping a close eye on it, but it’s all still pretty up in the air. We have a great relationship with our contributor community and our users, and we’re working very closely with them to figure out these copyright questions. We give ownership of the images to the users, but we don’t claim that they can be copyrighted by those users.”
Adams and Bhautik Joshi, Canva’s “principal picture and video nerd,” did stress, though, that one key addition they made to the tool was additional filters to stop users from generating NSFW output — particularly important if many users are school children. “We did find that Stable Diffusion’s were fairly easy to circumvent,” said Adams.
Joshi added that the company was “super cognizant that [the output] could be problematic and it’s something we’re actively engaging with — it’s not something we’ll sleep on.”