At its annual Summit in Las Vegas, Adobe unveiled Firefly, a "family of creative generative AI models" that promises to redefine the landscape of digital art and design.
In a rapidly evolving landscape where competitors already offer powerful tools for generating images, videos, and sound on the cheap, Adobe is determined to regain its position at the heart of the creative app ecosystem.
Today's beta release of Firefly includes two models, an AI image generator, and a stylized text generator. Eventually, Adobe plans to seamlessly blend these generative AI tools with its iconic creative app suite, such as Photoshop, Illustrator, and Premiere.
Image credit: Adobe
Firefly seeks to go beyond the capabilities of existing AI-powered creative tools like Midjourney, Stability AI's Stable Diffusion, and OpenAI's DALL-E2. What immediately sets Firefly apart is their transparency about the data used to train their AI models. By training Firefly's models on content that is either out of copyright, licensed for training, or part of the Adobe Stock library, the company ensures that artists' work remains respected and the system is brand-safe.
Adobe goes a step further by intending to compensate artists who contribute to the training data. Though not available at the outset, Adobe plans to develop a fair compensation strategy before the beta phase concludes.
Image credit: Adobe
User-friendliness is at the heart of Adobe's AI tools, offering built-in options for art styles, lighting, and aspect ratio. This flexibility allows users to apply these effects to already generated images, providing more creative freedom in the process.
Adobe's future plans for Firefly include AI-generated outpainting in Photoshop, vector variations on hand-drawn sketches in Illustrator, and image prompt-based color grading or restyling in Premiere. One particularly intriguing feature in development allows artists to train the system on their work, enabling AI-assisted content generation in the artist's personal style.
To prevent potential misuse, Adobe is working on a "Do Not Train" system that allows artists to embed the request into an image's metadata. This would prevent training systems from using the image if the creators respect the request. Adobe is in talks with other model creators to establish a no-training standard.
The public beta for Firefly's first two tools is now open, and access is not limited to Creative Cloud subscribers. However, an invite is still required and Adobe will carefully manage the number of users allowed into the beta.