Adobe photo/video editing software has long relied on the help of Sensei AI machine intelligence, and now the company is introducing the next generation of AI features – Firefly generative models that create fonts, images, audio, video and 3D models.
“Generative AI is the next step in the evolution of creativity and productivity, making artist-computer interactions more natural, intuitive and powerful,” said David Wadwani, president of Adobe Digital Media Business.
Firefly is multi-modal and can create audio, video, illustrations and 3D models from text prompts. The first model available now and can only generate images and fonts for nowis trained on “hundreds of millions” of images from the Adobe Stock Photo Catalog and other public domain materials, which virtually guarantees that its use will not lead to litigation (as in the case of StableDiffusion and Getty).
The system interface has a chat window for text prompts with a selection of generated fragments. As soon as the user enters the text, the system will show about a dozen initial sentences. Then you can change the styles and effects of the image or edit the hint until the system gives you what you need.
The quality of the resulting images is almost photorealistic, although hands are not visible in the demonstrations, so it was not possible to count the number of fingers on the hands, and this is a problem that users of other generative neural networks often encountered.
Adobe is considering introducing the ability to add custom portfolios to the model training database, which will allow artists to replicate the product with their own style.
Another feature of the first Firefly model is the creation of customized font effects and the generation of wireframe logos based on scanned drawings and sketches. This is all very cool, but has the potential to put a lot of digital artists out of work if their products are misused. The Adobe Content Authenticity Initiative (CAI) aims to prevent this.
CAI is an attempt to set the rules in a new industry. The proposed working standards govern ethical behavior and transparency in the AI learning process. For example, CAI will create a “do not exercise” tag, which works much the same as the robots.txt exclusion standard. Currently, about 900 organizations around the world, “including media and technology companies, non-governmental and academic organizations” have signed up to the plan.
True, Adobe customers will have to face another headache: figuring out if they really own the rights to the illustrations created by Firefly. In the US, the status of artificial intelligence copyright is not very clearly defined. The Copyright Office so far states that copyright protection “depends on the circumstances” – including “how the AI tool works and how it was used to create the final work.”