Adobe makes generative AI video model safe for enterprises

Video generation

The Firefly Video Model will power new Adobe Creative Cloud offerings.

They include Generative Extend for video editing, which enables users to extend clips to cover holes in footage, hold shots longer and make transitions smoother. It is now available in beta in Premier Pro.

Also, Adobe is rolling out Text to Video and Image to Video, also powered by the Firefly Video Model, in a limited public beta in the Firefly web app. Text to Video lets video editors generate video from text prompts, access camera controls and use images for B-roll footage generation. Image to Video enables users to transform images into live-action clips.

The Firefly Video Model is an example of fast-growing availability of multimodal capabilities in the generative AI market. More vendors are paying attention to video generation. For example, earlier this year, OpenAI introduced its text-to-video model Sora. On Oct. 4, social media giant Meta introduced Movie Gen, a video model that uses text inputs to generate new videos. It can also edit existing footage or still images.

Reliable training data

A differentiator between the Firefly Video Model and other AI video generators is the data Adobe trains its model on, Futurum Group analyst Keith Kirkpatrick said.

"Adobe does not train its video models on unverified or unsafe content," Kirkpatrick said. "The fact that Adobe doesn't train on unverified or unlicensed content is important for enterprises and others that want to use video for commercial purposes."

This is because enterprises do not want to ingest copyrighted content, which could lead them to infringe on the intellectual property rights of creators, Kirkpatrick continued.

For many enterprises, the main issue is trust and reliability, Constellation Research analyst Liz Miller said.

This is because many CIOs and CISOs are barring the introduction of LLMs and other generative AI tools because they have not been able to vet them for security, compliance and regulatory demands, Miller said.

"There's so many questions when it comes to how these LLMs are trained, " she said. "You're starting to see that nuance really become very important for a lot of enterprises. That's why I think you're starting to see this interesting shift that's happening in the workflow around utilizing these tools."

A model like Firefly Video that has been intentionally trained to be enterprise-ready and commercially friendly is appealing, Miller said.

Limited output

However, providing a model that is trained only on verified information limits Adobe on the data it uses for training. While the vendor has a lot of data, it doesn't use customers' content as part of its training model.

The limit on data might limit the output of what can be created. But that is a tradeoff most organizations will make, Kirkpatrick said.

There are so many questions when it comes to how these LLMs are trained.

Moreover, Adobe remains competitive by providing good results for its customers, Miller said.

"Where they become financially competitive is just how much harder they are down the road in understanding what enterprises want when it comes to AI image and video generation," she said. "If you are a business looking to utilize AI, whether it's image generation or the new generation, you are going to have to ask yourself some tough questions about what you willing to risk in an age where people like to get litigious."

Other than appealing to enterprises, Adobe is extending its reach to consumers and B2C.

"It's another way to penetrate and radiate the user base," Gartner analyst Frances Karamouzis said. "When you make something so easy for B2C customers, they're often a less sophisticated customer, and they're the masses. If you make it so easy and intuitive for people to do it on their own, then you're highly likely to also get them to use it in their enterprise at work."

Esther Ajao is a TechTarget Editorial news writer and podcast host covering artificial intelligence software and systems.