Hoping to attract more enterprise teams to its ecosystem, Adobe has launched a fresh model customization service called Adobe AI Foundry, which will create customized versions of its flagship AI model, Firefly.
Adobe AI Foundry will work with enterprise customers to rebuild and retrain Firefly models customer specific. AI Foundry models differ from custom Firefly models in that Foundry models understand multiple concepts compared to custom models that only contain a single concept. These models will do it also be multimodaloffering a wider range of applications than custom Firefly models, which can only accept images and respond with images.
Adobe AI Foundry models, powered by Firefly, will understand a company’s brand tone, image and video style, products and services, and all its intellectual property. The models will generate content based on this information for any utilize case the company wants.
Hannah Elsakr, vice president of GenAI Up-to-date Business Ventures at Adobe, told VentureBeat that the idea for AI Foundry came about because enterprise customers wanted more sophisticated, custom versions of Firefly. However, given how complicated the needs of enterprises are, Adobe will rebuild the architecture instead of handing over the reins to customers.
“We will be retraining our own commercially safe Firefly models using the enterprise IP. We keep that IP separate. We never move it back to the base model, and the enterprise itself owns those results,” Elsakr said.
Adobe will deploy the Foundry version of Firefly through its Firefly Services API solution.
Elsakr compared AI Foundry to a consulting service in that Adobe will employ teams that work directly with enterprise clients to retrain the model.
Deep tuning
Elsakr calls Foundry a deep tuning method because it goes beyond simply tuning the model.
“The way we think about it, perhaps in more lay terms, is that we are surgically reopening Firefly-based models,” Elsakr said. “This way you can take advantage of all the knowledge in the world with our image model or video model. We go back in time and transfer the intellectual property out of the enterprise, like a brand. It can be shot-style footage, regardless of what the footage is licensed for. Then we retrain. We call it continuous pre-training, where we stress the model to pick certain things differently. So we literally We retrain our base model and that’s why we call it deep tuning, not tuning.”
Part of the training plan includes built-in Adobe teams working with the company to identify the data needed. The data is then securely transferred and processed before being marked. This is passed to the base model and then Adobe starts running the model before training.
Elsakr maintains that Foundry versions of Firefly will not be compact or distilled models. Often, additional data from companies expands Firefly’s parameters.
Adobe AI Foundry’s first two customers were Home Depot and Walt Disney Imagineering, Disney’s theme park research and development arm.
“We are always looking for innovative ways to improve customer service and streamline creative workflows. Adobe’s AI Foundry is an exciting step forward in implementing cutting-edge technologies to deepen customer engagement and deliver impactful content across our digital channels,” said Molly Battin, senior vice president and chief marketing officer at The Home Depot.
More personalization
Businesses often turn to tuning and adapting the model bring immense language models with their extensive external knowledge closer to the needs of the company. Tuning also enables enterprise users to utilize models only in the context of their organization’s data, ensuring that the model does not respond with text that is completely unrelated to the business.
However, most organizations do the tuning themselves. They connect to the model’s API and start retraining it to respond based on ground truth or preference. There are several tuning methods, including some that can be done just with a hint. Other model providers are also trying to make it easier for their customers to tune their models, e.g OpenAI with his o4-mini reasoning model.
Elsakr said she expects some companies to have three versions of Firefly: a Foundry version for most projects, a custom version of Firefly for specific single-concept applications, and a basic version of Firefly because some teams want a model less burdened by enterprise knowledge.