Adobe has expanded its Firefly AI creative platform with custom model training, access to more than 30 third-party AI models, and a new conversational agentic interface called Project Moonlight.
Adobe reckons that its latest Firefly release is a further step along that path and positions Firefly as an all-in-one AI studio rather than a standalone generation tool. And there are definitely some interesting new features that support its case, with beta versions of Firefly Custom Models and the unveiling of the private beta of the long-anticipated Project Moonlight.
The headline addition, at least for now, is Firefly Custom Models, now in public beta. This lets creators train a model on their own image library to capture a specific visual style, character design, or photographic look. Once trained, the model is reusable across projects and private by default.
Adobe identifies three primary use cases: illustration styles (where consistency of stroke weight and color matters), character design (maintaining a recognizable character across scenes), and photographic looks (repeating a specific visual treatment across multiple images).
There's a really interesting document on how this is all done here: Best practices for creating custom models (beta). Costs are reported to be 500 credits per model with training taking between 30 minutes to a couple of hours, depending on complexity and how many models are being trained at the time.
The idea is that a trained model becomes a foundation for generating new assets that stay on-brand without manual intervention each time. Plus, because it is trained on your own output, you should be protected from future copyright lawyers knocking on your creative door and subpoenaing all your recent work.
In the context of recent news such as the UK government backtracking on its woeful proposition that AI companies could use copyrighted works to train their models with an opt-out option, not to mention ByteDance suspending the rollout of Seedance 2.0, this is starting to feel like an inflection point.
Adobe has long trumpeted its commercially safe AI models. But for those happy to work with mainstream models, Firefly now integrates more than 30 AI models from across the industry.
The idea is that different models have different strengths, whether that be cinematic motion, photorealism, stylised illustration, and so on. Firefly lets you generate with one, refine with another, and compare outputs within a single environment.
Current models supported include Google's Nano Banana 2 and Veo 3.1, Runway's Gen-4.5, Kling 2.5 Turbo, and Adobe's own Firefly Image Model 5, which moves to general availability with this release.
Adobe is currently offering unlimited video and image generations across the model range. Supported plans include Firefly Pro, Firefly Pro Plus, Firefly Premium, 4,000 Credits, 7,000 Credits, 10,000 Credits, 50,000 Credits, with variations on available models and resolutions. Unlimited generations run until April 22. More information here.
Adobe has talked for a while now of evolving generative AI beyond the prompt stage to become something far more nuanced and, dare we say it, genuinely useful to creators who don't want to type an essay to get something that may or may not be what they were looking for.
Project Moonlight is a conversational agentic interface now entering expanded private beta (a sign-up form is available on the Adobe site). Moonlight works across Adobe apps and is designed to take a described intent and turn it into actions using Adobe's tools. The idea is that it functions less like a prompt box and more like a creative collaborator that understands your assets, libraries, and style preferences.
Agentic AI assistants are also rolling out across Photoshop, Express, and Acrobat as part of the same initiative, and it will be interesting to not only chart feedback but see how long it takes them to move from beta to final release.