Updated September 7, 2026. This was written on March 28, 2026, and the phrase "this week" below refers to that week, not to now. Two things have moved since. The third party model list has turned over: Kling 3.0 replaced the 2.5 Turbo named here, and Google shipped Gemini Omni 1.1 Flash on August 27. Bigger than that, on August 20, 2026 Adobe made Firefly's audio tools generally available, so Firefly stopped being an image and video catalog and became a full studio. Treat every version number below as a March snapshot. The part that has not aged is the argument about the licensing clause, which is why the piece is still here.
The Wrong Headline
Adobe announced this week that Firefly now integrates over 30 third-party AI models: Google Veo 3.1, Runway Gen-4.5, Kling 2.5 Turbo among them. Every tech outlet covered it as "Adobe adds more AI models."
That's the wrong headline.
The story is Custom Models: Adobe expanded access to Firefly Custom Models, which lets you train a reusable model on your own images and visual style. For independent creators and small production companies, this is the feature that changes the economics. Let me explain why.
What is Adobe Firefly in 2026?
If you are arriving at this page from a search rather than from March, the honest short answer is that Firefly is no longer one thing. It started as Adobe's own image model. Today it is a hub: Adobe's own models for image, video, audio and design, sitting next to third party models from Google, ElevenLabs, Kling AI, Luma AI, OpenAI and Runway, billed through one account.
The change that matters most is recent. On August 20, 2026, Adobe moved Firefly's audio tools out of preview and made them generally available. Three of them: Generate Music, running the Firefly Music Model and producing tracks Adobe licenses for commercial use, Generate Speech, running the Firefly Speech Model with ElevenLabs available as an option inside it, and Generate Sound Effects. That is the whole audio post chain, sitting in the same browser tab as the picture.
I want to be careful about what that does and does not mean, because the coverage flattened it. It does not mean Firefly now beats a dedicated tool at any one of those jobs. I run voice through ElevenLabs directly, and the reason is control, not quality of the model, which is the same model. What consolidation buys you is the thing nobody writes headlines about: one bill, one license position, one place to answer the question a client asks at the worst possible moment, which is whether you can prove you were allowed to use the music.
The current lineup is worth checking on the day you plan around it rather than trusting any article, this one included. Adobe keeps it at the What's new in Adobe Firefly page, and the audio release itself is documented on the Adobe blog post from August 20. Firefly itself is at firefly.adobe.com, and there is a free tier with a daily generation allowance if you want to test the clause question below on your own account before committing a brand library to it.
Does Adobe Firefly Work With Multiple AI Models?
Yes. Firefly is no longer only Adobe's own model, it is a front end to a catalogue of outside models, and you pick which one runs each generation from the same interface.
As of September 2026 the image side includes Google's Gemini Nano Banana family, and the version numbering matters because Adobe lists them separately: Gemini 2.5 as Nano Banana, Gemini 3 as Nano Banana Pro and Gemini 3.1 as Nano Banana 2. Alongside those sit OpenAI's GPT Image and GPT Image 1.5, the FLUX family from Black Forest Labs including FLUX.2 and the Kontext variants, Google's Imagen, Ideogram 3.0 and Runway Gen-4 Image.
On video the catalogue lists Kling 2.5 Turbo and Kling 3.0, Luma's Ray line, and Gemini Omni Flash. Adobe also states plainly that neither it nor its partners train on your personal or project data, which is the sentence to point at when a client asks.
These are not locked inside the Firefly website either. Adobe extends partner models into Photoshop and Express, so the choice follows you into the app where you were already working.
The Honest Part, From Someone Who Runs Those Models Elsewhere
Here is what that list means to me, and it is the reason I have not moved.
Nano Banana Pro, Nano Banana 2, Kling 3.0, GPT Image and Gemini Omni are not names I read in a press release. They are what I generate with every week, through a different hub. So the question is not whether the catalogue is impressive. It is whether the same model, reached through Firefly, behaves like the model I know.
That is the thing nobody can answer for you from a feature page, and it is where I would spend your first hour rather than your first month.
Three specifics worth knowing before you commit a project:
- Version lag is the real risk, not quality. A hub carries the version it has integrated, not the one that shipped yesterday. When a model jumps a generation, native access gets it first. If your work depends on the newest capability the week it lands, a hub will always be behind, and no amount of unified billing fixes that.
- The controls are not always the full set. Models expose their own parameters, and a wrapper exposes the ones it chose to surface. The features I lean on hardest, motion control and holding a character across a whole piece, live in those deeper controls. Check that yours survived the trip before you plan a job around them.
- Unified billing is a genuine win and it is the actual argument. One invoice instead of six accounts is not a small thing for anyone who has explained five subscription lines to an accountant. If your models are already in the catalogue and you do not need bleeding edge versions, that alone can justify the move.
So my answer to whether Firefly works with multiple models is yes, and my answer to whether you should route your work through it is: run your own worst shot through the same model in both places, side by side, on the same afternoon. If the output matches, take the simpler invoice. If it does not, you just learned something a feature list was never going to tell you.
Adobe keeps the current list here, and it changes often enough to be worth rechecking rather than trusting this paragraph in six months: Firefly partner models and the partner models help page.
How do you prompt Adobe Firefly so the output is usable?
The question I see asked most about Firefly is not about features, it is how to prompt it so the result is something you can hand to a client. What works for me is to stop describing the picture and start describing the shot. Lens, camera height, where the light comes from, what the subject is doing. A prompt that says "a product on a table, nice lighting" gives you a different table every time. A prompt that says "35mm, camera at table height, single source from the left, product turned three quarters toward camera" gives you the same table and lets you change one variable at a time when the client asks. The second thing is to stop writing the look into the text once a job passes a handful of shots, and move it into a reference frame instead, because one approved frame holds across a sequence in a way that a repeated sentence never does. Custom Models is the industrial version of that same idea, which is why the clause below matters before you upload anything.
What Custom Models Do
Brand consistency has been the hardest problem in AI-assisted production. You can generate beautiful images. You cannot reliably generate images that look like they belong to the same visual universe (same lighting logic, same color temperature, same character design, same product representation) unless you're manually prompting for every detail, every time.
Custom Models changes this. You feed Adobe a set of reference images (your brand's approved visual library, your product photos, your established color palette) and it trains a model that encodes your aesthetic DNA. Every subsequent generation inherits that DNA without you having to re-specify it in the prompt.
For a brand with an established visual identity, this means: consistency at scale. The 50 pieces of content per month that would otherwise require a creative director reviewing every output can now start from a model that already knows what "on-brand" looks like.
The Production Case for Independent Creators
The feature was previously available to enterprise accounts. Expanding access is what makes this story relevant to anyone who isn't a Fortune 500 brand team.
The practical scenario: I've been working with AI generation tools for a production company that needs consistent brand imagery across 40-50 deliverables per month. The current process involves detailed prompt templates, a style guide document, and manual review at every stage. It works. It's slow.
With Custom Models, the workflow becomes: build the model once from approved reference images, generate at volume, review exceptions rather than every output. That's a fundamentally different labor equation. The creative director's time goes toward what requires creative judgment, not what can be pattern-matched.
The 30 Models Question
On the model catalog expansion: having Google Veo 3.1, Runway Gen-4.5, and Kling 2.5 Turbo inside Firefly is genuinely useful, not because of any single model's capabilities, but because of unified billing and workflow. Right now, serious creative teams are juggling accounts, APIs, and billing relationships across five or six different platforms. Firefly as a hub for multiple models simplifies the operational side.
That said, integration quality will determine whether this is useful or just a checkbox. A model that works well through its native API sometimes degrades when accessed through a third-party wrapper. I'll run comparative tests when I have more time with the new integrations.
Before this goes anywhere near a client account I would need one number Adobe does not publish, which is how long a retrain takes. Today Higgsfield is my hub and Nano Banana Pro is what holds six characters together across a whole piece, and it holds them without a training step, so the day a client repaints a package I am generating again that afternoon. A trained model hands me consistency I do not have to keep asking for, and it also hands me a training cycle every time the brief moves, and the brief always moves late. If a retrain runs overnight this fits inside a real schedule, and if it takes a weekend I keep prompting and absorb the drift.
What This Means for the Independent Creator Economy
I've argued consistently on this site that the most significant AI story for independent creators isn't the headline capabilities. It's the infrastructure maturation. Tools moving from "impressive demo" to "reliable workflow component."
Adobe's Firefly expansion is infrastructure maturation. Not a capability breakthrough. A reliability and accessibility upgrade. Custom Models moving from enterprise-only to broader access is exactly the kind of shift that changes what small teams can produce without enterprise budgets.
The creative director skill (knowing what good looks like, recognizing when an output is off-brand, making judgment calls that require taste) remains the scarce resource. The tools are becoming easier to direct well. That's a good problem to have if you have the taste to direct them.
How I would use Custom Models on a real account
For a brand client at Pichorra, the first thing I would feed the model is not the brand book. It is the rejected work. The dozen frames the creative director shot down in the last campaign for being "off." That negative training set teaches the model the line a brand draws between on and off, which is more useful than the official approved-asset library, because the official library is what survives. The rejects are where the actual taste lives.
Then I would feed the approved set on top. The model trained on both the floor and the ceiling produces output that holds up in client review better than a model trained only on the ceiling.
The honest concern
Custom Models trained on your visual library encode your brand's aesthetic into Adobe's infrastructure. Read the data agreement before uploading. Different tier of subscription, different data handling. For a brand with proprietary visual IP, this is a contract question, not a creative one. Get the legal team in the loop before you train.
Source: Adobe Blog: Firefly expands video and image creation with new AI capabilities and Custom Models
