Solving one of the biggest “identity crises” in generative AI, Adobe officially launched Firefly Custom Models into public beta on March 19, 2026. The new feature allows artists, photographers, and brand teams to move beyond “random luck” prompts and instead train a private AI model using their own unique visual assets to ensure total aesthetic consistency.
The Death of “Vibe-Guessing”
Until now, using AI meant hoping the machine guessed your style correctly. With Custom Models, you are the creative director. You can now teach Firefly exactly how you paint, sketch, or light a photograph.
- The Training Recipe: You only need 10 to 30 images to create a custom model. Firefly analyzes these to learn your “visual fingerprint”—from specific brush strokes and color palettes to complex lighting setups.
- Optimized Use Cases: Adobe has specifically tuned the beta for three high-demand areas:
- Illustration Styles: Maintains consistent stroke weight, fills, and textures.
- Consistent Characters: Ensures the same person or creature looks identical across different scenes and poses.
- Photographic Looks: Replicates a specific “look and feel” (e.g., moody noir or high-key product shots) across an entire campaign.
Privacy & Ethics: “Yours and Yours Alone”
In a direct attempt to win back the trust of the creative community, Adobe has built significant guardrails around the training process.
- Private by Default: Models you train are private. Adobe does not use your training images or the resulting outputs to improve its general Firefly model.
- Rights Confirmation: Users must confirm they own the rights to the uploaded images, preventing the “sideloading” of other artists’ work for training.
- Commercial Safety: Like the core Firefly model, Custom Models are designed to be commercially safe, utilizing Adobe’s “Content Credentials” to provide a digital trail of provenance.
| Feature | Details (Beta – March 2026) |
| Training Requirement | 10–30 JPG/PNG images (Min 1024×1024 res) |
| Cost | 500 Generative Credits per training session |
| Availability | Public Beta (Premium Subscribers) |
| Processing Time | ~20–40 minutes per model |
“Project Moonlight” & The Conversational Shift
Accompanying the Custom Models launch is an expanded private beta for Project Moonlight. This marks Adobe’s shift toward “Agentic AI,” where you no longer just type a static prompt.
- The Interface: A turn-by-turn chat interface that lives inside Photoshop, Express, and Acrobat.
- Actionable AI: Instead of just “generating an image,” you can tell the AI assistant, “Take that character I just trained, put her in a forest, and then adjust the lighting to match my Sunscoop reference.” The AI executes the steps autonomously across Adobe’s tools.
More Choice: 30+ Integrated Models
In a surprising move toward “model agnosticism,” Adobe also confirmed that Firefly now offers access to over 30 third-party models in one workspace. For a limited time through April 22, 2026, users can enjoy unlimited generations across this expanded library, which includes:
- Google’s Veo 3.1 & Nano Banana 2
- Runway’s Gen-4.5
- Kling’s 2.5 Turbo
- Adobe’s own Firefly Image Model 5 (now Generally Available)
