Adobe lets you train Firefly on your own art

Anyone who has tried using generative AI for actual brand work knows the frustration: generating one stunning image is easy, but making ten images with the exact same character and art style has been nearly impossible.

Adobe is finally addressing that consistency headache. On March 19th, the company rolled out Firefly Custom Models into public beta, giving creators and brands the ability to train AI image generators directly on their own artwork.

The feature was originally teased as a private beta during last year's Adobe Max, but now it is open to anyone looking to lock down a repeatable aesthetic.

Locking in styles and characters

Basically, Custom Models act as an in-house visual assistant trained specifically on your portfolio.

Instead of starting from scratch with random text prompts for every new asset, users can feed the system their own illustrations, character sheets, or photography. Adobe says the models can accurately preserve specific stroke weights, color palettes, lighting setups, and character details across future generations.

The goal is to help marketing teams and solo artists pump out high volumes of campaign assets without diluting what makes their brand recognizable.

Crucially, Adobe confirmed that all custom models remain private by default. Whatever proprietary artwork you upload will stay locked to your account and will not be used to train Adobe's general, public-facing Firefly base models.

The guardrails behind the training

Adobe has long marketed Firefly as the commercially safe choice because its base models were trained on licensed stock and public domain works. But allowing users to bring their own data opens the door to potential copyright abuse.

Before you start training, Adobe prompts users to legally certify that they own all necessary rights to the imagery and are not infringing on anyone else's IP or likeness.

Since a simple terms-of-service checkbox rarely stops bad actors, Adobe spokesperson Frankie Tobin told The Verge that the platform also uses automated scanning.

Firefly automatically reads uploaded images for Content Authenticity Initiative credentials. If an artist has embedded tags opting their work out of AI scraping and training, the system will actively block those files from being used.

Why it matters

If you run an agency, manage brand assets, or illustrate for a living, this beta is worth testing immediately. Style drift has been the single biggest bottleneck preventing generative tools from being used in serious production workflows, and this directly targets that friction.

While the open-source community has used LoRA weights to fine-tune image styles for years, risk-averse enterprise clients could never touch them due to copyright liability. Adobe is packaging that same capability inside a familiar, enterprise-friendly sandbox.

That said, we think the copyright safety net still has obvious holes. CAI metadata is far from universal across the internet today. If someone uploads scraped art that simply lacks those credentials, Adobe is essentially relying on that consent pop-up to shift legal liability onto the user. It solves the production consistency problem, but the ethical tightrope is far from gone.