Adobe's Firefly Foundry lets studios train AI on their own IP

Major entertainment companies are building custom generative AI models trained exclusively on their proprietary content, with Adobe providing the infrastructure through a new platform announced at Sundance.

At the Sundance Film Festival last month, directors David Ayer and Jaume Collet-Serra sat down with Adobe engineers to demonstrate something unusual: AI-generated content that studios could actually use without legal anxiety. The demos showed custom models producing franchise-specific imagery, including characters, settings, and props, all trained exclusively on IP the studios already owned.

Adobe is betting that entertainment companies care more about legal certainty than raw capability. Firefly Foundry, unveiled January 22, lets studios create what Adobe calls commercially safe generative models by training them only on content they have clear rights to use.

The partnerships read like a talent agency roster. CAA and UTA have signed on, according to IT Brief Australia. Walt Disney is working with Adobe to build models tuned to what B&T describes as their visual DNA. VFX studios are early adopters, per Mi3's reporting. The timing, announced during Sundance rather than at a tech conference, signals Adobe's intended audience.

Ethically trained appears throughout Adobe's blog announcement, a phrase that does heavy lifting here. In Adobe's framing, ethical means trained on content you own, not scraped from the internet. The definition sidesteps broader questions about AI's impact on artistic labor while addressing studios' immediate concern: litigation risk.

The technical approach differs from typical foundation models. Instead of one massive model trained on everything, Firefly Foundry creates smaller, specialized models for each client. According to Forbes, studios collaborate directly with Adobe engineers to build these bespoke systems. The models remain isolated. Disney's model never sees Warner Bros.' training data, and vice versa.

This isolation serves dual purposes. It protects each studio's IP from competitors while ensuring generated content stays within established visual boundaries. A Marvel model won't accidentally produce DC-style imagery. The constraint becomes the feature.

Adobe is positioning this as workflow integration rather than disruption. The new capabilities announced for Premiere Pro and After Effects, per B&T's reporting, suggest these models will generate assets within existing production pipelines, including concept art, previz, and marketing materials, rather than replacing entire productions.

The IP-safe framing reflects the industry's current anxieties. Studios want AI's efficiency gains without the copyright lawsuits that have plagued other generative AI deployments. They want to move fast without breaking contracts. Adobe is selling them permission to experiment.

But the approach raises its own questions. Training only on proprietary content means these models will excel at reproducing what already exists, generating variations on established franchises rather than creating genuinely new visual languages. It's AI as brand extension tool.

The partnerships also suggest a consolidation of AI capabilities among major studios. Smaller production companies without extensive IP libraries or Adobe partnerships may find themselves at a competitive disadvantage. The democratization narrative around AI tools meets the reality of IP ownership concentration.

Studios can now generate franchise-specific content without scraping the internet or licensing third-party models. Custom models remain isolated, protecting competitive advantages while limiting artistic cross-pollination. Early adopters include major agencies and studios, suggesting industry-wide momentum. The platform integrates with Adobe's existing tools rather than requiring new workflows. Smaller studios without extensive IP libraries may struggle to compete with custom-trained models.

Adobe hasn't disclosed pricing, model performance metrics, or specific release timelines beyond the January announcements. The company's emphasis on high-fidelity generation of video, audio, and 3D assets suggests capabilities beyond current Firefly offerings, though hands-on demonstrations remain limited to select partners. As awards season approaches and production schedules accelerate, we'll see whether commercially safe AI can deliver both the legal protection studios demand and the artistic capabilities they need.