An Adobe text-to-3D system that builds a scene you can walk through from a single written prompt is set out in a newly published US patent application. The filing, US 2026/0301328 A1, is titled “Query-Based Generation of Traversable Three-Dimensional Scenes”. The US Patent and Trademark Office published it on 1 October 2026; Adobe filed it on 30 March 2025. It describes a method, not a product, and Adobe has not announced any tool based on it.
Creative Bloq picked up the filing after it was flagged by the patent tracking site Patentlyze. The named inventors are Yiwei Hu, Zhaoyang Zhang, Yannick Hold-Geoffroy, Miloš Hašan and Fujun Luan.
How the Adobe text-to-3D pipeline works
The application splits the job into two steps. First, a generation model reads the prompt and produces what the filing calls a spherical representation: a sequence of 360 degree panoramic frames showing the scene from several viewpoints. In some versions this model is a text-to-video diffusion model that has been fine-tuned on panoramic video.
Second, a reconstruction model, for which the filing gives a long-LRM (large reconstruction model) as an example, turns those panoramas into geometry. The system pulls perspective views out of the panoramic frames, estimates a virtual camera pose for each, and uses them to assemble a traversable scene. The result is placed in a 3D environment where the user can move around and look from different positions.
Adobe argues in the filing that skipping the slow, repeated refinement steps of earlier methods makes the process lighter on computing resources, and that it keeps the scene consistent from one viewpoint to the next. These are the applicant’s claims; the document does not include independent test results.
An opera house as the worked example
The application’s main example is architectural. It imagines a user who wants to preview a design concept in virtual reality and types a request for a grand opera house with ornate balconies and a dimly lit stage. The generated environment would then let that user move between the stage, the seating and the balconies while the architecture, lighting and spatial relationships stay coherent.
The drawings show other prompts too: a village built inside a giant tree, a Greek temple on a sunny cliff, and a room with a fireplace of glowing stones and furniture carved from tree trunks.
A filing, not a launch
A published application is neither a granted patent nor a product roadmap. As Creative Bloq points out, companies file patents for many reasons, including keeping rivals away from an idea. Adobe already sells the Substance 3D apps, but Substance 3D Modeler has no Firefly generative features so far. Other recent Adobe filings cover generating new camera angles from one photo, finding matching materials across 3D models and previewing garments in 3D.
Leave a comment