Adobe Patents an AI That Builds 3D Scenes From a Text Prompt
Typing a description of a room and getting a fully arranged 3D scene back sounds like science fiction, but it is exactly what Adobe's latest patent filing lays out. The system uses a language model to plan a scene's layout before a second model renders it into geometry.
How Adobe's text-to-3D scene builder actually works
A 3D artist sits down to build a virtual coffee shop from scratch, placing every chair, table, and window by hand. That process can take hours before anything looks right.
Adobe's patent describes a system that short-circuits that process. You type what you want, something like "a cozy coffee shop with a counter near the window and seating in the back," and an AI figures out where everything should go. It first draws up a kind of blueprint, deciding which parts of the space belong to which zone and what objects fit in each zone. A second AI then turns that blueprint into actual 3D geometry.
The two-step approach matters because it separates planning from rendering. The planning model thinks about context (a stool belongs near the counter, not the reading nook), and the rendering model focuses on making objects look right in three dimensions. The result is a 3D scene built from plain language, not from hours of manual placement.
… subdividing the two-dimensional environment into one or more subparts, wherein the one or more subparts represent contextually distinct portions of the two-dimensional environment; and modifying a representation of a subpart of the one or more subparts by adding additional objects within the subpart, wherein the additional objects are contextually related to the subpart; …
Translation: The AI breaks down the layout into logical areas and populates them with relevant objects.
Inside Adobe's two-model scene graph pipeline
The system takes a natural language prompt and runs it through two machine-learning models in sequence.
Model one handles layout. It creates a scene graph, which is essentially a structured map of the scene's objects and how they relate to each other spatially. It places initial objects inside a flat 2D grid that represents a bird's-eye view of the space, then divides that grid into subparts (contextually distinct zones, think "dining area" vs. "kitchen corner"). For each zone, it adds more objects that make sense in that context. A zone labeled as a kitchen corner gets appliances and countertops; a dining area gets chairs and place settings. This zone-filling step is how the system adds density and realism beyond what a single prompt can specify.
Model two takes that scene graph blueprint and renders it into actual three-dimensional geometry, producing the final viewable scene.
Key steps in the pipeline:
- Receive plain-English instructions
- Generate a 2D scene graph with objects and zones
- Populate each zone with contextually appropriate objects
- Render the populated graph into 3D
- Present the finished scene
The separation of planning from rendering is the architectural choice that makes the patent distinct. Each model is specialized, which in theory makes each one easier to train and improve independently.
A large language model (LLM)-based scene generator receives a natural language prompt including instructions to generate a three-dimensional scene.
Translation: The system uses an artificial intelligence model to turn text descriptions into 3D environments.
What this means for 3D artists and game designers
For 3D artists, game designers, and film pre-visualization teams, the tedious part of building a scene is often the blocking phase: deciding where things go before you worry about how they look. A tool that automates that first step could turn hours of rough layout work into minutes, letting creators spend more time on the details that actually require human judgment.
Adobe's steady investment in AI-assisted creative tools fits squarely into the broader push to put generative AI inside products like Substance 3D and Firefly. If this pipeline finds its way into a shipping product, the people most likely to feel it are the ones who currently stare at an empty 3D viewport wondering where to even start.
This is the 21st Adobe filing we've tracked since May on controllable AI images, building on earlier applications like one on layered scene assembly and one on brand style training.
Claim 1 covers quite a bit of territory. It claims the entire pipeline: receive a text prompt, build a 2D scene graph, divide it into contextually distinct zones, fill those zones with related objects, and render the result into 3D. That chain of steps, taken together, is broad enough to describe a large class of text-to-3D systems that use any two-model architecture with an intermediate layout stage.
The specific detail that narrows it is the subdivision step, where the 2D environment is split into subparts whose objects must be "contextually related" to that subpart. That constraint is what separates this claim from a system that just drops random objects into a scene. Whether that subdivision idea is novel enough to survive a prior-art challenge is the real question; text-to-layout research has been active for years across academic and industry groups.
In practice, if this patent were granted as written, it could give Adobe a legal foothold over a generation of text-to-3D tools that use a planning-then-rendering two-step approach with zone-aware object population. That is a meaningful chunk of how people are currently building these systems, which makes this filing one to watch even if the claim eventually gets narrowed during examination.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
8 drawing sheets from US 2026/0268602 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →