Adobe · Filed Feb 28, 2025 · Published Sep 3, 2026 · verified — real USPTO data

Adobe Patents an AI That Splits Image Questions Between Two Specialized Models

Most AI image tools answer questions about photos in one shot, with a single model doing all the thinking. Adobe's new patent splits that job between two specialized AIs, one that plans and one that looks, with the idea that two heads are better than one.

A user interacts with a visual perception system, providing an image and a question, and receiving an answer. Drawing from patent filing US 2026/0260478 A1.
A user interacts with a visual perception system, providing an image and a question, and receiving an answer.
See all 11 drawings from this filing ↓
Publication number US 2026/0260478 A1
Applicant ADOBE INC.
Filing date Feb 28, 2025
Publication date Sep 3, 2026
Inventors Zhehao Zhang, Ryan A. Rossi, Tong Yu, Franck Dernoncourt, Ruiyi Zhang, Jiuxiang Gu, Sungchul Kim, Xiang Chen, Zichao Wang, Nedim Lipka
CPC classification 382/155
Grant likelihood Medium
Examiner TC 4100, DOCKET (Art Unit 4100)
Status Docketed New Case - Ready for Examination (Apr 9, 2025)
Document 20 claims

What Adobe's two-model image analysis actually does

Ever tried to ask an AI to describe exactly what's in a complex chart or a dense document scan, only to get a vague or wrong answer? That's a real limitation of today's image-reading AI: one model handles everything, and it can't always juggle big-picture reasoning with fine-grained visual detail at the same time.

Adobe's patent describes a system with two roles. A coordinator AI reads your question and the image, then decides what kind of visual inspection needs to happen. It passes that instruction to a specialist AI trained to do that specific type of visual work. The specialist reports back, and the coordinator uses that report to write your answer.

Think of it like asking a question at a library. The reference librarian figures out which section you need, sends you to a subject expert, and then explains the answer you bring back. Neither person has to know everything.

From the filing · CLAIM 1
… generating, using an orchestrator vision-language model, an action based on the image and the query; generating, using a vision expert model, a vision analysis result based the image and the action …

Translation: A main AI model decides what to inspect, and a specialized model performs the detailed visual check.

How the orchestrator and vision expert divide the work

The system is built around two distinct model types working in a loop.

The orchestrator vision-language model (a large AI that understands both text and images) receives the image and a natural-language query. Rather than immediately generating an answer, it produces an action, essentially a structured instruction describing what visual analysis needs to be performed to answer the question well.

The vision expert model then takes the image and that action as its inputs. Vision expert models are purpose-built for specific tasks such as object detection, reading text inside images, or identifying regions of interest. The expert runs its specialized analysis and returns a vision analysis result back to the orchestrator.

The orchestrator then uses that result, combined with its own understanding of the original query, to generate the final answer. The claim covers the full loop:

  • Receive image and query
  • Orchestrator generates a task-specific action
  • Vision expert executes that action on the image
  • Orchestrator synthesizes the expert's output into an answer

This architecture allows the system to call on whichever expert is most appropriate for a given question, rather than forcing a single general-purpose model to handle every visual task.

From the filing · THE ABSTRACT
An image perception system then generates an action based on the image and the query using an orchestrator vision-language model.

Translation: The system uses an organizing AI model to figure out what needs to be examined in the picture.

What this means for AI tools that read images

The gap between what people expect AI image tools to do and what they actually deliver is frustrating and consequential. A paralegal uploading a scanned contract, a designer asking about a client's reference image, or a researcher trying to extract data from a chart all need accurate answers, not confident-sounding guesses. A system that decomposes the question before trying to answer it has a better shot at getting the details right.

Adobe's broader push into AI-assisted creative and document tools makes this filing a natural fit. Adobe already ships AI features in Acrobat, Firefly, and Creative Cloud, and a more reliable image-question pipeline would directly improve those products for everyday users who work with complex visuals.

Adobe's third patent we've tracked since July in our AI models working in teams builds on one catching AI misreads and one rewriting vague questions.

Editorial take

When an automated system misreads a number in a contract or misses a clause in a legal document, the cost isn't measured in seconds lost. It's measured in bad decisions made with false confidence.

The problem Adobe is attacking is that image-reading AI tends to blur fine details when it scans large, complex documents, and it has no way to recognize when it's looking at the wrong thing entirely. For anyone relying on these tools to review medical records, financial filings, or dense legal text, that's not a tolerable margin of error.

Splitting the job between one model that decides what to examine and another that actually examines it adds coordination overhead, and this document doesn't measure exactly how much accuracy that buys. But for the people using Adobe's document tools professionally, being correct is the only metric that matters, and the design reflects that priority.

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

11 drawing sheets from US 2026/0260478 A1 · click any drawing to enlarge

Patent filing page

Source. Full patent text and figures from the official USPTO publication PDF.