Get the newsletter
Adobe · AI / ML · Filed May 19, 2025 · USPTO published Nov 19, 2026 · This week's release

Adobe Files Patent for AI Images That Carry a Hidden Artist Signature

Ask an AI to paint 'in the style of Van Gogh' and Adobe's filing would slip a hidden code into the picture that points back to the artist. It is a way to give credit to the creators whose styles AI tools imitate.

A computer setup sends an artist-style prompt to cloud servers to produce a generated picture. Drawing from patent filing US 2026/0368946 A1.
A computer setup sends an artist-style prompt to cloud servers to produce a generated picture.All 17 drawings ↓
Publication number
US 2026/0368946 A1(US20260368946A1)
Applicant
ADOBE INC.
Filed · Published
May 19, 2025 · Nov 19, 2026
Inventors
Vishal Asnani, Shruti Agarwal, John Philip Collomosse
Status when we published
Examiner raised objections; the company can respond
Document
20 claims
US classification
345/619
Examiner
SUO, JOSHUA JUNGWOOK (Art Unit 2616)

What Adobe's hidden artist signature does

You type 'a picture of a dog in the style of Van Gogh' into an AI image tool and hit enter. A few seconds later you have a nice painting. What you can't see is whether the artist behind that style gets any credit for it, and today the answer is usually no.

Adobe's filing describes a way to change that. When an artist asks for credit, the system gives that artist a short secret code, a string of 12 or 16 ones and zeros. It then builds the code into the picture while it is being made, so the image still looks like the artist's style but carries an invisible tag.

Later, a second piece of software can read the tag back out and say which artist's style the image was built on. The goal is a trail back to the artist's work, without rebuilding the whole AI each time a new artist signs up.

From the filing · CLAIM 1
… generating, using a concept encoder network, a concept token based on the input prompt and the target information, wherein the concept token represents the image concept and the target information; …

Translation: An AI tool combines your text prompt with hidden artist data into a single digital token.

How a secret code rides inside the image

The filing covers an image generator that takes two inputs: your text prompt and a hidden message the filing calls target information. In the examples, that message is a short sequence of 12 or 16 bits (ones and zeros) tied to one artist or style. A clean, unmarked image uses all zeros.

In the main version (claim 1), a concept encoder network (a small neural network) nudges the internal representation of a word like 'Van Gogh' so it also carries the code. That altered version is called a concept token. It is handed to a latent diffusion model (the type of AI that makes a picture by starting from random static and gradually cleaning it up), which produces an image that depicts the concept and carries the code.

A second version (claim 11) skips the token. A sequence mapper network instead tweaks the starting static itself based on the code. The filing also describes a decoder that reads the bits back out of a finished image, and a way to work out an image's provenance (where it came from) from them.

New artists can be added by fine-tuning only the model's attention layers (the parts that link prompt words to image regions) instead of retraining everything. One prompt can also carry codes for more than one artist.

From the filing · THE ABSTRACT
… an image generation model generates a synthetic image based on the concept token, wherein the synthetic image depicts the image concept and includes an encoding of the target information.

Translation: The system then creates the final picture so that it secretly contains the embedded data.

What this means for artists and Adobe's AI tools

If this idea ever shipped in an image tool, an artist whose style gets imitated could request a hidden tag and later check whether a suspicious image carries it. For you as a user, the picture would look the same, because the tag is invisible. The filing says the mark is built to hold up under non-editorial changes to the image.

It speaks to a real sore spot: the filing itself says artists worry their work is used without recognition or pay. A tag only helps if tools actually add it and someone checks it. The examiner's first review rejected the claims as filed, which is routine and usually followed by amendments.

That makes this Adobe's 51st filing we've tracked since May in our controllable AI image generation watchlist, which already covers work like AI-made fabric and stone textures and styled letters that keep their shape.

Source: US 2026/0368946 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Adobe patent news breaks.

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 · 17 sheets

17 drawing sheets from US 2026/0368946 A1 · click any drawing to enlarge

Sending a prompt to the cloud to make art, drawing sheet 1 from US 2026/0368946 A1
Sheet 1 Sending a prompt to the cloud to make art
Finding the artist name to add a watermark, drawing sheet 2 from US 2026/0368946 A1
Sheet 2 Finding the artist name to add a watermark
Turning a text prompt into an artist-inspired picture, drawing sheet 3 from US 2026/0368946 A1
Sheet 3 Turning a text prompt into an artist-inspired picture
Making pictures that match different requested painting styles, drawing sheet 4 from US 2026/0368946 A1
Sheet 4 Making pictures that match different requested painting styles
Inside the AI system that reads prompts to make art, drawing sheet 5 from US 2026/0368946 A1
Sheet 5 Inside the AI system that reads prompts to make art
Using text guidance to clear static into a finished picture, drawing sheet 6 from US 2026/0368946 A1
Sheet 6 Using text guidance to clear static into a finished picture
An alternative AI setup that predicts noise during generation, drawing sheet 7 from US 2026/0368946 A1
Sheet 7 An alternative AI setup that predicts noise during generation
The image cleanup network that restores detail step by step, drawing sheet 8 from US 2026/0368946 A1
Sheet 8 The image cleanup network that restores detail step by step
Show the other 9 sheets
Turning random static step by step into a clear picture, drawing sheet 9 from US 2026/0368946 A1
Sheet 9 Turning random static step by step into a clear picture
Steps for embedding artist attribution directly into a generated image, drawing sheet 10 from US 2026/0368946 A1
Sheet 10 Steps for embedding artist attribution directly into a generated image
Using a hidden noise pattern to encode artist details, drawing sheet 11 from US 2026/0368946 A1
Sheet 11 Using a hidden noise pattern to encode artist details
Weaving an artist signature into the generated artwork, drawing sheet 12 from US 2026/0368946 A1
Sheet 12 Weaving an artist signature into the generated artwork
The training routine used to prepare the AI model, drawing sheet 13 from US 2026/0368946 A1
Sheet 13 The training routine used to prepare the AI model
Teaching the generator to recreate pictures by comparing stages, drawing sheet 14 from US 2026/0368946 A1
Sheet 14 Teaching the generator to recreate pictures by comparing stages
Hardware components that power the image processing computer, drawing sheet 15 from US 2026/0368946 A1
Sheet 15 Hardware components that power the image processing computer
A server setup that stores and trains the generation model, drawing sheet 16 from US 2026/0368946 A1
Sheet 16 A server setup that stores and trains the generation model
The AI parts that embed and read the hidden signature, drawing sheet 17 from US 2026/0368946 A1
Sheet 17 The AI parts that embed and read the hidden signature
Patent filing page

Watchlists this filing belongs to

All watchlists →
Reader comments

No comments yet

Start the discussion

Real name or a handle, either is fine. Comments are read by a person before they appear, so allow a little time. Keep it about the filing.