Get the newsletter
Google · AI / ML · Filed Apr 1, 2025 · USPTO published Oct 1, 2026 · Oct 1 release

Google Patents a System That Teaches AI to Copy Text Styles in Images

Getting AI to reproduce a specific font style, color treatment, or animated text effect accurately requires enormous amounts of labeled training data. Google's new patent describes a way to manufacture that data automatically, at scale.

Examples of images and videos demonstrating various text styles and backgrounds, including animated text. Drawing from patent filing US 2026/0301260 A1.
Examples of images and videos demonstrating various text styles and backgrounds, including animated text.All 22 drawings ↓
Publication number
US 2026/0301260 A1(US20260301260A1)
Applicant
Google LLC
Filed · Published
Apr 1, 2025 · Oct 1, 2026
Inventors
Sourish Chaudhuri, Luciano Sbaiz, Daniel V. Klein
Status when we published
Approved; patent expected soon
Document
20 claims
US classification
345/619
Examiner
CHEN, FRANK S (Art Unit 2611)

What Google's synthetic text-style training actually does

Imagine opening a design app and asking its AI to write a movie-poster title in that same cracked, neon-lit font from the template you uploaded. Today, that kind of AI usually stumbles, producing blurry letters or ignoring the style entirely.

The reason is simple: teaching an AI to handle stylized text requires thousands of example images, each carefully labeled. Collecting those by hand is slow and expensive. Google's patent describes a system that builds that training library automatically, by taking existing styled-text templates and swapping in new words and backgrounds while preserving the original look.

The result is a large dataset of image-and-description pairs, all generated by software rather than human annotators. Another AI then trains on that dataset, learning to produce or recognize stylized text on its own.

From the filing · CLAIM 1
… generating, with a generative language model, a natural language description based on a plurality of tags that comprise data representing a plurality of attributes associated with the graphical text and the animation template; …

Translation: The AI writes out a detailed text description using tags that describe the original image and its text style.

How the pipeline swaps words while locking in the style

The system works in two connected stages Google calls a template description generation pipeline and an instance rendering pipeline.

In the first stage, the system takes an existing animation template (an image or video clip containing styled graphical text) and reads its attributes automatically. Those attributes, things like font weight, color, motion type, and layout, are stored as tags. A generative language model (essentially a text-generation AI similar to the kind behind chatbots) then turns those tags into a plain-English description of the template.

In the second stage, the system picks new words and a new background, then builds an augmented description that combines the original style description with those new choices. A rendering engine uses that description to produce a new image: the new words, drawn in exactly the same stylistic treatment as the original template, placed on the new background.

  • Each output image is paired with its augmented description.
  • Thousands of these pairs form a synthetic training dataset.
  • A separate machine-learning model then trains on that dataset, learning to generate or recognize styled text in new images and videos.
From the filing · THE ABSTRACT
The machine-learned models can utilize the template descriptions associated with the graphical text to generate images and/or videos with graphical text.

Translation: The system then uses these descriptions to create new images and videos featuring the exact same visual text style.

What this means for AI-generated images with readable text

For you as someone who uses AI image or video tools, the payoff is text that actually looks right. Right now, asking most AI image generators to place styled words into a scene often produces garbled letters or text that ignores the requested look. That failure comes directly from too little training data covering the full range of fonts, effects, and animations that exist in real design work.

A pipeline that manufactures that training data automatically, rather than relying on hand-labeled examples, is one of the more practical routes to fixing the problem. Google's run of synthetic-data filings suggests the company sees data generation itself as a core engineering challenge, not just a preprocessing step. If the approach works as described, future AI tools could handle animated text overlays, styled titles, and branded graphics with far more accuracy than today.

Google's 63rd filing we've tracked since May in the AI photo editing race builds on earlier applications covering how light affects photos and facial expression edits.

Source: US 2026/0301260 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 Google 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 · 22 sheets

22 drawing sheets from US 2026/0301260 A1 · click any drawing to enlarge

Machine learning model training data., drawing sheet 1 from US 2026/0301260 A1
Sheet 1 Machine learning model training data.
Machine learning model input data., drawing sheet 2 from US 2026/0301260 A1
Sheet 2 Machine learning model input data.
Generating training data flowchart., drawing sheet 3 from US 2026/0301260 A1
Sheet 3 Generating training data flowchart.
Template description generation pipeline., drawing sheet 4 from US 2026/0301260 A1
Sheet 4 Template description generation pipeline.
Instance rendering pipeline., drawing sheet 5 from US 2026/0301260 A1
Sheet 5 Instance rendering pipeline.
Media content management systems., drawing sheet 6 from US 2026/0301260 A1
Sheet 6 Media content management systems.
Image and video examples., drawing sheet 7 from US 2026/0301260 A1
Sheet 7 Image and video examples.
User devices and servers., drawing sheet 8 from US 2026/0301260 A1
Sheet 8 User devices and servers.
Show the other 14 sheets
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 9
Sheet 9
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 10
Sheet 10
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 11
Sheet 11
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 12
Sheet 12
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 13
Sheet 13
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 14
Sheet 14
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 15
Sheet 15
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 16
Sheet 16
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 17
Sheet 17
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 18
Sheet 18
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 19
Sheet 19
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 20
Sheet 20
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 21
Sheet 21
Google Patents a System That Teaches AI to Copy Text Styles in Images patent drawing, sheet 22
Sheet 22
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.