Apple has a pending US patent application for a system that takes a marketing image or screenshot for an app, keeps the headline and the phone in it, and regenerates everything else around a keyword before a scoring model filters the results. The application, US20250265066A1, was published on August 21, 2025. It drew fresh attention in early November 2026, when the ASO Patents site published a breakdown on November 5 and app marketer Ramazan Tugay Kahraman, one of the site's three listed operators, promoted it in a LinkedIn post.

In Short

Apple filed a patent application describing software that turns one app marketing image into several keyword-specific versions, keeping the headline and device mockup and repainting the background with Stable Diffusion models. It matters to app marketers and Apple Ads buyers because the design treats one creative per search keyword as the unit of work, and dependent claims cover keywords and images generated from the app itself rather than typed in by a developer. Nothing has been granted or launched: the application has been pending since 2025, contains no performance data and never names the App Store.

What the filing covers

The document is titled "Techniques for managing images associated with software applications." Apple Inc. is the assignee, and eleven inventors are named: Jia Huang, Anikesh Kamath, Babak Seyed Aghazadeh, Dayvid Victor Rodrigues de Oliveira, Justin M. Platz, Kanishka Bhaduri, Milind Pandurang Jagre, Robert J. Monarch, Yuan Yen Tai, Xiaoyuan Gu and Alex J. Kim.

The paper trail is short and regular. Apple filed provisional application 63/556,382 on February 21, 2024, and the full application, 19/060,399, exactly a year later on February 21, 2025. The inventors signed the assignment between February 24 and March 14, 2025, and it was recorded on March 18. Three days later, on March 21, 2025, the case was docketed as ready for examination. Publication followed on August 21, 2025, 18 months after the priority date, which is the standard interval for US applications. Google Patents lists the status as pending, and the most recent legal event it records is that March 2025 docketing. The application has not been granted, so it confers no rights.

Google Patents files it under CPC code G06F8/63, a class titled "Image based installation; Cloning; Build to order" - an odd home for a document about marketing pictures, since that code concerns software installation.

The problem statement is plain. According to the background section, "producing such digital assets is typically costly and time consuming. For example, to reflect the keywords 'fast food' and 'vegetarian,' distinct images need to be gathered/produced, merged together in an aesthetically pleasing way, and so on, which can be a challenging endeavor for those who lack design expertise or confidence." The next paragraph adds the scale argument: "In view of the large volume of search terms frequently provided by users, trying to create digital assets, customize them, and evaluate them, presently requires substantial human involvement."

The same passage names the weakness of off-the-shelf tools. Unsupervised text-to-image models, the filing says, "frequently generate distorted, unrealistic, or even inappropriate images, which can lead to a poor user experience." Much of what follows is an attempt to fence those models in.

Claim 1 in six steps

The application carries 20 claims. Three are independent - claim 1 (a method), claim 9 (a non-transitory storage medium) and claim 17 (a computing device) - and they recite the same sequence. A computing device:

  • receives an original image associated with a software application, containing at least one foreground element;
  • extracts that element, producing the extracted element and a "modified original image" with the element removed;
  • generates two or more candidate images from three inputs: the extracted element, the modified original and at least one keyword associated with the original image;
  • generates a quality score for each candidate, "based at least in part on the at least one keyword";
  • filters the candidates by score;
  • associates at least one surviving candidate with the software application.

The dependent claims narrow it. Claim 2 defines the foreground element as a screenshot of the app, "a hardware device on which the software application is capable of executing," descriptive text, or a combination. Claim 3 attaches masking information to the modified image. Claim 4 adds a step in which filtered candidates are offered "for selection" and a selection is received. Claim 5 places the chosen image in "an online application store profile." Claim 8 describes generating "at least one additional element" from the keyword and blending it with the foreground and the modified original.

Claims 6 and 7 deal with where the inputs come from, and they matter more than their length suggests. They are covered below.

How the pipeline is built

Three components sit on a server: a masking engine, an image generator and an image selector. The filing lists a long menu of possible model types for them, from small and large language models to "biological models," and states that rules-based systems would also qualify. The worked example, though, names specific open models.

Cutting the image apart

The masking engine runs "foreground element detection, segmentation, and mask post-processing operations." Detection uses "a pre-trained detection model (e.g., a CenterNet model)" to find interface elements - the filing lists dozens, from buttons and sliders to carousels, avatars and chat interfaces. According to the document, the detection step also covers "screenshots of software applications, computing devices that execute the software applications, and so on, which are commonly included in marketing materials for software applications." In other words, the standard store screenshot - a phone frame with a caption above it - is the expected input.

Text in the image is not only preserved. The filing says OCR "can be utilized to extract text content from the text elements," and that "the text content can be used to supplement, refine, etc., the keywords 105 that are provided along with the original image 104." The words on a screenshot can therefore feed back into the keyword set that drives generation.

Bounding boxes then go to "a pre-trained segmentation model (e.g., a Segment Anything Model)" for refined masks. A pixel entropy step decides which masks survive: masks covering high-entropy regions that meet "a predefined threshold" are kept, while others "can be disregarded." Finally, "dilation and Gaussian blurring" soften the mask edges so the new background blends rather than cuts.

Repainting the background

The image generator takes the original image, the keywords and the masks. According to the filing, it "can implement three different weights of the stable diffusion model, including a pre-trained Stable Diffusion inpainting model, a pre-trained Stable Diffusion XL (SDXL) model, and a fine-tuned Stable Diffusion model." The masks tell it "where inpainting should be performed."

Keywords build the prompt, which is padded with "broad terms, e.g., 'best quality,' 'clear image,' 'abstract,' 'realistic style.'" Negative prompts act "as safeguards," with examples including "'not safe for work (NSFW),' 'text,' 'logo,' 'low quality,' 'weird,' 'blurry.'" A pre-trained aesthetic scoring model then rates each generated image for relevance to the original, whether it "enhances" it, and appropriateness.

The constraint is deliberate. One of the models is tuned "using a set of (image, text) pairs that have been deemed to be relevant/appropriate to the types of images described herein," which, according to the filing, "can effectively establish a bounded generation process, which can help avoid the generation of inappropriate or policy-violating images." The document does not say what those pairs are, where they come from, or under what terms.

The worked example in FIG. 2 uses a food delivery image with the headline "Affordable Meals At Your Fingertips!" and a smartphone showing the app. The headline and the phone are masked as foreground. One candidate adds "a coffee icon and a utensil icon," and in it "the text included in the original image 104 can be reproduced in italic form." Another adds star icons and "additional/modified decorative lines." The first candidate scores highest and becomes the selected image.

Scoring and filtering

The image selector is where the keyword returns as a judge. According to the filing, it "can implement a pre-trained cross-modal image-text matching architecture, e.g., Learning Cross-Modality Encoder Representations from Transformers (LXMERTs)." Text passes through a WordPiece tokenizer; candidate images go through "a Faster R-CNN object detection model" into object-level embeddings, and a transformer-based cross-modal encoder predicts the quality score.

The fine-tuned version is spelled out layer by layer: a linear layer, "a Gaussian Error Linear Unit (GELUE) activation function," a normalization layer, another linear layer, and "a sigmoid activation function as the last layer to output a binary classification prediction." A binary output means the selector asks whether an image and a keyword match, not how attractive the image is.

The selector also screens out what the filing calls hallucinations. It "can also be configured to identify and eliminate any candidate images 113 that are considered to be 'AI hallucinations.'" And an "explanation agent" can describe why candidates were dropped, which the document says can help "demonstrate to end-users that time is being saved by intelligently eliminating certain results for good/explainable reasons." A second explanation agent, attached to the generator, can draw "visible bounding boxes around areas of the generated image that were modified," delivered as metadata that can be toggled on and off.

What else the system can consult

All three components "can be configured to interface with knowledge sources 118." These include web search engines that "crawl" the internet, question-and-answer knowledge bases, knowledge graphs and approximate nearest-neighbour indexes. The context for a generated image is therefore not limited to what a developer uploads.

The system can also keep "conversation history information" between the developer and the server, which the filing says can include "user account information, demographic-related information, device-related information" about the developer side, as well as past "selections, feedback, etc." The stated purpose is to produce images that "appropriately align with observed preferences." Paragraph 37 extends the generator beyond still images to "audio content, video content, document content, web content (e.g., hypertext markup language (HTML) content)."

The submission screens

FIGS. 3A and 3B walk through a store submission. Interface 310 takes a brief app description - one "that will be displayed on a product page" - and one or more keywords. Interface 320 asks for "at least one marketing image that was generated for the software application" and shows that it is being processed. Interface 330 reports that "Enhanced Image 1" and "Enhanced Image 2" were generated and lets the developer view them beside the original "for comparison purposes." Interface 340 previews one; interface 350 offers "an option to select different images"; and after "Submit," interface 360 shows the product page carrying the chosen image.

The stated benefit is effort, not autonomy. According to the filing, the result "constitutes an enhancement to the original image 104 with exceptionally little input, work, etc., required from the distribution entity 102." The screens are, in the document's words, "merely exemplary and should not be construed as limiting."

Where would the selected image appear? Paragraph 18 gives three places: "within a product page for the software application within an online software application store, within a website for the software application, within an advertisement associated with the software application, etc." Advertising is named alongside the store listing.

Where the inputs come from

The passage that drew the most attention in Kahraman's post sits in paragraph 23. Developers can supply images and keywords at submission. "Alternatively, or additionally," the filing continues, "original images 104 and keywords 105 can be automatically generated for the software application 103. For example, one or more entities (e.g., the distribution entity 102, the server computing device 108, and/or other entities) can execute the software application 103, analyze its components, assets, logic, etc., to extract relevant screenshots from the software application 103 (e.g., screenshots of a primary UI, screenshots of relevant features, etc.), to identify relevant keywords, and so on."

According to the filing, this approach "effectively eliminates or reduces the need for distribution entities 102 to generate, review, and provide original images 104 and keywords 105 when submitting software applications 103." Paragraph 24 adds that rasterization engines "can be useful to generate screenshots from a given software application."

This is not only description. Claim 6 covers an original image that is "automatically generated based on one or more aspects of the software application." Claim 7 covers a keyword "automatically generated based on one or more aspects of the software application" or "automatically generated based on the original image." Claims 14 and 15 repeat both for the storage-medium version. What the claims do not recite is the mechanism - running the app and reading its components - which appears only in the description.

Note also who does the executing in that passage. The examples include the "distribution entity," which the filing defines as the developer or company behind the app, as well as the store's server and "other entities." The document leaves open which side would run the app.

How the ASO Patents reading compares with the document

Kahraman's LinkedIn post, written under his headline "Senior Performance Marketing & Organic Growth Manager | ASO, Apple Ads, Meta Ads, Google Ads," compressed the filing into a few lines. "Apple has filed for a way to make new App Store screenshots out of the ones you already uploaded," he wrote. "It keeps your headline and your phone mockup, repaints the background, and builds the new version around a keyword. You get a few options back and you pick one."

He then linked it to a product Apple shipped. "Then I remembered Asset Library, which shipped in October and lets you get creative assets approved without an app update," the post said. "No proof the two are connected. Just hard to unsee." On the date, the post is consistent with Apple's own record: Apple's developer news announced the Asset Library in App Store Connect on October 5, 2026, saying developers can submit creative assets for approval and use them later, for example in a seasonal campaign. On the connection itself, the post's own caveat stands. Nothing in the patent mentions an asset library, creative assets as a format, or seasonal campaigns.

The post closed with: "US20250265066A1. Still pending, so nothing here is promised."

The longer ASO Patents article, published November 5, 2026 under the headline "US20250265066A1: Apple's AI Image Generator for the App Store," carries no byline. Its About page says the site is managed by Ramazan Tugay Kahraman, Hakan Can İncili and Emre Kahraman. The article is careful in many places. It states that the application "confers nothing, the claims have not survived examination, and an application describes what a team designed rather than what the store runs." Its list of what the patent "does not tell you" - no keyword-generation method, no ranking use, no evaluation numbers, no commitment to ship - matches the document.

Several statements go further than the filing, or sit awkwardly with it.

Claims. ASO Patents says the automatic derivation of images and keywords "is not in any claim" and, in its conclusion, that this part "is one line and claimed nowhere." Claims 6, 7, 14 and 15 do cover automatically generated images and keywords. The narrower point - that executing the app is not claimed - holds.

Who decides. ASO Patents writes that "The developer is in the loop at every step" and that "The store is not quietly rewriting your product page." The interfaces in FIGS. 3A-3B do keep the developer choosing. But claim 1 contains no human selection step; that arrives only in dependent claim 4. Paragraph 18 says images with the strongest scores "can be selected (e.g., automatically, by one or more users, etc.)," and in FIG. 2C the top-scoring candidate "is output as selected image 116" without a human choice described.

Text preservation. ASO Patents says "Your words stay yours because they were lifted out as foreground before any of this ran." The words do survive in the worked example, but their styling does not necessarily: one candidate reproduces the headline "in italic form."

Apple's earlier work. ASO Patents states that LXMERT "is the same architecture Apple used in its cross modal creative selection work, the one trained on advertising creatives that earned clicks and conversions." The article does not identify or link that work, and the patent itself makes no reference to advertising performance data, clicks or conversions.

Naming the store. Both the post and the article speak of the App Store. The filing never does. It refers throughout to an "online software application store (hereinafter, 'OSAS')," and the word Apple appears only in the assignee data.

Timing. Kahraman's post opens: "Looking for something else in Apple's patent filings last week, I ran into this one." The application had been public since August 21, 2025. By PPC Land's calculation, 441 days passed between publication and the ASO Patents write-up.

Where such a tool would land

The filing predates, by more than a year, most of the App Store marketing changes that give it practical relevance.

Apple previewed creative assets at WWDC26 on June 8, 2026, describing rich images and videos for the product page header and search results, an Asset Library and a product page preview tool. On August 5, Apple's best practices placed creative assets in the product page header, in search results on the App Store and in the Apple Games app, in custom product pages, in In-App Events and in Apple Ads campaigns on the Today tab and in search, and required a 4+ rating for every asset even when the app itself is rated higher. On October 16, Apple moved part of that guidance into the rules with guideline 2.3.14, which bars specific or discounted pricing and unverifiable claims in creative assets.

Screenshot workloads are rising at the same time. On October 5, Apple told developers that apps and games submitted from April 2027 must include iPhone Duo screenshots, on product pages that hold up to 10 screenshots and up to three app previews, with up to three of each shown in search results. Each new device size, language and campaign multiplies the set of images a developer maintains - precisely the "substantial human involvement" the patent's background section describes.

Keywords are the other half. At WWDC 2025, Apple introduced App Store tags generated by large language models from app metadata and other data, and let developers assign keywords to custom product pages, so that a tailored page appears in organic results for those terms. The tags go through human review, and developers can deselect them. Apple has, in other words, already let a model describe apps on its store. The patent extends machine generation to pictures and keyword lists. It does not mention tags or custom product pages.

Paid placement completes the picture. Apple rebranded Search Ads as Apple Ads on April 14, 2025, listing placements on the Search tab, the Today tab and other developers' product pages. From March 3, 2026, multiple ad positions in App Store search results rolled out, starting in the UK, with ads using either default product pages built from App Store metadata or custom product pages made in App Store Connect. An image generator that produces a keyword-matched creative from one upload would fit that chain. The filing does not say it will, and Apple has not said so either.

Platforms that already generate creative

Apple would not be first. Google expanded its Asset Studio in May 2026 so that it generates assets from a marketer's brief, brand guidelines and website, and added Gemini Omni video creation and 1-Click Creative Testing. More pointedly, Google will turn on image extraction by default for new Performance Max campaigns from December 4, 2026, scanning landing pages for images to add to campaigns unless advertisers opt out in the campaign payload. That is the web-ads equivalent of the patent's paragraph 23: the platform derives creative from the advertiser's own property.

Meta has built reporting around the same question. In July 2025, it added a Creative breakdown in Ads Manager that compares AI-generated image variations with original content. The Apple filing contains no equivalent measurement. Nowhere does it report whether enhanced images convert better, rank better or perform at all.

The choice of Stable Diffusion has its own history. On November 4, 2025, the High Court of England and Wales dismissed Getty Images' secondary copyright claims against Stability AI over Stable Diffusion training, while finding trademark liability for some watermark outputs. The patent names three Stable Diffusion variants, one of them fine-tuned, but is silent on licensing and on the source of the image-text pairs used for tuning.

What the document does not establish

The filing gives no accuracy figures, no test results and no comparison between generated and original images. It does not say whether the system exists inside Apple, whether it was ever tested with developers, or whether it will ship. It never names the App Store, Apple Ads or App Store Connect.

It does not say generated keywords would be shown to developers, used in search ranking or used in ad auctions. It does not describe how an automatically generated keyword would be checked against what an app actually does. And it does not say whether shoppers in a store would be told that an image was machine-generated; the explanation-agent annotations are described as metadata for the people producing the image, not as consumer labels.

The privacy section, paragraphs 64 to 68, is standard language about personal information data, opt-in and opt-out controls and de-identification. It does not explain how demographic information about developers, mentioned in paragraph 25, would be used in generating images.

There are also signs of hurried drafting. Paragraph 26 lists "image generator 112" twice where the selector appears to be meant. Paragraph 40 names the activation function "GELUE." And paragraph 43 states that the embodiments "primarily pertain to mapping-based search queries 106 and selected image 116" - a reference numeral that appears nowhere else in the document, followed by a claim that the techniques could apply to "movies, television shows, books, documents, web pages." None of this affects the claims. It does suggest the specification shares text with other filings.

For app marketers, the substance comes down to two things. First, the design treats the keyword - not the campaign or the app - as the unit of creative production. Second, Apple has written, and partly claimed, a path in which the store side can derive both the image and the keyword from the app. Whether either becomes a product is a separate question, and the document does not answer it.

Timeline

Summary

Who: Apple Inc., assignee of US patent application US20250265066A1 with eleven named inventors; ASO Patents, the site that analysed it; and Ramazan Tugay Kahraman, an app marketer and one of the site's three listed operators, who promoted the analysis on LinkedIn.

What: A pending application describing a system that keeps the headline and device mockup of an app marketing image, regenerates the background around a keyword using three Stable Diffusion variants, scores candidates with an LXMERT-based image-text matcher, filters out hallucinations and attaches a chosen image to an app store listing, website or advertisement. Dependent claims cover images and keywords generated automatically from the app.

When: Provisional filing February 21, 2024; full filing February 21, 2025; publication August 21, 2025; ASO Patents analysis November 5, 2026; LinkedIn post in early November 2026. The application remains pending.

Where: The United States Patent and Trademark Office, with the record on Google Patents. The filing describes an "online software application store" and never names the App Store.

Why: The filing argues that producing marketing images per search keyword "presently requires substantial human involvement" and that general image generators produce "distorted, unrealistic, or even inappropriate images." For app marketers, it describes a path in which creative is produced per keyword and, in part, derived from the app itself, at a time when Apple has opened creative assets, the Asset Library and more Apple Ads search slots.