Are AI-Upscaled Artworks Copyrightable in 2026?

Short Answer: Usually Only the Human-Created Parts May Qualify

Yes, an artwork can be partly copyrightable after it has been enlarged or improved with AI upscaling, but the upscaled result is not automatically protected as a new photograph or painting. U.S. copyright law requires a work to originate in an identifiable human author; as of September 27, 2026, purely AI-generated images are generally not copyrightable in the United States. If an artist makes substantive human modifications in Photoshop before or after upscaling, copyright may cover the human-authored elements, but it normally will not cover pixels or details invented by the tool. AI upscaling is usually better understood as image processing than authorship: it can increase resolution, alter texture, or reconstruct details, yet those changes do not necessarily represent the human creative choices protected by copyright. The legal answer therefore depends less on the size of the file and more on what the human actually created.

Also worth reading: How Do You Create Accurate, Marketplace-Ready Product Images With AI? · What Are Vendor Master Controls for AI Product Images, and How Should Teams Implement Them? · How Should Ecommerce Teams Control AI Product Image Quality in 2026?

For businesses using AI product images, the safest position is to preserve the source design, edit records, prompt logs, and evidence of substantial human selection, arrangement, painting, or compositing. Do not describe an upscaled image as “100% human-made” if the model invented visual features. Commercial use is not the same as copyrightability: a company may lawfully sell or advertise a non-copyrightable image, yet it may lack an exclusive copyright license and face weaker enforcement against copying. This distinction matters especially when the image depicts a product whose photograph, package, logo, or industrial design may carry separate rights. The conclusion is not that AI upscaling destroys every copyright claim; it is that protection must be tied to qualifying human expression rather than the final upscaled file as a whole.

Why Upscaling Does Not Automatically Create Copyrightable Expression

AI upscaling tools such as Gigapixel AI analyze a supplied image and produce a larger version by predicting plausible pixels, edges, and textures. The user may select an output resolution, denoising strength, face restoration, or detail level, but these technical settings are not automatically creative authorship in the legal sense. Traditional enlargement can also add or replace pixels, and courts and agencies have long distinguished mechanical reproduction from independently created expression. The deciding issue is whether the user contributed original and independently identifiable expression to the work, not whether software calculated a more complex result. Consequently, merely running a 1,024-pixel product mock-up through a 4× or 16× upscaler is unlikely to turn it into a new copyrightable artwork.

The U.S. Copyright Office’s 2025 report on copyrightability states that purely AI-generated material cannot satisfy the statutory requirement of human authorship. The agency also uses a case-by-case method when human and AI-generated material are combined: protection may extend to the human contributions that can be separated from AI assistance. A human might create the original composition and use AI only to enlarge it, in which case the original work remains protected even if the enlarged duplicate is not separately registered. A human might instead add painted highlights, materially redesign the background, and combine several elements into a new arrangement, producing a stronger claim to the human-created additions. Copyright does not arise because the final image looks polished or because the user selected a preferred output; selection alone may not contain enough original expression.

This approach does not mean that every choice in an AI-assisted workflow is legally irrelevant. A creator who spends substantial time iteratively masking, painting, arranging, correcting, and compositing parts can make a persuasive authorship record. The problem is that “substantial” cannot be reduced to a reliable percentage such as 10%, 30%, or 50%. No general safe-harbor threshold exists, and an office examining the claim will focus on the actual creative and expressive contribution. Keeping drafts and documenting revisions is therefore more useful than asserting a numerical AI-use threshold that the law does not recognize.

What Counts as Human Authorship in an AI-Upscaling Workflow?

The strongest copyright claims arise when the human contribution is both creative and identifiable. Drawing a product in a recognizable pose, painting background shadows, designing typography, arranging cut-out components, or substantially changing composition can supply human expression. Using conventional software such as Photoshop for these tasks does not weaken authorship; copyright law does not reserve creation to traditional media. What matters is that the person determines the expressive result rather than delegating the creative work to a generative system. An artist can also create one version, upscale it, and then use the enlargement as a working canvas for new manual edits, provided the later edits are sufficiently identifiable.

By contrast, generating a product scene from a text prompt creates a more difficult authorship case. If the person describes “a red bottle on a marble table with cinematic lighting,” the software may still decide the bottle shape, reflections, shadows, labels, and scene layout. The fact that the prompt uses a few adjectives does not establish that the user authored every visual element. The Théâtre D’opéra Spatial dispute illustrates this divide: an image produced through Midjourney and subsequent digital editing led to a U.S. copyright claim being rejected because the claimant’s human contribution was not sufficiently separable from the AI-generated material. That case should not be generalized into a rule that Photoshop edits never help; it shows that edits must do real expressive work, not merely select or finish an AI-generated design.

A useful test is to ask whether a person could identify what was manually or conceptually created by the human without examining every generic-looking pixel. Product poses, custom illustrations, a hand-painted set, graphic overlays, and original arrangements are more likely to qualify. Fabricated micro-details, decorative patterns, and background objects generated by the model are less likely to qualify. Where human and AI material are inseparably blended, registrability may be limited or refused, and the owner may need to identify only the protectable portions. Prompt logs can help explain the process, but a prompt itself is not a substitute for visible human authorship.

What About Copyright Registration, Registrability, and Infringement Claims?

Copyright exists under statute when a qualifying work is fixed in a tangible medium, while registration provides federal benefits and creates a practical record for enforcement. Some purely AI-generated images may still contain enough factual or human-created material to support a narrower registration, depending on the submission and the evidence. The Copyright Office has not accepted “AI-assisted” or “AI-upscaled” as magic labels that settle the legal question. Applicants are expected to explain the human author’s contribution, and they may need to disclaim material generated by AI. This means a platform cannot simply rely on a receipt saying “made in Photoshop” or “upscaled with AI” as conclusive proof.

Failure to obtain copyright in a portion of an image does not automatically make use of that portion unlawful. It can, however, make an exclusive-rights theory harder to enforce. Copyright owners alleging copying often identify protectable selection, arrangement, graphic elements, or modified expression and compare those features with the accused work. If the allegedly copied material is a generic bottle, surface texture, or AI-generated shadow, those features may be weak candidates for copyright. Trademarks can offer a different claim when a protected logo or source-identifying mark is reproduced, although misuse rules and likelihood-of-confusion standards differ. Patent and trade-dress questions may also arise for functional product features, but those are outside ordinary image copyright analysis.

A rejected registration does not always end every possible copyright suit; it may simply remove an existing registration as the basis for a federal claim, and factual disputes can still be litigated. Nevertheless, a registrant who gives inaccurate information about AI creation can create separate exposure involving 17 U.S.C. § 506(c), under which a knowingly false representation of ownership or exclusive rights made in a Copyright Office matter can lead to statutory damages. Businesses should therefore distinguish three facts in their records: who created the source image, which elements humans materially modified, and who owns the resulting file. Those records should describe the actual workflow accurately instead of treating an AI output as though a photographer personally painted every visible pixel.

AI Upscaling Compared with Conventional Retouching and Full Generation

There is no single category covering every function marketed as “AI upscaling.” Some tools enlarge an image while retaining much of its existing detail, while others hallucinate new texture, faces, lettering, or product geometry. That technical difference can matter to authorship, but even a large visual change is not automatically copyrightable. A conventional clone or sharpening filter also changes pixels, yet the user ordinarily does not claim a new work merely because the file became larger. The relevant question remains whether the human made original expressive choices through manual drawing, design, arrangement, or other creative authorship.

Upscaling is also not interchangeable with creating a new product photograph. If a real photographer takes the source image, that photograph may contain copyright even before enhancement, subject to ordinary authorship and fixation rules. If a designer composes a wholly synthetic product scene, the file may be commercially usable but lack copyright in the generated details. For catalogs, the answer can therefore be operationally simple: retain the licensed photography or owned illustration, use upscaling only for a technically usable rendition, and manually correct any altered product shape. A file’s 4K or 8K size says nothing by itself about its legal status.

Practical Steps for Copyrightable AI-Product Image Workflows

First, create as much of the product depiction as possible through identifiable human work. Photograph products owned or lawfully shot by the business, draw custom labels, or build the base scene manually. If a generator is needed for a background, keep the product itself separate as a protected layer and use a masked or constrained tool that preserves its geometry. Second, document the order of operations. Save the original, the intermediate output, every major revision, and the final export, while retaining layer files that show painted or composited areas. Timestamps, cloud histories, and version-control commits can corroborate authorship, whereas a statement written years later is weaker evidence.

Third, inspect upscaled images at high magnification. AI systems can invent package text, change logos, merge control buttons, add extra perforations, or smooth away safety markings. Those defects create more than an aesthetic concern: a consumer may receive a materially different product, and a trademark reproduced inaccurately can produce brand and consumer-protection risk. Manually redraw essential features and compare the output with a physical sample. Fourth, do not submit misleading registration applications. Describe “AI-assisted enlargement” accurately, identify the human author or authors, and disclaim generated material when the Office’s procedures or the work’s actual structure require it. Fifth, obtain licenses for all source assets. Copyrightability and permission are different questions, and an image without copyright may still violate trademark, contract, publicity, or trade-secret rights.

These steps do not guarantee registration or eliminate infringement risk, because agencies and courts assess facts case by case. They do make the process more defensible and help a team explain what was human-made. The most important operational principle is separation: isolate product-specific, custom, and manually edited expression from generated detail. That allows a marketer to make a narrower and more credible claim without implying that every synthetic pixel is exclusively owned.

Common Mistakes, Costs, and the Point at Which a Business Should Act

A common mistake is assuming that more human supervision converts every AI-generated pixel into human authorship. Selecting among multiple outputs, naming a model version, or entering a detailed prompt can establish control over the process, but it does not necessarily establish copyright in the final visual expression. Another mistake is treating Photoshop as a cure-all. Applying a final filter, increasing contrast, or making a handful of small cleanup changes may not be enough to support a claim across an image dominated by generated material. A third error is focusing only on U.S. copyright and ignoring logos, product-design rights, licenses, or the rules of the country where the image will appear.

Costs vary by workflow. Many conventional upscalers offer a limited free tier, while commercial desktop products and subscription services commonly range from about $10 to more than $100 per month or per annual license. Generative subscriptions can also run from free tiers to roughly $20–$30 or more per month, with charges for additional usage. Prices are not copyright thresholds: paying $200 for software cannot license copyrights the user did not create, and a low-cost human illustration may carry stronger protection than an expensive generated image. Budget should therefore be allocated partly for source photography, editing software, provenance storage, and legal review, not only for generation credits.

A business should act before publication when a product image will become part of a permanent brand campaign, paid advertisement, packaging, marketplace listing, or merchandise. The minimum response is to save provenance and inspect commercial identifiers. Larger brands should also compare AI output against product specifications, document who approved the final design, and obtain specialist review when a model materially changed protected graphics or a disputed source asset was used. If the image is a small, generic background with a low commercial stakes, the analysis may be simpler, but records are still inexpensive. By contrast, selling high-resolution prints, licensing the image to others, or advertising it as exclusively owned calls for a much stronger authorship and rights record.

The Defensible 2026 Position

The definitive U.S. answer is that AI-upscaled artworks can be partly copyrightable, but upscaling by itself does not create a new copyright in the entire image. Existing human expression remains potentially protected, and substantial identifiable human painting, compositing, arrangement, or design can support protection for qualifying additions. Purely AI-generated pixels ordinarily do not qualify because the Copyright Office requires human authorship. International results can differ, and a business distributing globally should check the law in each relevant market rather than assume that the U.S. rule settles every question.

For AI product images, the defensible claim is not “AI made this product image, so all of it is copyrightable” or “AI was used, so all of it is public domain.” The accurate claim is narrower: identify the human-created source and modifications, preserve the generated or machine-enhanced material as a separate issue, and verify other rights attached to the product. This approach may not produce an exclusive right over every polished detail, but it avoids overclaiming and makes the asset more useful in evidence, licensing, and commercial decisions. As of September 27, 2026, provenance and human creative control are therefore more important than output resolution alone.