The AI features in photo editing apps have become genuinely useful — one-tap background removal, automatic sky replacement, generative fill, noise reduction that would have taken a professional half an hour in 2020. They’re also, in many cases, sending your photos to cloud servers to run.
That’s been the default arrangement for several years. What’s changed in 2026 is that a handful of major apps have started moving AI processing back to your device — quietly, without much fanfare, driven partly by technical progress and partly by user pressure on privacy. Understanding where things stand now matters, especially if you’re editing photos that contain sensitive personal content.
Why AI Photo Editing Went to the Cloud in the First Place
Running a generative AI model requires significant computational resources. The models behind features like Photoshop’s Generative Fill, VSCO’s AI adjustments, and Google Photos’ Magic Eraser are too large to run on most consumer hardware — especially smartphones.
So the solution for the past few years was simple: when you tap “Generative Fill” or “Remove Object,” the app sends your image (or a cropped portion of it) to the company’s cloud servers, runs the AI model there, and returns the result. The processing happens in seconds from the user’s perspective.
The privacy implication is straightforward: your photo briefly leaves your device and lives on a server operated by the app developer. What happens to it there depends entirely on the company’s policies.
What Adobe Does With Your Photos
Adobe’s suite — Photoshop, Lightroom, and Firefly — has been the most-discussed case, partly because of Adobe’s market share and partly because of a terms-of-service controversy in 2023 that alarmed many photographers.
The current state, as of mid-2026:
Photoshop’s Remove Tool is now on-device. Adobe’s June 2026 update moved the Remove Tool (formerly one of its most-used generative AI features) to an on-device model. It now works without an internet connection, which means it works without your photo going to Adobe’s servers. This is a significant change.
Generative Fill and Generative Expand still use cloud processing. These features, which generate new image content from text prompts, still route through Adobe’s Firefly cloud infrastructure. When you use them, a portion of your image goes to Adobe’s servers.
Adobe’s stated policy on training: Adobe’s Generative AI terms state that content submitted by paid subscribers is not used to train Firefly or other AI models. That’s a meaningful policy commitment. It does not mean the images don’t touch their servers — they do — it means Adobe asserts they don’t train on them.
Lightroom AI adjustments are mixed. Lens blur, noise reduction, and most sliders in Lightroom run locally. The “Select Subject” and “Select Sky” features use on-device ML. Some more sophisticated masking features may still use cloud processing depending on your version.
The trend at Adobe is clearly toward local processing, driven by user complaints and competition from on-device alternatives. But “generative” features — the ones that synthesize new content — remain cloud-dependent.
Google Photos: Always-On Cloud AI
Google Photos is the largest photo AI platform in the world, and its approach differs from desktop tools: almost all AI features are tied to Google’s cloud infrastructure, by design.
When Google Photos recognizes faces, creates auto-albums, lets you search by “cats” or “beach sunset,” or generates a “memory” from a trip, that analysis runs in Google’s data centers. Your photos are already uploaded to Google’s servers for storage, so this is the same data already in Google’s cloud.
The implication is less about “sending photos somewhere new” and more about understanding that your entire library is continuously analyzed by Google’s AI systems. Google uses this analysis for features, not for advertising directly from photo content (per their privacy policy), but your photo library’s content informs everything from your Google account profile to the AI models Google trains.
Magic Eraser, Photo Unblur, and generative editing in Google Photos run via cloud inference. There is no on-device version of these.
If you use Google Photos, the cloud AI processing is intrinsic to the product, not optional.
VSCO, Snapseed, and Mobile AI Editors
The major mobile-only editors vary significantly.
Snapseed (Google-owned) processes all adjustments locally, on your device. It has no generative AI features. Everything you do in Snapseed stays on your phone. This is why Snapseed remains popular with privacy-conscious photographers — it is a capable non-cloud tool.
VSCO has added AI-powered tools. Depending on the feature, processing may use cloud inference. VSCO’s privacy policy permits using content to “improve our products and services,” which is vague enough to warrant scrutiny. Their current terms are less explicit than Adobe’s about not training on subscriber content.
PicsArt and Canva both offer AI editing features that are cloud-processed. Both are primarily consumer apps with advertising-supported tiers, and their data policies reflect that business model.
Lightroom on mobile follows the same pattern as desktop: most basic adjustments are local, generative features are cloud-processed.
Aftershoot and the On-Device Alternative
An interesting counterpoint to the major apps is Aftershoot, an AI photo editing tool aimed at professional photographers. Aftershoot runs its culling and editing AI models entirely on your local machine. Your photos never leave your computer.
Aftershoot made a deliberate product decision around privacy: building for photographers who shoot weddings, portraits, and other sensitive professional work, and who cannot afford to have client photos uploaded to a vendor’s cloud without client knowledge.
It’s slower than cloud-based tools for the initial model download and requires reasonably modern hardware, but the processing speed is competitive once the model is local. It represents the direction AI photo tools can go when the constraint is privacy rather than minimal hardware requirements.
What Gets Sent When AI Processes Your Photo
The mechanics matter. When an app sends your photo to a cloud AI endpoint, it’s typically not sending your whole camera roll — it’s sending the image or cropped region you’re actively editing.
What that image may contain:
- The full visual content of your photo
- In some cases, EXIF metadata including GPS coordinates, timestamps, and device model
- Your account identifier (so the result can be returned to you)
What happens on the receiving server:
- The image is decoded
- The AI model runs inference
- The result (edited image or mask) is returned
- The image may be cached temporarily, then deleted according to policy
“May be cached temporarily” is where policy commitments matter most. If a company’s policy says images are deleted after processing, you’re relying on them to actually do this. Companies with SOC 2 or ISO 27001 certifications have third-party audits of their security practices, though these audits focus on security controls, not privacy policy compliance specifically.
The Practical Questions to Ask About Any AI Editor
Before using an AI photo editing tool on photos you care about, it’s worth asking:
Does this feature run locally or in the cloud? Check the app’s help documentation or settings. Some apps label this explicitly; others don’t. A reliable signal is whether the feature works in airplane mode — if it does, it’s local.
What is the data retention policy for uploaded images? Some companies specify “deleted immediately after processing.” Others say “deleted within 30 days.” Some don’t specify at all. No specification is a signal worth noting.
Does the company train models on customer content? Adobe explicitly says no for paid accounts. Google says it uses data to improve products but not to build advertiser profiles from photos. Many smaller apps are vague.
Are you editing photos with information you wouldn’t want on a vendor’s server? Medical records, legal documents photographed for reference, photos of children, or anything commercially sensitive deserves extra care before you tap “AI edit.”
A Tiered Approach That Works
Most people don’t need to use the same editing tool for all photos. A tiered approach is practical:
Everyday photos — sunsets, food, landscapes — fine to use cloud AI editing. The privacy stakes are low and the results are good.
Photos with people, sensitive context, or identifiable personal information — use tools that explicitly process locally: Snapseed for mobile edits, Lightroom’s local-only features, Aftershoot for serious professional work.
Documents photographed for OCR or reference — avoid AI editing tools entirely. These are better handled by dedicated secure document apps.
The underlying storage question is separate from the editing question. Where you keep your photo library — in whose cloud, backed up how — matters regardless of which editor you use. Editing a photo in Snapseed doesn’t help if the original is stored in a cloud with looser policies than your editor.
The Direction Things Are Moving
The trajectory in 2026 is clearly toward more on-device AI processing. Hardware is fast enough, model compression has improved, and user demand for privacy has put real pressure on major developers.
Adobe’s move of the Remove Tool to on-device in June 2026 is a meaningful benchmark — a few years ago, that was considered too computationally demanding for local processing. Apple’s on-device photo ML (which powers Photos search and the clean-up tool) demonstrates what’s possible at the chip level.
Within the next few years, most basic generative features will likely run locally on flagship hardware. The cloud advantage for AI photo editing is shrinking.
But “moving in the right direction” and “already there” are not the same thing. For now, if you use generative AI features in the major editing apps, your photos are touching cloud servers. Understanding that — and making intentional choices about which photos you put through those features — is the practical response.