Affinity on iPad Draws the AI Line at the Cloud, and That Line Is the Point

Canva released Affinity for iPad on 2 November, and the pricing is the part everyone will repeat: free for individuals, with a free Canva account, no subscription required for the core app. The more interesting decision is buried one paragraph deeper. The masking AI runs on the device and costs nothing. The generative tools run in the cloud and need Canva Pro at 12 euros a month or 110 a year, with monthly caps on some of them.

That is a line drawn between two kinds of AI, and it happens to be the right line. Selection work is a chore you repeat fifty times in an afternoon. Generative fill is an editorial decision you make twice. Charging for the first and metering the second matches how the work actually feels, and it is a useful signal for anyone deciding where AI belongs in their own pipeline.

What actually shipped

Affinity for iPad puts vector drawing, photo editing and multi-page layout into one application, the same consolidation the desktop version went through after Canva bought Serif. The interface is built for touch and Pencil rather than ported from the mouse, and Canva claims edits stay live on large files with many layers.

The new part is a dedicated drawing area. Brush strokes respond to Pencil pressure and tilt, and hand-drawn shapes can be converted into clean vector shapes that stay editable with the vector tools. That conversion is the one feature that is genuinely better with a pencil than with a mouse, because the input is a sketch and the output is a path.

The photo workspace imports and edits RAW files, and carries selection and retouch tools. The masking tool uses AI to cut difficult edges such as hair, fur or leaves. That model runs locally, and heise's report is explicit that it does not count as one of the paid generative features.

Layout handles multi-page documents with repeating page templates and grids, aimed at photo books and magazines. Everything a project contains lives in a single file. Affinity reads files from the older Photo and Designer apps, but those older apps cannot open the new format.

Why the local/cloud split is the right one

Strip the branding and there are two AI workloads here with completely different cost shapes.

Masking is inference on a small model over one image, run maybe two hundred times a day by a working designer, and the output is a mask you can inspect and correct. It is deterministic enough to be a tool. Putting it on the device means no per-use cost, no upload of client material, and no round trip in the middle of a selection.

Generative fill is a large model producing new pixels that did not exist in the source. It is expensive per call, it changes the meaning of the image, and it is the part that carries disclosure obligations when the image involves a real person or a real product. Metering it is not just commercially convenient; it keeps a visible boundary around the operation that actually needs a decision.

Designers who have lived through the last two years of tool churn will recognize why this matters. The failure mode is not AI in the editor. It is AI folded into a shortcut you already press without thinking, so that generated content enters a file with no marker and no memory of the decision. Affinity's split at least makes the generative step a place where you consciously spend something.

The format asymmetry is the real switching cost

The file news deserves more attention than the feature list. A single project file is genuinely better than juggling Photo, Designer and Publisher documents. Reading the old formats means a migration path exists.

But the compatibility is one-directional. New files do not open in the old apps. If you are mid-project across a team, or you keep a client archive that someone else still opens in Designer, that asymmetry is where the cost lands, not in the subscription.

The practical rule is boring and saves a week: finish one real project end to end on the new app before you commit a client archive to it, keep the old apps installed until the migration is actually complete, and export a flattened deliverable alongside the working file for anything with a long tail. If you have ever had to reconstruct a layout from a PDF because the only copy of a source file was in a format the new version broke, you already know the rest.

Where a tablet actually fits

The honest placement is a first-pass device. Sketching, roughing a layout, cutting a subject out of a phone photo, marking up a client's draft on the train: all of that is faster on glass, and the Pencil-specific features are the reason.

Long layout sessions, color-critical retouching and print preparation still belong at a desk with a calibrated display. That is not a knock on the app. It is the same division that has held for every tablet creative tool since the first stylus.

For anyone running an AI-heavy pipeline, the interesting test is whether the local masking is good enough to remove a hosted background-removal step. That is a measurable question, not a matter of taste: run twenty of your own hard edges through both, at the size you actually export, and compare how many need manual repair.

What the announcement does not answer

Three things I would want before trusting it for production work, none of which the release covers.

Print color management. A layout app that cannot carry a proper CMYK workflow to press is a mockup tool, and the announcement says nothing about color spaces, separations or output intents. Verify on your own file before a print job depends on it.

Round-trip fidelity. "Reads older files" is not the same as "opens them without reflowing the master pages". Test the ugliest file you own, which is always the one with fifteen linked assets and a table.

Whether the local model is actually local. On-device inference that quietly falls back to a server when the model is missing or the device is old is a different product, and worth checking with the network monitor on before you promise a client their unreleased product shots never leave the tablet.

The line worth copying

The free tier is generous, and that is fine, but the structural decision is the one to remember: chores are local and free, synthesis is remote and metered. Whatever tool you build or buy next, that is a reasonable test of whether its AI belongs in the loop at all.

If you want the wider argument about where tool value actually sits, a Photoshop clone at 45 percent parity is the same question asked of a very different project. And once generative fill is in your file, the disclosure maths from the German trust study is the part of this that has nothing to do with software.