[FOOD]

Log a Meal by Photo With the AI Scanner

[Published 2026-08-27 · 827 words]

[THE SHORT ANSWER]

The AI scanner photographs a plate and returns an itemised estimate you edit before it lands. It is opt-in: it runs on a Google Gemini API key you supply in Settings, sends nothing but the one resized photo, and does nothing at all until you add a key. Every item it returns is renameable, resizeable and deletable, and nothing reaches your log until you commit the reviewed list.

Setting it up

Settings, then AI Photo Scanner. Paste a Gemini API key from Google AI Studio.

Until a key exists, opening the scanner shows an explanation of what the feature would do and a button through to that settings screen. It does not half-work and it does not nag.

This is the one place the app's otherwise strict network policy has an asterisk, and the settings screen states exactly what leaves the device and why.

What actually gets sent

One photograph, resized to 1024 pixels wide, posted directly from your phone to Google's Gemini API using your key.

That is the whole payload. No account identifier, no training history, no food log, no device identifier. Base does not proxy the call through a server of its own — there is no Base server in the path at all — which is both a privacy property and the reason the feature costs you nothing beyond your own Google usage.

The photo is not retained by Base. What Google does with it is governed by your own API key's terms, which is worth reading once if the answer matters to you. See what Base does with your data for the complete picture.

Using it

The AI scanner before a key is set. It explains what it would send and links to Settings rather than half-working — the feature is genuinely opt-in.

Shoot the plate, or pick a photo from your library. The scan runs through three visible stages — camera, analysing, review — and any failure lands on an explicit error with a reason rather than an empty list.

The scale warning

If the model could not find a scale reference in the frame, Base says so before you review.

This matters more than any other caveat about photo estimation. A photograph carries no weight information whatsoever. Portion size is inferred from apparent volume, and apparent volume depends entirely on knowing how big something in the frame is. A bowl of rice photographed close up and a bowl of rice photographed from standing look identical to a model with nothing to calibrate against.

Practical fix: get something of known size in the shot. A fork, a standard dinner plate, a hand. It measurably improves the portion estimate and costs nothing.

Reviewing before committing

The review list is the part that makes the feature usable rather than merely impressive.

ActionEffect
Rename an itemFixes a misidentification without losing its macros
Change gramsRescales that item's macros proportionally from a stable baseline
Remove an itemDrops it entirely

The rescaling is linear off the original estimate rather than compounding, so nudging a portion up and back down returns exactly where you started:

detected:  chicken thigh, 180 g, 412 kcal
corrected: 130 g
scaled:    412 × (130 ÷ 180) = 298 kcal

A running total sits below the list so you can see the whole plate before committing.

When to use it, and when not to

Good for: restaurant meals, food someone else cooked, anything with no packaging and no obvious recipe, a plate you want logged in fifteen seconds rather than five minutes.

Bad for: anything with a barcode, which the barcode scanner handles far more accurately; anything you eat regularly, which is better saved as a meal; anything where the exact number matters, because the exact number is not what this produces.

The honest framing is that a photo estimate you actually log beats a precise entry you do not. Adaptive maintenance is computed from your average logged intake, so a missing meal shifts your measured maintenance permanently while a roughly-right meal barely moves it.

Common mistakes

Treating the output as measured. It is an estimate from a picture. The scale warning exists because Base knows when it is estimating badly, and it tells you.

Committing without reviewing. The review step is not a formality — misidentifications are common and take one tap to fix.

Shooting from directly overhead with nothing else in frame. That is the worst possible input for portion estimation. Angle it and include a utensil.

Using it for packaged food. A barcode gives you the manufacturer's own numbers. There is no contest.

Re-scanning the same home meal every day. Scan it once, correct it, then save the plate as a meal from the Meal tab. After that it is one tap and no network.

Expecting it to work without a key. It is opt-in by design and will keep pointing at Settings until you add one.

What to do next

If you eat out regularly, add a Gemini key and scan one restaurant meal to see how the review list behaves. If you mostly eat at home, skip this feature entirely — saved meals are faster, free and exact.

Questions

Is the AI food scanner free?

The feature is part of the app, but it runs on a Google Gemini API key that you supply yourself and add in Settings. Google's free tier covers ordinary personal use comfortably. Base does not proxy the request through any server of its own, which is also why there is no cost to you beyond whatever Google charges your key.

What gets sent when I scan a plate?

One photograph, resized to 1024 pixels wide, sent directly from your phone to Google's Gemini API using your own key. Nothing else goes with it — no account identifier, no training history, no log data. The photo is not stored by Base and never touches Base's own servers.

How accurate is a photo calorie estimate?

Treat it as a starting point rather than a measurement. A photograph carries no weight information, so portion size is inferred from apparent volume and any scale reference in frame. Base explicitly warns when it could not find a scale reference, because that is the condition under which the estimate is least reliable. Every item is editable before you commit.

Can I edit what the scanner found?

Yes, and you should. The review list lets you rename any item, change its weight in grams — which rescales its macros proportionally — or remove it entirely. Nothing is written to your log until you commit the reviewed list.

Do I have to use the AI scanner?

No. It is entirely opt-in and does nothing until you add a key. Without one the screen explains what it would do and links to Settings. Barcode scanning, search and Quick Add all work with no network account of any kind.

[KEEP READING]