Thomas: AI Camera

Thomas is a camera whose film is language. Point it at the world and press the shutter. Instead of only saving a photograph, a vision model running on your iPhone says what it sees, and those words are the recording. Then a second model draws the scene again from the words alone, having never seen your photograph. One press, three frames: • The photograph: what the lens actually saw. • The words: what the machine perceived, in its own voice. • The drawing: the machine re-imagining the scene from that description. Reality, machine perception, machine re-imagining. The gap between the three is the point: a legible, pointable look at how a machine sees, which is not how you see. Not a worse eye. A different one. ON-DEVICE AND PRIVATE No accounts, no servers, no tracking, no ads. Both models can run entirely on your iPhone: the drawing always does, and choosing a downloaded eye keeps the seeing there too. (The built-in Apple Intelligence eye may use Apple's Private Cloud Compute when you're online.) For an even more private shot, stamp raw coordinates instead of a place name, or turn location off, and nothing is looked up. The photograph and the words work on a fresh install with no download at all, through Apple Intelligence. The drawing is an optional model you download when you want it. THE EYE Choose which machine looks. Apple Intelligence is built in and needs no download. Qwen is downloadable and sees differently. Neither is better; they are different machines. The eye's system prompt and its temperature are yours to read and to change. Experiment! THE HAND A diffusion model draws the scene again, and it never sees your photograph, unless you want it to. It can work from the eye's words alone. That is the whole reason the third frame is worth looking at. IT SAYS WHAT IT SEES No hedging, no confidence scores, no apology chrome. The machine states its perception plainly, and your photograph is sitting right there beside it. The juxtaposition is the disclaimer. Sometimes the words will be wrong. Thomas reports a perception rather than grading an answer, and honest perception is the art. HOW THE FRAMES MEET Words laid on the world, a diptych, a triptych of all three, or the words alone. Every frame is stamped with the place and the time, the way cameras always have: a visible stamp, never hidden GPS written into the file. THE SHOT IS ATOMIC You configure once, then just shoot. Finished frames land in Photos like a print surfacing in a chemical bath. The latency is the film developing. Requires iOS 27 on an Apple-Intelligence-capable iPhone. Powered by Stability AI. Thomas is open source under the MIT License.