Raychis

What comes after garden sharing

Garden sharing has shipped, and the work underway now is multi-angle capture. It identifies a plant from several photos of different parts and chooses the species they agree on, which helps most where a single photo is ambiguous.

Garden sharing has shipped. The next identification work is multi-angle capture.

Garden sharing has shipped

Garden sharing landed in 1.2.0, the first big feature since launch. If you are going away, you can send a plain care list to whoever is watering your plants, pause your own reminders for the dates you are gone, and come back to a summary of what came up while you were away. It all runs on your phone, and it is all free. The release notes have the detail.

Multi-angle capture

Raychis identifies a plant from a single photo, on your device, in a second or two. That works well most of the time. But one photo gives the model one look at the plant, and a lot of plants look alike from one angle. When that look is ambiguous, the model has to commit on thin evidence.

We are not making the underlying model larger for now. That may change later, but the model that ships today is close to the limit of what its current design can do. Past a certain point, a bigger model adds accuracy too small to notice while making the app slower and heavier. So the next accuracy work is about giving the model more to look at.

Multi-angle capture steps in when a first identification comes back unsure. Rather than leave you with a shaky single guess, Raychis asks for a few more photos of different parts of the plant. Those prompts ask for a close-up of one leaf, the whole plant in its pot, the newest growth at the tip, and the underside of a leaf. Each photo runs through the same on-device model on its own. Raychis then looks at which species the photos agree on and shows you that. Flowering and woody plants get a flower or bark card in place of one of these, since that is where their diagnostic detail sits.

Why more angles help the identification

Different parts of a plant carry different clues. A leaf close-up shows the veins and texture. The whole plant shows its shape and habit, and the underside of a leaf often shows what the top hides. Two species that look identical from one angle are frequently obvious apart from another, so more angles give the model more of the features that separate them.

Agreement between the photos is a second clue. When several photos of the same plant independently land on the same species, that is a stronger signal than any one photo on its own. When they disagree, the first guess was worth a second look. Raychis weighs the photos by how consistently they point the same way, rather than taking the most confident single shot.

What we will and will not claim

The confidence figure Raychis shows you still comes from a single photo. We do not inflate it because you took more pictures, so it still means what it says. We are not putting a percentage on the improvement either, because how much a second photo helps depends on how different it is from the first. Another shot of the same leaf from the same angle adds almost nothing. A photo of a different part adds a lot. That is why the flow asks for different parts of the plant rather than more of the same.

Multi-angle capture is in build now. It will be free, and it will run entirely on your device, like the rest of Raychis. No account, no upload. We will post here when it ships, and support is the fastest way to tell us what is working and what is not.