Do Pixelmator Pro ML models improve with use?

Talk about Pixelmator Pro, share tips & tricks, tutorials, and other resources.
User avatar

2020-05-19 06:53:12

I wonder whether Pixelmator Pro ML models improve with use? In other words, is on-device personalization of these models enabled?
User avatar

2020-05-19 12:55:00

Hi, Yuka.

Unfortunately, there is no on-device personalization of the ML Models in Pixelmator Pro. Nevertheless, we do tweak the models for some Pixelmator Pro releases.

However, some users may feel that the models become better as they get used to ML functions and know what to expect from ML and when to use it.

Meantime we investigating the ways to do some on-device model tweaking/personalization.
User avatar

2020-05-19 14:49:05

This surprises me a bit, because Core ML models are capable of on-device personalization for a year now. The bigger question here, however, is whether users will want this type of model. You do a great job improving your app and, obviously, want users to use the latest version. But, once user updates to the latest version, the personalized model is gone and he/she will need to start from scratch personalizing it. Unless Apple introduces some way to store this model on iCloud or something and keep in sync with updated versions of the app, this looks like an impossible problem to solve. What is your opinion on this?
User avatar

2020-05-21 19:57:28

The capability of Core ML model personalization (on-device training) does not solve the task of a certain function personalization in our app. Apple frameworks simplify some aspects of the development process, but we still need to spend months developing ML functions in our apps. You can think of the following analogy: computers have hardware capabilities and software libraries to display 3D graphics, however, it takes months or even years to develop a video game.

Taking technology and human resources aspects aside, in order for personalization to work the app should know what is good for a particular user and what is not. Naive way to solve this problem will look like this in our case: «Do you like how the colors are automatically adjusted in your image? If not, then please adjust the colors the way you like and press the "Train" button».

We can invent some non-intrusive ways of personalization. For example, if the user adjusts colors manually after automatic color adjustment, then such manual adjustments can be added to personalization dataset. So, next time the ML algorithm will try to work better for this user, making him to do less manual tweaking. Such personalization dataset can be stored in application settings and will survive app update.

Also people afraid the word "personalization". When they hear it, they immediately think about things like "personal data collection", "tracking", "advertising", and so on.