<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Alexander G</title>
    <description>The latest articles on DEV Community by Alexander G (@alexander_g_9261e19aef5c3).</description>
    <link>https://dev.to/alexander_g_9261e19aef5c3</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4252415%2F5acad2fe-708e-4e0f-9da6-1c7e86b9887a.jpg</url>
      <title>DEV Community: Alexander G</title>
      <link>https://dev.to/alexander_g_9261e19aef5c3</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/alexander_g_9261e19aef5c3"/>
    <language>en</language>
    <item>
      <title>Turning a sentence into a Wear OS watch face: what I learned building an AI face creator</title>
      <dc:creator>Alexander G</dc:creator>
      <pubDate>Wed, 18 Nov 2026 18:00:12 +0000</pubDate>
      <link>https://dev.to/alexander_g_9261e19aef5c3/turning-a-sentence-into-a-wear-os-watch-face-what-i-learned-building-an-ai-face-creator-87o</link>
      <guid>https://dev.to/alexander_g_9261e19aef5c3/turning-a-sentence-into-a-wear-os-watch-face-what-i-learned-building-an-ai-face-creator-87o</guid>
      <description>&lt;p&gt;I make Watchland, a watch face app for Wear OS. For the last few months I've been building an AI face creator for it: you type a sentence like "neon digital time glowing cyan and magenta over a rainy city at night", it designs 4 faces, you change anything by asking ("gold hands", "add seconds"), and it installs on your watch.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fb01kd1166fducmdobqk0.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fb01kd1166fducmdobqk0.jpg" alt="The app's second step: four AI designs generated from one prompt" width="800" height="1000"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Generating the pictures turned out to be the easy part. Here's what took the time.&lt;/p&gt;

&lt;h2&gt;
  
  
  The watch doesn't take pictures, it takes Watch Face Format
&lt;/h2&gt;

&lt;p&gt;New Wear OS watch faces have to use Watch Face Format (WFF): a declarative XML file plus images, rendered by the system. No code runs on the watch. So the AI can't just hand over an image. The hour hand, the minute hand, the date window and the step counter all have to be separate parts the system can move and fill in.&lt;/p&gt;

&lt;p&gt;So the AI doesn't write XML. It produces a recipe: the dial art, the hand style, where the complications go, which fonts and colours. A deterministic builder turns that recipe into WFF XML. Every face goes through two validators before it can ship. If the model dreams up something WFF can't express, the builder refuses it rather than shipping a broken face.&lt;/p&gt;

&lt;h2&gt;
  
  
  The preview lies
&lt;/h2&gt;

&lt;p&gt;What looks right in a flat preview isn't always what the watch draws. Round screens clip corners, the always-on mode needs its own dim version, and complications show real data of different lengths. So every face also gets rendered on a Wear OS emulator before I trust it, and some get checked on a real watch.&lt;/p&gt;

&lt;h2&gt;
  
  
  Animation is maths in XML
&lt;/h2&gt;

&lt;p&gt;WFF has expressions, so you can animate without code. A tourbillon cage that turns once a minute is a rotation driven by the seconds value. One gotcha: the trigonometry functions work in radians, not degrees. My first orbiting balance wheel flew off in a very wrong direction.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fz9s1i4phcq6k1spos87g.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fz9s1i4phcq6k1spos87g.jpg" alt="A photo-real tourbillon dial made by the AI face creator" width="800" height="1422"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Picking the image model was a grading job
&lt;/h2&gt;

&lt;p&gt;Photo-real dials were the hardest style. I graded about a dozen image models on the same set of test dials, scoring things like straight indices, readable numerals and no invented logos, and looked at cost per image alongside quality. The cheapest good model and the best-looking model were not the same one, and the gap mattered at scale.&lt;/p&gt;

&lt;h2&gt;
  
  
  Every click costs money
&lt;/h2&gt;

&lt;p&gt;Each AI image has a real cost, so "unlimited for subscribers" was never going to work. The creator runs on credits from day one. The rule I set myself: the user is only charged what the button showed before they pressed it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where it is now
&lt;/h2&gt;

&lt;p&gt;It's in a small closed beta, by invitation, while I fix what testers find. It needs a Wear OS 6 watch and an Android phone. If you're curious, there's a launch list here: &lt;a href="https://watchland.ca/ai-face-creator/" rel="noopener noreferrer"&gt;https://watchland.ca/ai-face-creator/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Happy to answer questions about WFF, the builder, or the model grading in the comments.&lt;/p&gt;

</description>
      <category>showdev</category>
      <category>android</category>
      <category>wearos</category>
      <category>ai</category>
    </item>
  </channel>
</rss>
