diff --git a/vite.config.js b/vite.config.js
index e499635..44da18d 100644
--- a/vite.config.js
+++ b/vite.config.js
@@ -119,6 +119,14 @@ export default defineConfig({
         // which precaching would undo. They're served via the
         // `sticker-library` runtimeCaching rule below instead.
         globIgnores: ['**/stickers/**'],
+        // Note: we deliberately do NOT raise `maximumFileSizeToCacheInBytes`
+        // here. The default 2 MiB cap is a useful regression alarm — if
+        // a future change pushes any chunk past it, the build fails loudly
+        // and we know to either split further or investigate why a chunk
+        // grew. Earlier this file overrode the cap to 5 MiB because the
+        // pre-split bundle was 2.32 MB; that override was removed once
+        // `build.rollupOptions.output.manualChunks` (below in this config)
+        // brought every chunk under the default cap.
         runtimeCaching: [
           {
             urlPattern: /^https:\/\/cdn\.huggingface\.co\/.*/i,
@@ -198,5 +206,95 @@ export default defineConfig({
       '/exports': { target: 'http://localhost:3001', changeOrigin: true },
     },
   },
-  build: { outDir: 'dist' },
+  build: {
+    outDir: 'dist',
+    // ────────────────────────────────────────────────────────────────
+    // Manual chunk splitting
+    // ────────────────────────────────────────────────────────────────
+    //
+    // Why we split rather than ship one bundle
+    // ─────────────────────────────────────────
+    // 1. Workbox precache cap. The PWA plugin's default
+    //    `maximumFileSizeToCacheInBytes` is 2 MiB. A single bundle
+    //    containing Konva + react-konva + @huggingface/transformers +
+    //    react + filerobot + everything else easily clears 2.3 MB,
+    //    which fails the precache step. Splitting keeps every chunk
+    //    comfortably under the cap without needing to override it.
+    //
+    // 2. Browser cache reuse across deploys. When app code changes
+    //    (the common case during active development), only the
+    //    `index` chunk's hash flips; vendor chunks (Konva, React,
+    //    transformers) keep their old hash and stay cached in users'
+    //    browsers. That means returning users only re-download the
+    //    relatively-small app chunk, not the 1-MB-each library
+    //    chunks that haven't changed.
+    //
+    // 3. Parallel download. Browsers can fetch multiple chunks
+    //    concurrently on HTTP/2, so loading 4 × 600 KB chunks is
+    //    typically faster than 1 × 2.3 MB chunk on broadband, even
+    //    before considering caching.
+    //
+    // What goes where
+    // ───────────────
+    //   transformers/      @huggingface/transformers \u2014 the in-browser
+    //                     ML runtime used for background removal.
+    //                     Biggest single dependency. Only the
+    //                     inference engine is bundled here; the
+    //                     actual ML model weights stream from the
+    //                     huggingface CDN at runtime (see
+    //                     `transformers-models` / `transformers-lfs`
+    //                     runtimeCaching rules above for offline
+    //                     re-use).
+    //   konva/            konva + react-konva + use-image \u2014 the canvas
+    //                     stack. All three are needed together
+    //                     because react-konva wraps konva and
+    //                     use-image wraps it for React consumption.
+    //                     Grouping prevents one from being inlined
+    //                     into a chunk that doesn't otherwise need
+    //                     the others.
+    //   react-vendor/     react + react-dom \u2014 stable, every visitor
+    //                     needs them, perfect for long-term caching.
+    //   filerobot/        react-filerobot-image-editor \u2014 mid-size,
+    //                     used only when the user opens the advanced
+    //                     image editor. Worth its own chunk so the
+    //                     rest of the app doesn't pay for it.
+    //   index/            Everything else: app code, react-select,
+    //                     styled-components, zod, uuid, plus all
+    //                     the small utilities. This stays the
+    //                     "main" chunk that changes with each deploy.
+    //
+    // What we deliberately did NOT split
+    // ──────────────────────────────────
+    // The smaller deps (react-select, styled-components, zod, uuid,
+    // @emotion/is-prop-valid) stay in the index chunk on purpose.
+    // Pulling them out would create chunks under 50 KB each, which
+    // adds HTTP overhead (per-request connection cost, separate
+    // hash entries in the precache manifest) without meaningfully
+    // helping cache reuse. The "many small chunks" antipattern is
+    // worse than a moderately-sized index chunk that includes the
+    // long tail.
+    //
+    // If a future change makes one of these grow significantly (say,
+    // a styled-components major upgrade that doubles its size), it
+    // becomes worth pulling into its own chunk \u2014 but only then.
+    //
+    // Maintenance: when adding a new heavy dependency
+    // ───────────────────────────────────────────────
+    // 1. Build and check `dist/assets/` for which chunk it landed in.
+    // 2. If it bloated the `index` chunk past comfort (rule of thumb:
+    //    if index alone is > 1 MiB), add a new entry to manualChunks
+    //    targeting that package name.
+    // 3. The 2 MiB workbox cap acts as the loud alarm \u2014 if you forget,
+    //    the build fails and tells you exactly which chunk overflowed.
+    rollupOptions: {
+      output: {
+        manualChunks: {
+          'transformers': ['@huggingface/transformers'],
+          'konva': ['konva', 'react-konva', 'use-image'],
+          'react-vendor': ['react', 'react-dom'],
+          'filerobot': ['react-filerobot-image-editor'],
+        },
+      },
+    },
+  },
 });