diff --git a/components/sections/PuntosInflexion.tsx b/components/sections/PuntosInflexion.tsx
index 4ff8e39..3d4d36c 100644
--- a/components/sections/PuntosInflexion.tsx
+++ b/components/sections/PuntosInflexion.tsx
@@ -80,6 +80,17 @@ function Escenificado() {
   // frame is a dozen lines, needs no measurement pass from the library, and can
   // be verified in the browser, which the black box could not.
   const bruto = useMotionValue(0);
+
+  // Whether the stage is close enough to be worth paying for its photographs.
+  // Measured on the static export: the six of them are 1,24 MB and every one
+  // was downloading before the reader had scrolled a pixel — `loading="lazy"`
+  // did not save us, because Chrome's distance threshold is generous enough to
+  // swallow a section that starts under 2.000px down. Two megabytes before the
+  // first gesture, on a page that gets opened from a WhatsApp link on mobile
+  // data.
+  const [cerca, setCerca] = useState(false);
+  const yaCerca = useRef(false);
+
   useEffect(() => {
     let pendiente = 0;
     const medir = () => {
@@ -89,6 +100,12 @@ function Escenificado() {
       const { top, height } = el.getBoundingClientRect();
       const recorrido = height - window.innerHeight;
       bruto.set(recorrido > 0 ? Math.min(1, Math.max(0, -top / recorrido)) : 0);
+      // Latch, never unset: the images are cached by then, and flipping it back
+      // would tear down the stage every time the reader scrolled past.
+      if (!yaCerca.current && top < window.innerHeight * 1.25) {
+        yaCerca.current = true;
+        setCerca(true);
+      }
     };
     const alScroll = () => {
       if (!pendiente) pendiente = requestAnimationFrame(medir);
@@ -138,8 +155,11 @@ function Escenificado() {
       <div ref={contenedor} className="relative" style={{ height: `${TOTAL * 100}vh` }}>
         <div data-escena={activo} data-tiempo={tiempo} className="sticky top-0 h-screen overflow-hidden bg-azul-profundo">
           {/* The photograph of whichever point is on screen, crossfading with
-              the beat. Only the active one is opaque; the rest sit at zero, so
-              the browser still holds them decoded and the swap costs nothing. */}
+              the beat. Only the neighbours of the active scene are mounted:
+              the next one is in the document a whole scene before it is needed,
+              which is time enough to arrive, and the far ones cost nothing.
+              Mounting all six was 1,24 MB spent up front for five pictures the
+              reader might never reach. */}
           {items.map((item, i) => (
             <motion.div
               key={item.name}
@@ -148,9 +168,11 @@ function Escenificado() {
               transition={{ duration: 0.6, ease: [0.22, 1, 0.36, 1] }}
               className="absolute inset-0"
             >
-              <motion.div style={{ y: fotoY }} className="absolute -inset-y-24 inset-x-0">
-                <FondoFoto src={FOTOS[i]} wash="" className="inset-0" />
-              </motion.div>
+              {cerca && Math.abs(i - activo) <= 1 && (
+                <motion.div style={{ y: fotoY }} className="absolute -inset-y-24 inset-x-0">
+                  <FondoFoto src={FOTOS[i]} wash="" className="inset-0" />
+                </motion.div>
+              )}
             </motion.div>
           ))}