diff --git a/docker/frontend/src/engine/inpaint.ts b/docker/frontend/src/engine/inpaint.ts
index 832c00c..d76db54 100644
--- a/docker/frontend/src/engine/inpaint.ts
+++ b/docker/frontend/src/engine/inpaint.ts
@@ -71,6 +71,23 @@ async function warmsUp(ort: typeof Ort, session: Ort.InferenceSession): Promise<
   }
 }
 
+// Whether the page has a device of its own to run the graph on. Asked of the
+// browser rather than of the runtime: creating a session of any kind — the
+// WebGPU one included — builds the wasm backend underneath it, and
+// `env.wasm.proxy` is read at that moment. A proxy set after it is a proxy that
+// never took, and every run of the fallback then ends the way it did the first
+// time this was tried: `no available backend found. ERR: [wasm] Error: worker
+// not ready`. So the question is put before anything is created.
+async function hasDevice(): Promise<boolean> {
+  const gpu = (navigator as Navigator & { gpu?: { requestAdapter(): Promise<unknown> } }).gpu;
+  if (!gpu) return false;
+  try {
+    return !!(await gpu.requestAdapter());
+  } catch {
+    return false;
+  }
+}
+
 // WebGPU first, the processor as the answer to every way that can fail: no
 // adapter, no `shader-f16` needed here but an op the provider never
 // implemented, a driver that opens a session and dies on the first dispatch.
@@ -99,6 +116,16 @@ function load(): Promise<Loaded> {
       ort.env.wasm.numThreads = self.crossOriginIsolated
         ? Math.min(4, navigator.hardwareConcurrency || 1)
         : 1;
+      // The fallback's turn is taken in a worker of the page's own, wherever
+      // there is no device to take it on: a wasm run is seconds of arithmetic
+      // that never hands the thread back in the middle of it, and a thread that
+      // cannot answer a finger for three seconds is a tab a phone takes away
+      // (measured: 3.5s of frozen page per repair, on a desktop). Proxied, the
+      // page keeps its frames and its touches while the repair is worked out,
+      // and the repair lands under the spinner the studio already shows. It is
+      // left off where a device is there, because the WebGPU path is a device of
+      // its own and must be built on the page's own thread.
+      ort.env.wasm.proxy = !(await hasDevice());
       ort.env.webgpu.powerPreference = 'high-performance';
       const session = await openSession(ort);
       return { ort, session };