fix(heal): enable MI-GAN neural inpaint on mobile with bounded memory to prevent grey blob

docker/frontend/src/engine/inpaint.ts

@@ -133,6 +133,13 @@ async function hasDevice(): Promise<boolean> {
 // 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.
 async function openSession(ort: typeof Ort): Promise<Ort.InferenceSession> {
+  const mobile = isMobile();
+  const wasmOptions: Ort.InferenceSession.SessionOptions = {
+    executionProviders: ['wasm'],
+    enableCpuMemArena: !mobile,
+    enableMemPattern: !mobile,
+    graphOptimizationLevel: 'all',
+  };
   if (await hasDevice()) {
     try {
       const gpu = await ort.InferenceSession.create(MODEL_URL, { executionProviders: ['webgpu'] });
@@ -142,7 +149,7 @@ async function openSession(ort: typeof Ort): Promise<Ort.InferenceSession> {
       // Whatever went wrong, the wasm provider is the fallback and not a failure.
     }
   }
-  return ort.InferenceSession.create(MODEL_URL, { executionProviders: ['wasm'] });
+  return ort.InferenceSession.create(MODEL_URL, wasmOptions);
 }
 
 // Loads the runtime and the model once per page, and runs one warm-up pass so
@@ -160,9 +167,6 @@ function load(): Promise<Loaded> {
         : self.crossOriginIsolated
         ? Math.min(4, navigator.hardwareConcurrency || 1)
         : 1;
-      // On mobile and desktop browsers, proxy=false runs direct in-thread WASM execution,
-      // avoiding Web Worker proxy script loading failures and cross-origin worker restrictions.
-      ort.env.wasm.proxy = false;
       ort.env.webgpu.powerPreference = 'high-performance';
       const session = await openSession(ort);
       return { ort, session };
@@ -177,9 +181,7 @@ function load(): Promise<Loaded> {
 
 // Start fetching the model before the brush that needs it, so the visitor waits
 // for the repair and not for the download. Called when HEAL is switched on.
-// Skipped on iOS to avoid loading unnecessary 56MB into the mobile WebKit process.
 export function preloadInpaint(): void {
-  if (isIOS() || isMobile()) return;
   void load().catch(() => {});
 }
 
@@ -554,15 +556,6 @@ export async function inpaintCrop(
   height: number,
   hole: Uint8Array
 ): Promise<Uint8Array | null> {
-  // On mobile browsers (iOS WebKit and Android), compiling the 27MB ONNX model and
-  // executing WASM tensors alongside CanvasKit exceeds tab memory limits (~384MB ceiling),
-  // causing iOS Jetsam process termination and browser page reload.
-  // On mobile, fast local inpaint provides instantaneous (<5ms) Poisson rim relaxation
-  // and grain synthesis using <1MB RAM, ensuring 100% reliable repair without tab crashes.
-  if (isIOS() || isMobile()) {
-    return fastLocalInpaint(rgba, width, height, hole);
-  }
-
   const maxSide = isMobile() ? 256 : INPAINT_MAX_SIDE;
   let w = Math.max(8, Math.round(width));
   let h = Math.max(8, Math.round(height));
@@ -616,6 +609,9 @@ export async function inpaintCrop(
     // model reads under the hole comes from.
     const { src, hole: mask } = cropPlate(rgba, width, height, hole, w, h);
 
+    let in0: Ort.Tensor | null = null;
+    let in1: Ort.Tensor | null = null;
+    let out: Record<string, Ort.Tensor> | null = null;
     try {
       // Allow up to 15000ms for neural inpainting session to initialize and synthesize scene context
       const sessionPromise = load();
@@ -624,8 +620,11 @@ export async function inpaintCrop(
         new Promise<never>((_, reject) => setTimeout(() => reject(new Error('Load timeout')), 15000)),
       ]);
       const { ort, session } = await loadWithTimeout;
-      const in0 = new ort.Tensor('uint8', src, [1, 3, h, w]);
+      // Brief yield to keep UI responsive and prevent browser watchdog timeouts on mobile
-      const in1 = new ort.Tensor('uint8', mask, [1, 1, h, w]);
+      await new Promise((resolve) => setTimeout(resolve, 10));
+
+      in0 = new ort.Tensor('uint8', src, [1, 3, h, w]);
+      in1 = new ort.Tensor('uint8', mask, [1, 1, h, w]);
 
       const runWithTimeout = Promise.race([
         session.run({
@@ -637,7 +636,7 @@ export async function inpaintCrop(
         ),
       ]);
 
-      const out = await runWithTimeout;
+      out = await runWithTimeout;
       const tensor = out[session.outputNames[0]];
       const oplane = w * h;
       const data = tensor.data as Uint8Array;
@@ -648,12 +647,6 @@ export async function inpaintCrop(
         patch[i * 4 + 2] = data[2 * oplane + i];
         patch[i * 4 + 3] = 255;
       }
-      try {
-        in0.dispose?.();
-        in1.dispose?.();
-        tensor.dispose?.();
-        for (const k in out) out[k]?.dispose?.();
-      } catch {}
 
       if (patchCache.size > 50) {
         const first = patchCache.keys().next().value;
@@ -661,11 +654,20 @@ export async function inpaintCrop(
       }
       patchCache.set(key, patch.slice());
       return patch;
-    } catch {
+    } catch (err) {
+      console.warn('Inpaint model execution failed, falling back to local inpaint:', err);
       // Fallback: fast local inpaint with multi-step Poisson harmonic propagation.
       // Do NOT cache the fallback in patchCache so that once neural inpainting finishes
       // initializing, subsequent renders will execute neural context interpolation.
       return fastLocalInpaint(rgba, width, height, hole);
+    } finally {
+      try {
+        in0?.dispose?.();
+        in1?.dispose?.();
+        if (out) {
+          for (const k in out) out[k]?.dispose?.();
+        }
+      } catch {}
     }
   })();