diff --git a/lib/ai/providers/vip98k.ts b/lib/ai/providers/vip98k.ts
index 1e5ba33cb..4e250175b 100644
--- a/lib/ai/providers/vip98k.ts
+++ b/lib/ai/providers/vip98k.ts
@@ -13,6 +13,7 @@ import { VIP98K_BASE_URL, normalizeVip98kBaseUrl } from "@/lib/ai/vip98k-models"
 import { shouldResolveVip98kResultDelivery } from "@/lib/ai/vip98k-result-delivery";
 import {
   getStorageObjectKeyFromPublicUrl,
+  getVip98kStagedMediaRetryDelay,
   shouldStageVip98kReferenceUrl,
 } from "@/lib/ai/vip98k-reference-staging";
 import { createLogger } from "@/lib/observability/logger";
@@ -499,6 +500,14 @@ function getUrlHost(value: string) {
   }
 }
 
+function isStagedReferenceUrl(value: string) {
+  try {
+    return new URL(value).pathname.startsWith(`/${REFERENCE_STAGING_KEY_PREFIX}`);
+  } catch {
+    return false;
+  }
+}
+
 async function downloadReferenceImage(mediaUrl: string) {
   let lastError: unknown;
   for (let attempt = 1; attempt <= REFERENCE_DOWNLOAD_ATTEMPTS; attempt += 1) {
@@ -791,25 +800,55 @@ async function submitTask(params: {
   const baseURL = normalizeVip98kBaseUrl(params.baseURL);
   const path = String(params.path || "/api/external/task/generate").trim();
   if (!path.startsWith("/")) throw new Error("VIP98K invocation path is invalid");
-  const response = await fetch(`${baseURL}${path}`, {
-    method: String(params.method || "POST").toUpperCase(),
-    headers: {
-      Authorization: `Bearer ${params.apiKey}`,
-      "Content-Type": params.contentType || "application/json",
-    },
-    body: JSON.stringify(params.body),
-    signal: AbortSignal.timeout(120_000),
-  });
-  const payload = await parseJsonResponse(response);
-  if (!response.ok) {
-    throw new Error(`${extractErrorMessage(payload)} (HTTP ${response.status})`);
+  const stagedReferences = Array.isArray(params.body.file_urls)
+    ? params.body.file_urls.filter(
+        (value): value is string =>
+          typeof value === "string" && isStagedReferenceUrl(value),
+      )
+    : [];
+  const maxAttempts = stagedReferences.length > 0 ? 3 : 1;
+
+  for (let attempt = 1; attempt <= maxAttempts; attempt += 1) {
+    const response = await fetch(`${baseURL}${path}`, {
+      method: String(params.method || "POST").toUpperCase(),
+      headers: {
+        Authorization: `Bearer ${params.apiKey}`,
+        "Content-Type": params.contentType || "application/json",
+      },
+      body: JSON.stringify(params.body),
+      signal: AbortSignal.timeout(120_000),
+    });
+    const payload = await parseJsonResponse(response);
+    if (response.ok) {
+      return {
+        payload,
+        taskId: extractTaskId(payload),
+        resultPath: extractResultPath(payload),
+        outputs: collectOutputUrls(payload),
+      };
+    }
+
+    const errorMessage = extractErrorMessage(payload);
+    const retryDelayMs = getVip98kStagedMediaRetryDelay({
+      status: response.status,
+      message: errorMessage,
+      attempt,
+      maxAttempts,
+      stagedReferenceCount: stagedReferences.length,
+    });
+    if (retryDelayMs !== null) {
+      logger.warn("staged reference not ready for provider; retrying submission", {
+        attempt,
+        retryDelayMs,
+        stagedReferenceCount: stagedReferences.length,
+      });
+      await delay(retryDelayMs);
+      continue;
+    }
+    throw new Error(`${errorMessage} (HTTP ${response.status})`);
   }
-  return {
-    payload,
-    taskId: extractTaskId(payload),
-    resultPath: extractResultPath(payload),
-    outputs: collectOutputUrls(payload),
-  };
+
+  throw new Error("VIP98K request failed after media readiness retries");
 }
 
 export type Vip98kExternalTaskSubmission = {
diff --git a/lib/ai/vip98k-reference-staging.ts b/lib/ai/vip98k-reference-staging.ts
index ee7d5e5a5..1c49e111e 100644
--- a/lib/ai/vip98k-reference-staging.ts
+++ b/lib/ai/vip98k-reference-staging.ts
@@ -1,3 +1,5 @@
+import { isMediaInputFailure } from "@/lib/ai/runtime/request-failure-classification";
+
 function parseUrl(value: string) {
   try {
     return new URL(String(value || "").trim());
@@ -49,3 +51,21 @@ export function getStorageObjectKeyFromPublicUrl(
     return encodedKey;
   }
 }
+
+export function getVip98kStagedMediaRetryDelay(params: {
+  status: number;
+  message: string;
+  attempt: number;
+  maxAttempts: number;
+  stagedReferenceCount: number;
+}) {
+  if (
+    params.status !== 400 ||
+    params.stagedReferenceCount < 1 ||
+    params.attempt >= params.maxAttempts ||
+    !isMediaInputFailure(params.message)
+  ) {
+    return null;
+  }
+  return params.attempt * 20_000;
+}
diff --git a/tests/routes/vip98k-reference-routing.test.ts b/tests/routes/vip98k-reference-routing.test.ts
index 289cc289b..3776e08fb 100644
--- a/tests/routes/vip98k-reference-routing.test.ts
+++ b/tests/routes/vip98k-reference-routing.test.ts
@@ -9,8 +9,11 @@ import {
   buildVip98kModelPresets,
   isVip98kImageModelVisible,
 } from "@/lib/ai/vip98k-models";
-import { shouldStageVip98kReferenceUrl } from "@/lib/ai/vip98k-reference-staging";
-import { getStorageObjectKeyFromPublicUrl } from "@/lib/ai/vip98k-reference-staging";
+import {
+  getStorageObjectKeyFromPublicUrl,
+  getVip98kStagedMediaRetryDelay,
+  shouldStageVip98kReferenceUrl,
+} from "@/lib/ai/vip98k-reference-staging";
 import { shouldResolveVip98kResultDelivery } from "@/lib/ai/vip98k-result-delivery";
 
 const textToImageInvocation: Vip98kTaskInvocation = {
@@ -25,7 +28,7 @@ const textToImageInvocation: Vip98kTaskInvocation = {
 };
 
 test.describe("VIP98K reference-image routing", () => {
-  test("stages references from the configured default storage into backup storage", () => {
+  test("identifies default-storage references that need provider staging", () => {
     expect(
       shouldStageVip98kReferenceUrl({
         mediaUrl: "https://default.example/uploads/reference.jpg",
@@ -55,6 +58,27 @@ test.describe("VIP98K reference-image routing", () => {
     ).toBe("folder name/reference.jpg");
   });
 
+  test("retries transient media rejection while a staged URL propagates", () => {
+    expect(
+      getVip98kStagedMediaRetryDelay({
+        status: 400,
+        message: "The media request is invalid.",
+        attempt: 1,
+        maxAttempts: 3,
+        stagedReferenceCount: 2,
+      }),
+    ).toBe(20_000);
+    expect(
+      getVip98kStagedMediaRetryDelay({
+        status: 400,
+        message: "The media request is invalid.",
+        attempt: 3,
+        maxAttempts: 3,
+        stagedReferenceCount: 2,
+      }),
+    ).toBeNull();
+  });
+
   test("shows the newly published Nano Banana Pro image editor", () => {
     expect(
       isVip98kImageModelVisible({