diff --git a/lib/ai/model-classification.ts b/lib/ai/model-classification.ts
index f0fd20229..8e45c9552 100644
--- a/lib/ai/model-classification.ts
+++ b/lib/ai/model-classification.ts
@@ -57,7 +57,6 @@ const VIDEO_MODE_KEYS = new Set([
   "audio-to-video",
   "first-frame",
   "first-last-frame",
-  "reference-images",
 ]);
 
 const VIDEO_KIND_KEYS = new Set([
diff --git a/tests/routes/model-classification.test.ts b/tests/routes/model-classification.test.ts
index 0289d8932..c0b50796b 100644
--- a/tests/routes/model-classification.test.ts
+++ b/tests/routes/model-classification.test.ts
@@ -18,6 +18,34 @@ const atlasHunyuan3D = {
 };
 
 test.describe("AI model media classification", () => {
+  test("keeps multi-reference BananaImage models classified as image models", () => {
+    const descriptor = {
+      providerType: "bananaimage",
+      modelId: "gemini-3-pro-image-preview",
+      capabilities: {
+        supportedModes: [
+          "text-to-image",
+          "image-to-image",
+          "reference-images",
+        ],
+      },
+      supportsImageGen: true,
+    };
+
+    expect(isVideoModelDescriptor(descriptor)).toBe(false);
+    expect(isImageGenerationModelDescriptor(descriptor)).toBe(true);
+  });
+
+  test("still recognizes explicit video generation modes", () => {
+    expect(
+      isVideoModelDescriptor({
+        providerType: "custom",
+        modelId: "custom-video-model",
+        capabilities: { supportedModes: ["text-to-video"] },
+      })
+    ).toBe(true);
+  });
+
   test("keeps Atlas 3D entries out of image and video categories", () => {
     expect(inferAtlasCloudMediaKind(atlasHunyuan3D)).toBe("3d");