diff --git a/server-rs/crates/platform-image/src/image_provider/runtime/executor.rs b/server-rs/crates/platform-image/src/image_provider/runtime/executor.rs
index 8cec1ae9d..0f25bcd58 100644
--- a/server-rs/crates/platform-image/src/image_provider/runtime/executor.rs
+++ b/server-rs/crates/platform-image/src/image_provider/runtime/executor.rs
@@ -44,7 +44,13 @@ pub async fn create_image_generation(
     create_image_generation_with_model(
         http_client,
         settings,
-        GPT_IMAGE_2_5_GENERATION_MODEL,
+        // 带参考图时执行路径会走 /v1/images/edits，必须与任务边界的
+        // `editor_image_generation_concrete_model` 判据一致，提交编辑 concrete model。
+        if reference_images.is_empty() {
+            GPT_IMAGE_2_5_GENERATION_MODEL
+        } else {
+            GPT_IMAGE_2_5_EDIT_MODEL
+        },
         prompt,
         negative_prompt,
         size,
diff --git a/server-rs/crates/platform-image/tests/image_provider.rs b/server-rs/crates/platform-image/tests/image_provider.rs
index ce56436fb..c1ca9439d 100644
--- a/server-rs/crates/platform-image/tests/image_provider.rs
+++ b/server-rs/crates/platform-image/tests/image_provider.rs
@@ -639,6 +639,38 @@ async fn image_edit_does_not_cross_model_when_unsupported() {
     server.abort();
 }
 
+#[tokio::test]
+async fn image_generation_with_references_uses_the_edit_concrete_model() {
+    let (base_url, server, requests) = start_http_response_sequence(vec![MockResponse {
+        status: "200 OK",
+        content_type: "application/json",
+        body: r#"{"data":[{"b64_json":"iVBORw0KGgpyZXN0"}]}"#,
+    }])
+    .await;
+    let settings = test_image_provider_settings(base_url);
+    let http_client = build_image_http_client(&settings).expect("client should build");
+
+    let generated = create_image_generation(
+        &http_client,
+        &settings,
+        "测试提示词",
+        None,
+        "1024x1024",
+        1,
+        &["data:image/png;base64,iVBORw0KGgpyZXN0".to_string()],
+        "测试 ImageProvider 图片生成失败",
+    )
+    .await
+    .expect("referenced generation should go through the edit endpoint");
+
+    assert_eq!(generated.images.len(), 1);
+    let requests = requests.lock().await;
+    assert_eq!(requests.len(), 1);
+    assert!(requests[0].starts_with("POST /v1/images/edits"));
+    assert!(requests[0].contains(&format!("\r\n\r\n{GPT_IMAGE_2_5_EDIT_MODEL}\r\n")));
+    server.abort();
+}
+
 #[tokio::test]
 async fn image_generation_does_not_fallback_on_auth_failure() {
     let (base_url, server, requests) = start_http_response_sequence(vec![MockResponse {