diff --git a/components/messages.tsx b/components/messages.tsx
index d577ae8..4d8a0ca 100644
--- a/components/messages.tsx
+++ b/components/messages.tsx
@@ -147,6 +147,12 @@ function PureMessages({
     imageGenerationPhase === "queued" ||
     imageGenerationPhase === "processing" ||
     imageGenerationPhase === "finalizing";
+  const stripInternalMarkers = (text: string) =>
+    text
+      .replace(/\n*<!-- GENERATING_IMAGE -->\n*/g, "")
+      .replace(/\n*<!-- IMAGE_GENERATION_START -->\n*/g, "")
+      .replace(/\n*<!-- IMAGE_GENERATION_END -->\n*/g, "")
+      .trim();
 
   useEffect(() => {
     return () => {
@@ -288,11 +294,16 @@ function PureMessages({
       // - 收到 AI 文字回复后 3 秒: true（开始显示生图占位图）
       // - 收到图片消息: false（隐藏占位图）
       
-      // 检查最后一条消息是否是 AI 的文字回复（包含 GENERATING_IMAGE 标记）
-      const hasAssistantTextReply = messages.some(msg => 
-        msg.role === "assistant" && msg.parts?.some(
-          (part: any) => part.type === "text" && part.text?.includes("<!-- GENERATING_IMAGE -->")
-        )
+      // 检查当前轮次是否已收到 AI 文字回复（只看最后一条用户消息之后）
+      const lastUserIndex = messages.findLastIndex((m) => m.role === "user");
+      const assistantMessagesAfterUser = lastUserIndex >= 0
+        ? messages.slice(lastUserIndex + 1).filter((m) => m.role === "assistant")
+        : messages.filter((m) => m.role === "assistant");
+      const hasAssistantTextReply = assistantMessagesAfterUser.some((msg) =>
+        msg.parts?.some((part: any) => {
+          if (part.type !== "text" || !part.text) return false;
+          return stripInternalMarkers(part.text).length > 0;
+        })
       );
       
       if (hasImageContent) {
@@ -353,22 +364,19 @@ function PureMessages({
     const isAutoModeWithImage = selectedModelId === "auto" && 
       lastUserMessage?.parts?.some((p: any) => p.type === "file" || p.type === "image");
     
-    const hasAssistantTextReplyWithContent = messages.some(msg =>
-      msg.role === "assistant" && msg.parts?.some((part: any) => {
+    const lastUserIndex = messages.findLastIndex((m) => m.role === "user");
+    const assistantMessagesAfterUser = lastUserIndex >= 0
+      ? messages.slice(lastUserIndex + 1).filter((m) => m.role === "assistant")
+      : messages.filter((m) => m.role === "assistant");
+    const hasAssistantTextReplyWithContent = assistantMessagesAfterUser.some((msg) =>
+      msg.parts?.some((part: any) => {
         if (part.type !== "text" || !part.text) return false;
-        const cleanedText = part.text
-          .replace(/\n*<!-- GENERATING_IMAGE -->\n*/g, "")
-          .replace(/\n*<!-- IMAGE_GENERATION_START -->\n*/g, "")
-          .replace(/\n*<!-- IMAGE_GENERATION_END -->\n*/g, "")
-          .trim();
-        return cleanedText.length > 0;
+        return stripInternalMarkers(part.text).length > 0;
       })
     );
     
     let hasImageContent = false;
     if (isAutoModeWithImage) {
-      const lastUserIndex = messages.findLastIndex(m => m.role === "user");
-      const assistantMessagesAfterUser = messages.slice(lastUserIndex + 1).filter(m => m.role === "assistant");
       hasImageContent = assistantMessagesAfterUser.some(msg => 
         msg.parts?.some((part: any) => part.type === "text" && part.text && (
           part.text.includes("![thumb|") || 
@@ -587,19 +595,24 @@ function PureMessages({
                 (part: any) => part.type === "text" && part.text?.includes("<!-- IMAGE_GENERATION_END -->")
               )
             );
-            const hasAssistantTextReplyWithContent = messages.some(msg =>
-              msg.role === "assistant" && msg.parts?.some((part: any) => {
+            const assistantMessagesAfterUser = lastUserIndexForMarkers >= 0
+              ? messages.slice(lastUserIndexForMarkers + 1).filter((msg) => msg.role === "assistant")
+              : messages.filter((msg) => msg.role === "assistant");
+            const hasAssistantTextReplyWithContent = assistantMessagesAfterUser.some((msg) =>
+              msg.parts?.some((part: any) => {
                 if (part.type !== "text" || !part.text) return false;
-                const cleanedText = part.text
-                  .replace(/\n*<!-- GENERATING_IMAGE -->\n*/g, "")
-                  .replace(/\n*<!-- IMAGE_GENERATION_START -->\n*/g, "")
-                  .replace(/\n*<!-- IMAGE_GENERATION_END -->\n*/g, "")
-                  .trim();
-                return cleanedText.length > 0;
+                return stripInternalMarkers(part.text).length > 0;
               })
             );
+            const shouldWaitForAgentReplyBeforeImagePlaceholder =
+              userMessageHasImage &&
+              (isAutoMode || showGemini || showVolcano || showAlibabaImage) &&
+              !hasAssistantTextReplyWithContent;
             
             if (showVideo) return <QianwenVideoLoading />;
+            if (shouldWaitForAgentReplyBeforeImagePlaceholder) {
+              return <ThinkingMessage />;
+            }
             if (showGemini) return <GeminiImageLoading imageCount={imageCount} />;
             if (showVolcano) return <VolcanoVisionLoading imageCount={imageCount} />;
             if (showAlibabaImage) {