Update app.py

app.py

@@ -546,15 +546,22 @@ def process_image_analysis_stream(image_dict, user_prompt, max_new_tokens, top_k
     
     # 检查图像输入
     if image_dict is None or image_dict.get('image') is None:
-        yield [["请先上传图片！", "我需要图片才能进行分析。"]], None
+        yield [[{"role": "user", "content": "请先上传图片！"}, {"role": "assistant", "content": "我需要图片才能进行分析。"}]], None
         return
     
     # 处理图像
     image = image_dict['image']
     if not isinstance(image, str):
         import tempfile
+        from PIL import Image as PILImage
         temp_path = os.path.join(tempfile.gettempdir(), f"temp_image_{hash(str(image))}.png")
-        image.save(temp_path)
+        # 将 numpy 数组转换为 PIL Image
+        if hasattr(image, 'shape'):  # 确认是 numpy 数组
+            pil_img = PILImage.fromarray(image)
+            pil_img.save(temp_path)
+        else:
+            # 如果已经是 PIL Image
+            image.save(temp_path)
         image = temp_path
     
     # 处理提示词
@@ -631,14 +638,14 @@ def process_image_analysis_stream(image_dict, user_prompt, max_new_tokens, top_k
     thread.start()
     
     # 初始化聊天历史
-    chat_history = [[user_prompt, ""]]
+    chat_history = [[{"role": "user", "content": user_prompt}, {"role": "assistant", "content": ""}]]
     yield chat_history, None
     
     # 流式接收输出，参考mm_qwen2vl.py的bot函数
     full_response = ""
     for new_token in streamer:
         full_response += new_token
-        chat_history[-1][1] = full_response
+        chat_history[-1][1]["content"] = full_response
         yield chat_history, None
     
     # 处理完成后的文件生成