Update app.py

app.py

@@ -1,23 +1,21 @@
-import gradio as gr
 from transformers import pipeline
+import gradio as gr
 from PIL import Image, ImageEnhance
 from pypdf import PdfReader
 
+# ===================== 1. 加载模型 =====================
 
-# ===================== 1. 加载模型（适配 CPU Basic） =====================
-
-# 文档 / 图表问答，用 Donut 做图表理解
-chart_qa_pipe = pipeline(
-    "document-question-answering",
-    model="naver-clova-ix/donut-base-finetuned-docvqa",
+# 图像理解：BLIP（不会报 sentencepiece 错，对 CPU 友好）
+vision_pipe = pipeline(
+    "image-to-text",
+    model="salesforce/blip-image-captioning-base",
     device=-1
 )
 
-# 文本生成：学术解释 / caption / 摘要
-# 如果觉得慢，可以把 large-ssm 换成 flan-t5-base
+# 文本生成（学术解释）
 text_pipe = pipeline(
     "text2text-generation",
-    model="google/flan-t5-large-ssm",
+    model="google/flan-t5-base",  # 注意：换回 base，更轻量
     max_length=512,
     device=-1
 )
@@ -101,32 +99,16 @@ def summarise_context(snippet: str) -> str:
 
 # ===================== 3. 图表理解：用 Donut 做 QA =====================
 
-def describe_chart_with_qa(image: Image.Image) -> str:
+def describe_chart_with_qa(image):
     """
-    利用 document-question-answering 管线，对图表提问。
-    Donut 会尝试读坐标轴 / 图例 / 文本框，从而给出一句话描述。
+    使用 BLIP 做图像描述，作为图表结构的原始信息。
     """
-    if image is None:
-        return ""
-
-    question = (
-        "This is a scientific bar chart. Describe what is on the horizontal axis "
-        "and vertical axis, what groups or treatments are compared, and the overall trend."
-    )
-
     try:
-        result = chart_qa_pipe(image=image, question=question)
-        # 有的实现返回 list，有的返回 dict，这里做个兼容
-        if isinstance(result, list):
-            answer = result[0].get("answer", "")
-        else:
-            answer = result.get("answer", "")
-        return answer.strip()
-    except Exception as e:
-        print("Chart QA error:", e)
+        result = vision_pipe(image)[0]["generated_text"]
+        return result
+    except:
         return ""
 
-
 # ===================== 4. 核心工作流：ChartSmith v2 =====================
 
 def analyze_and_explain_v2(image, keywords, language, paper_file):