better：新增log前缀映射，优化emoji的显示，加强了emoji的识别
This commit is contained in:
SengokuCola
@@ -6,6 +6,7 @@ import uuid
|
||||
import io
|
||||
import asyncio
|
||||
import numpy as np
|
||||
import jieba
|
||||
|
||||
from typing import Optional, Tuple
|
||||
from PIL import Image
|
||||
@@ -94,7 +95,7 @@ class ImageManager:
|
||||
logger.error(f"保存描述到数据库失败 (Peewee): {str(e)}")
|
||||
|
||||
async def get_emoji_description(self, image_base64: str) -> str:
|
||||
"""获取表情包描述，带查重和保存功能"""
|
||||
"""获取表情包描述，使用二步走识别并带缓存优化"""
|
||||
try:
|
||||
# 计算图片哈希
|
||||
# 确保base64字符串只包含ASCII字符
|
||||
@@ -107,33 +108,66 @@ class ImageManager:
|
||||
# 查询缓存的描述
|
||||
cached_description = self._get_description_from_db(image_hash, "emoji")
|
||||
if cached_description:
|
||||
return f"[表情包，含义看起来是：{cached_description}]"
|
||||
return f"[表情包：{cached_description}]"
|
||||
|
||||
# 调用AI获取描述
|
||||
# === 二步走识别流程 ===
|
||||
|
||||
# 第一步：VLM视觉分析 - 生成详细描述
|
||||
if image_format in ["gif", "GIF"]:
|
||||
image_base64_processed = self.transform_gif(image_base64)
|
||||
if image_base64_processed is None:
|
||||
logger.warning("GIF转换失败，无法获取描述")
|
||||
return "[表情包(GIF处理失败)]"
|
||||
prompt = "这是一个动态图表情包，每一张图代表了动态图的某一帧，黑色背景代表透明，使用1-2个词描述一下表情包表达的情感和内容，简短一些，输出一段平文本，只输出1-2个词就好，不要输出其他内容"
|
||||
description, _ = await self._llm.generate_response_for_image(prompt, image_base64_processed, "jpg")
|
||||
vlm_prompt = "这是一个动态图表情包，每一张图代表了动态图的某一帧，黑色背景代表透明，描述一下表情包表达的情感和内容，描述细节，从互联网梗,meme的角度去分析"
|
||||
detailed_description, _ = await self._llm.generate_response_for_image(vlm_prompt, image_base64_processed, "jpg")
|
||||
else:
|
||||
prompt = "图片是一个表情包，请用使用1-2个词描述一下表情包所表达的情感和内容，简短一些，输出一段平文本，只输出1-2个词就好，不要输出其他内容"
|
||||
description, _ = await self._llm.generate_response_for_image(prompt, image_base64, image_format)
|
||||
vlm_prompt = "这是一个表情包，请详细描述一下表情包所表达的情感和内容，描述细节，从互联网梗,meme的角度去分析"
|
||||
detailed_description, _ = await self._llm.generate_response_for_image(vlm_prompt, image_base64, image_format)
|
||||
|
||||
if description is None:
|
||||
logger.warning("AI未能生成表情包描述")
|
||||
return "[表情包(描述生成失败)]"
|
||||
if detailed_description is None:
|
||||
logger.warning("VLM未能生成表情包详细描述")
|
||||
return "[表情包(VLM描述生成失败)]"
|
||||
|
||||
# 第二步：LLM情感分析 - 基于详细描述生成简短的情感标签
|
||||
emotion_prompt = f"""
|
||||
请你基于这个表情包的详细描述，提取出最核心的情感含义，用1-2个词概括。
|
||||
详细描述：'{detailed_description}'
|
||||
|
||||
要求：
|
||||
1. 只输出1-2个最核心的情感词汇
|
||||
2. 从互联网梗、meme的角度理解
|
||||
3. 输出简短精准，不要解释
|
||||
4. 如果有多个词用逗号分隔
|
||||
"""
|
||||
|
||||
# 使用较低温度确保输出稳定
|
||||
emotion_llm = LLMRequest(model=global_config.model.utils, temperature=0.3, max_tokens=50, request_type="emoji")
|
||||
emotion_result, _ = await emotion_llm.generate_response_async(emotion_prompt)
|
||||
|
||||
if emotion_result is None:
|
||||
logger.warning("LLM未能生成情感标签，使用详细描述的前几个词")
|
||||
# 降级处理：从详细描述中提取关键词
|
||||
import jieba
|
||||
words = list(jieba.cut(detailed_description))
|
||||
emotion_result = "，".join(words[:2]) if len(words) >= 2 else (words[0] if words else "表情")
|
||||
|
||||
# 处理情感结果，取前1-2个最重要的标签
|
||||
emotions = [e.strip() for e in emotion_result.replace("，", ",").split(",") if e.strip()]
|
||||
final_emotion = emotions[0] if emotions else "表情"
|
||||
|
||||
# 如果有第二个情感且不重复，也包含进来
|
||||
if len(emotions) > 1 and emotions[1] != emotions[0]:
|
||||
final_emotion = f"{emotions[0]}，{emotions[1]}"
|
||||
|
||||
logger.info(f"[二步走识别] 详细描述: {detailed_description[:50]}... -> 情感标签: {final_emotion}")
|
||||
|
||||
# 再次检查缓存，防止并发写入时重复生成
|
||||
cached_description = self._get_description_from_db(image_hash, "emoji")
|
||||
if cached_description:
|
||||
logger.warning(f"虽然生成了描述，但是找到缓存表情包描述: {cached_description}")
|
||||
return f"[表情包，含义看起来是：{cached_description}]"
|
||||
return f"[表情包：{cached_description}]"
|
||||
|
||||
# 根据配置决定是否保存图片
|
||||
# if global_config.emoji.save_emoji:
|
||||
# 生成文件名和路径
|
||||
# 保存表情包文件和元数据（用于可能的后续分析）
|
||||
logger.debug(f"保存表情包: {image_hash}")
|
||||
current_timestamp = time.time()
|
||||
filename = f"{int(current_timestamp)}_{image_hash[:8]}.{image_format}"
|
||||
@@ -146,11 +180,11 @@ class ImageManager:
|
||||
with open(file_path, "wb") as f:
|
||||
f.write(image_bytes)
|
||||
|
||||
# 保存到数据库 (Images表)
|
||||
# 保存到数据库 (Images表) - 包含详细描述用于可能的注册流程
|
||||
try:
|
||||
img_obj = Images.get((Images.emoji_hash == image_hash) & (Images.type == "emoji"))
|
||||
img_obj.path = file_path
|
||||
img_obj.description = description
|
||||
img_obj.description = detailed_description # 保存详细描述
|
||||
img_obj.timestamp = current_timestamp
|
||||
img_obj.save()
|
||||
except Images.DoesNotExist: # type: ignore
|
||||
@@ -158,17 +192,17 @@ class ImageManager:
|
||||
emoji_hash=image_hash,
|
||||
path=file_path,
|
||||
type="emoji",
|
||||
description=description,
|
||||
description=detailed_description, # 保存详细描述
|
||||
timestamp=current_timestamp,
|
||||
)
|
||||
# logger.debug(f"保存表情包元数据: {file_path}")
|
||||
except Exception as e:
|
||||
logger.error(f"保存表情包文件或元数据失败: {str(e)}")
|
||||
|
||||
# 保存描述到数据库 (ImageDescriptions表)
|
||||
self._save_description_to_db(image_hash, description, "emoji")
|
||||
# 保存最终的情感标签到缓存 (ImageDescriptions表)
|
||||
self._save_description_to_db(image_hash, final_emotion, "emoji")
|
||||
|
||||
return f"[表情包：{description}]"
|
||||
return f"[表情包：{final_emotion}]"
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"获取表情包描述失败: {str(e)}")
|
||||
return "[表情包]"
|
||||
|
||||
Reference in New Issue
Block a user