import os import time import gradio as gr from phi.assistant import Assistant from phi.llm.openai import OpenAIChat from pydantic import BaseModel from .log import logger qwen_model = 'qwen-plus' qwen_api_key = os.getenv("QWEN_DS_API_KEY") qwen_url = 'https://dashscope.aliyuncs.com/compatible-mode/v1' class TranslateFormat(BaseModel): en: str class PormptExtendFormat(BaseModel): prompt: str trans_assist = Assistant( llm=OpenAIChat(model=qwen_model, max_tokens=3000, temperature=0.3, api_key=qwen_api_key, base_url=qwen_url), description='你是专业的英语翻译,请将用户输入的一句中文翻译成英文', output_model=TranslateFormat ) requirements = [ "1.对于过于简短的用户输入,在不改变原意及主体外观的前提下,合理补充具象化细节,避免使用过于抽象描述。", "2.如果涉及人像,完善与用户输入相符的人物外貌、表情、种族、姿态、景别、穿着、影调、质感等方面描述。", "3.匹配符合用户意图且精准详细的风格描述。如果用户未指定,则根据画面选择最恰当的风格,或使用纪实摄影照片。", "4.若输入为中文,则整体中文输出;若输入为全英文,则整体英文输出;保留引号中原文以及重要的输入信息,不要改写。", "5.在语义完整前提下,改写后prompt字数小于200字,避免prompt冗长。", """ 6.改写后 prompt 示例: (1) 日系小清新写真照片,扎着双麻花辫的波西米亚小女孩坐在船边。女孩穿着白色方领泡泡袖连衣裙,裙子上有褶皱和纽扣装饰。她皮肤白皙,五官精致,眼泪汪汪直视镜头。她双手扶船,长发刘海遮住部分额头。背景是清澈明亮的户外场景,可见蓝天、山峦和一些干枯植物。高清写实摄影,近景中心对称构图。\n (2) 二次元厚涂动漫插画,一个猫耳东亚萌妹手持文件夹,怒气冲冲走向电脑。她深紫色爆炸头,红色眼睛,头顶有一个粉色光圈。少女身穿深灰色短裙和浅灰色上衣,腰间系着白色系带,胸前佩戴名牌,居中写着黑体中文"紫阳"。室内背景中摆放了很多办公桌。粗线条的日系赛璐璐风格。近景半身略仰视视角。\n (3) 美剧艺术海报风格,身穿黄色防护服的老年Walter White坐在金属折叠椅上,头顶无衬线英文写着"Breaking Bad",周围是成堆的美元和蓝色塑料储物箱。他戴着眼镜目光直视镜头,左手拿着一支雪茄,右手放在膝盖上。背景是废弃阴暗的厂房,阳光透过窗户照射进来。画面带有明显颗粒质感纹理。长焦人物平视特写。 (4) CG game concept digital art featuring three giant mutant crocodiles with wide-open mouths, revealing pink tongues and sharp teeth. Their rough skin resembles grayish-white stone. On the back of the crocodile to the left, lush trees, shrubs, and some thorn-like protrusions grow. The background of the scene shows a dusk sky and a shimmering pond. The overall atmosphere is dark and cold, with a tilted composition that creates a strong sense of depth and layers. \n """ ] prompt_writer = Assistant( llm=OpenAIChat(model=qwen_model, max_tokens=8000, temperature=0.3, api_key=qwen_api_key, base_url=qwen_url), description="你是一位prompt优化师,旨在将用户输入改写为优质prompt,在不影响原意前提下更高质量生图。", instructions=requirements, debug_mode=False, output_model=PormptExtendFormat, ) def contains_chinese(text): """ 判断文本中是否包含中文字符(基于 Unicode 范围) """ # 定义中文的 Unicode 范围(基础汉字 + 扩展区) cjk_ranges = [ (0x4e00, 0x9fff), # CJK 基本汉字区(含常用汉字) (0x3400, 0x4dbf), # CJK 扩展 A (0x20000, 0x2a6df), # CJK 扩展 B (0x2a700, 0x2b73f), # CJK 扩展 C (0x2b740, 0x2b81f), # CJK 扩展 D (0x2b820, 0x2ceaf), # CJK 扩展 E ] for c in text: code = ord(c) for start, end in cjk_ranges: if start <= code <= end: return True return False def translate_prompt(prompt, max_attempts=5): attempts = 0 while attempts < max_attempts: try: if contains_chinese(prompt): res = trans_assist.run(prompt) logger.info(f"translate Chinese prompt into English: {prompt} -> {res.en}") return res.en else: return prompt except Exception as e: print(f"尝试 {attempts + 1} 失败:{e}") attempts += 1 if attempts < 5: time.sleep(1) # 延迟重试 raise gr.Error(f"提示词扩写达到最大尝试次数 {max_attempts},任务失败。") def extend_prompt(prompt, max_attempts=5): attempts = 0 while attempts < max_attempts: try: res = prompt_writer.run(prompt, stream=False) logger.info(f"extend prompt finished: {prompt} -> {res.prompt}.") return res.prompt except Exception as e: print(f"尝试 {attempts + 1} 失败:{e}") attempts += 1 if attempts < 5: time.sleep(1) # 延迟重试 raise gr.Error(f"提示词扩写达到最大尝试次数 {max_attempts},任务失败。") if __name__ == '__main__': # test_trans() pass