import gradio as gr import random import json import os import re from datetime import datetime from openai import OpenAI import subprocess import torch from PIL import Image from transformers import AutoProcessor, AutoModelForCausalLM, pipeline # 설치 과정은 생략 가능하며 필요한 경우에만 실행 subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True) # OpenAI 클라이언트 초기화 (API 키는 환경 변수에서 가져옴) openai_api_key = os.getenv("OPENAI_API") client = OpenAI(api_key=openai_api_key) # 번역 모델 추가 translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en") # Florence 모델 초기화 device = "cuda" if torch.cuda.is_available() else "cpu" florence_model = AutoModelForCausalLM.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True).to(device).eval() florence_processor = AutoProcessor.from_pretrained('microsoft/Florence-2-base', trust_remote_code=True) # 한글 프롬프트 번역 함수 def translate_prompt(prompt): if any("\uAC00" <= char <= "\uD7A3" for char in prompt): # 한글이 포함된 경우 translated = translator(prompt, max_length=512)[0]['translation_text'] return translated return prompt # Florence 캡션 함수 def florence_caption(image): if not isinstance(image, Image.Image): image = Image.fromarray(image) inputs = florence_processor(text="", images=image, return_tensors="pt").to(device) generated_ids = florence_model.generate( input_ids=inputs["input_ids"], pixel_values=inputs["pixel_values"], max_new_tokens=1024, early_stopping=False, do_sample=False, num_beams=3, ) generated_text = florence_processor.batch_decode(generated_ids, skip_special_tokens=False)[0] parsed_answer = florence_processor.post_process_generation( generated_text, task="", image_size=(image.width, image.height) ) return parsed_answer[""] # JSON 파일 로드 함수 def load_json_file(file_name): file_path = os.path.join("data", file_name) with open(file_path, "r") as file: return json.load(file) # JSON 데이터 로드 ARTFORM = load_json_file("artform.json") PHOTO_TYPE = load_json_file("photo_type.json") BODY_TYPES = load_json_file("body_types.json") DEFAULT_TAGS = load_json_file("default_tags.json") ROLES = load_json_file("roles.json") HAIRSTYLES = load_json_file("hairstyles.json") ADDITIONAL_DETAILS = load_json_file("additional_details.json") PHOTOGRAPHY_STYLES = load_json_file("photography_styles.json") DEVICE = load_json_file("device.json") PHOTOGRAPHER = load_json_file("photographer.json") ARTIST = load_json_file("artist.json") DIGITAL_ARTFORM = load_json_file("digital_artform.json") PLACE = load_json_file("place.json") LIGHTING = load_json_file("lighting.json") CLOTHING = load_json_file("clothing.json") COMPOSITION = load_json_file("composition.json") POSE = load_json_file("pose.json") BACKGROUND = load_json_file("background.json") # 프롬프트 생성 클래스 class PromptGenerator: def __init__(self, seed=None): self.rng = random.Random(seed) def split_and_choose(self, input_str): choices = [choice.strip() for choice in input_str.split(",")] return self.rng.choices(choices, k=1)[0] def get_choice(self, input_str, default_choices): if input_str.lower() == "disabled": return "" elif "," in input_str: return self.split_and_choose(input_str) elif input_str.lower() == "random": return self.rng.choices(default_choices, k=1)[0] else: return input_str def clean_consecutive_commas(self, input_string): cleaned_string = re.sub(r',\s*,', ',', input_string) return cleaned_string def process_string(self, replaced, seed): replaced = re.sub(r'\s*,\s*', ',', replaced) replaced = re.sub(r',+', ',', replaced) original = replaced first_break_clipl_index = replaced.find("BREAK_CLIPL") second_break_clipl_index = replaced.find("BREAK_CLIPL", first_break_clipl_index + len("BREAK_CLIPL")) if first_break_clipl_index != -1 and second_break_clipl_index != -1: clip_content_l = replaced[first_break_clipl_index + len("BREAK_CLIPL"):second_break_clipl_index] replaced = replaced[:first_break_clipl_index].strip(", ") + replaced[second_break_clipl_index + len("BREAK_CLIPL"):].strip(", ") clip_l = clip_content_l else: clip_l = "" first_break_clipg_index = replaced.find("BREAK_CLIPG") second_break_clipg_index = replaced.find("BREAK_CLIPG", first_break_clipg_index + len("BREAK_CLIPG")) if first_break_clipg_index != -1 and second_break_clipg_index != -1: clip_content_g = replaced[first_break_clipg_index + len("BREAK_CLIPG"):second_break_clipg_index] replaced = replaced[:first_break_clipg_index].strip(", ") + replaced[second_break_clipg_index + len("BREAK_CLIPG"):].strip(", ") clip_g = clip_content_g else: clip_g = "" t5xxl = replaced original = original.replace("BREAK_CLIPL", "").replace("BREAK_CLIPG", "") original = re.sub(r'\s*,\s*', ',', original) original = re.sub(r',+', ',', original) clip_l = re.sub(r'\s*,\s*', ',', clip_l) clip_l = re.sub(r',+', ',', clip_l) clip_g = re.sub(r'\s*,\s*', ',', clip_g) clip_g = re.sub(r',+', ',', clip_g) if clip_l.startswith(","): clip_l = clip_l[1:] if clip_g.startswith(","): clip_g = clip_g[1:] if original.startswith(","): original = original[1:] if t5xxl.startswith(","): t5xxl = t5xxl[1:] return original, seed, t5xxl, clip_l, clip_g def generate_prompt(self, seed, custom, subject, artform, photo_type, body_types, default_tags, roles, hairstyles, additional_details, photography_styles, device, photographer, artist, digital_artform, place, lighting, clothing, composition, pose, background, input_image): kwargs = locals() del kwargs['self'] seed = kwargs.get("seed", 0) if seed is not None: self.rng = random.Random(seed) components = [] custom = kwargs.get("custom", "") if custom: custom = translate_prompt(custom) # 번역 적용 components.append(custom) is_photographer = kwargs.get("artform", "").lower() == "photography" or ( kwargs.get("artform", "").lower() == "random" and self.rng.choice([True, False]) ) subject = kwargs.get("subject", "") if is_photographer: selected_photo_style = self.get_choice(kwargs.get("photography_styles", ""), PHOTOGRAPHY_STYLES) if not selected_photo_style: selected_photo_style = "photography" components.append(selected_photo_style) if kwargs.get("photography_style", "") != "disabled" and kwargs.get("default_tags", "") != "disabled" or subject != "": components.append(" of") default_tags = kwargs.get("default_tags", "random") body_type = kwargs.get("body_types", "") if not subject: if default_tags == "random": if body_type != "disabled" and body_type != "random": selected_subject = self.get_choice(kwargs.get("default_tags", ""), DEFAULT_TAGS).replace("a ", "").replace("an ", "") components.append("a ") components.append(body_type) components.append(selected_subject) elif body_type == "disabled": selected_subject = self.get_choice(kwargs.get("default_tags", ""), DEFAULT_TAGS) components.append(selected_subject) else: body_type = self.get_choice(body_type, BODY_TYPES) components.append("a ") components.append(body_type) selected_subject = self.get_choice(kwargs.get("default_tags", ""), DEFAULT_TAGS).replace("a ", "").replace("an ", "") components.append(selected_subject) elif default_tags == "disabled": pass else: components.append(default_tags) else: if body_type != "disabled" and body_type != "random": components.append("a ") components.append(body_type) elif body_type == "disabled": pass else: body_type = self.get_choice(body_type, BODY_TYPES) components.append("a ") components.append(body_type) components.append(subject) params = [ ("roles", ROLES), ("hairstyles", HAIRSTYLES), ("additional_details", ADDITIONAL_DETAILS), ] for param in params: components.append(self.get_choice(kwargs.get(param[0], ""), param[1])) for i in reversed(range(len(components))): if components[i] in PLACE: components[i] += "," break if kwargs.get("clothing", "") != "disabled" and kwargs.get("clothing", "") != "random": components.append(", dressed in ") clothing = kwargs.get("clothing", "") components.append(clothing) elif kwargs.get("clothing", "") == "random": components.append(", dressed in ") clothing = self.get_choice(kwargs.get("clothing", ""), CLOTHING) components.append(clothing) if kwargs.get("composition", "") != "disabled" and kwargs.get("composition", "") != "random": components.append(",") composition = kwargs.get("composition", "") components.append(composition) elif kwargs.get("composition", "") == "random": components.append(",") composition = self.get_choice(kwargs.get("composition", ""), COMPOSITION) components.append(composition) if kwargs.get("pose", "") != "disabled" and kwargs.get("pose", "") != "random": components.append(",") pose = kwargs.get("pose", "") components.append(pose) elif kwargs.get("pose", "") == "random": components.append(",") pose = self.get_choice(kwargs.get("pose", ""), POSE) components.append(pose) components.append("BREAK_CLIPG") if kwargs.get("background", "") != "disabled" and kwargs.get("background", "") != "random": components.append(",") background = kwargs.get("background", "") components.append(background) elif kwargs.get("background", "") == "random": components.append(",") background = self.get_choice(kwargs.get("background", ""), BACKGROUND) components.append(background) if kwargs.get("place", "") != "disabled" and kwargs.get("place", "") != "random": components.append(",") place = kwargs.get("place", "") components.append(place) elif kwargs.get("place", "") == "random": components.append(",") place = self.get_choice(kwargs.get("place", ""), PLACE) components.append(place + ",") lighting = kwargs.get("lighting", "").lower() if lighting == "random": selected_lighting = ", ".join(self.rng.sample(LIGHTING, self.rng.randint(2, 5))) components.append(",") components.append(selected_lighting) elif lighting == "disabled": pass else: components.append(", ") components.append(lighting) components.append("BREAK_CLIPG") components.append("BREAK_CLIPL") if is_photographer: if kwargs.get("photo_type", "") != "disabled": photo_type_choice = self.get_choice(kwargs.get("photo_type", ""), PHOTO_TYPE) if photo_type_choice and photo_type_choice != "random" and photo_type_choice != "disabled": random_value = round(self.rng.uniform(1.1, 1.5), 1) components.append(f", ({photo_type_choice}:{random_value}), ") params = [ ("device", DEVICE), ("photographer", PHOTOGRAPHER), ] components.extend([self.get_choice(kwargs.get(param[0], ""), param[1]) for param in params]) if kwargs.get("device", "") != "disabled": components[-2] = f", shot on {components[-2]}" if kwargs.get("photographer", "") != "disabled": components[-1] = f", photo by {components[-1]}" else: digital_artform_choice = self.get_choice(kwargs.get("digital_artform", ""), DIGITAL_ARTFORM) if digital_artform_choice: components.append(f"{digital_artform_choice}") if kwargs.get("artist", "") != "disabled": components.append(f"by {self.get_choice(kwargs.get('artist', ''), ARTIST)}") components.append("BREAK_CLIPL") prompt = " ".join(components) prompt = re.sub(" +", " ", prompt) replaced = prompt.replace("of as", "of") replaced = self.clean_consecutive_commas(replaced) return self.process_string(replaced, seed) def add_caption_to_prompt(self, prompt, caption): if caption: return f"{prompt}, {caption}" return prompt class OpenAIGenerationNode: def __init__(self): self.client = OpenAI() self.prompts_dir = "./prompts" os.makedirs(self.prompts_dir, exist_ok=True) def save_prompt(self, prompt): filename_text = "openai_" + prompt.split(',')[0].strip() filename_text = re.sub(r'[^\w\-_\. ]', '_', filename_text) filename_text = filename_text[:30] timestamp = datetime.now().strftime("%Y%m%d_%H%M%S") base_filename = f"{filename_text}_{timestamp}.txt" filename = os.path.join(self.prompts_dir, base_filename) with open(filename, "w") as file: file.write(prompt) print(f"Prompt saved to {filename}") def generate(self, input_text, happy_talk, compress, compression_level, poster, custom_base_prompt=""): try: # 고정 모델: gpt-4.1-mini openai_model = "gpt-4.1-mini" default_happy_prompt = """Create a detailed visually descriptive caption of this description, which will be used as a prompt for a text to image AI system (caption only, no instructions like "create an image").Remove any mention of digital artwork or artwork style. Give detailed visual descriptions of the character(s), including ethnicity, skin tone, expression etc. Imagine using keywords for a still for someone who has aphantasia. Describe the image style, e.g. any photographic or art styles / techniques utilized. Make sure to fully describe all aspects of the cinematography, with abundant technical details and visual descriptions. If there is more than one image, combine the elements and characters from all of the images creatively into a single cohesive composition with a single background, inventing an interaction between the characters. Be creative in combining the characters into a single cohesive scene. Focus on two primary characters (or one) and describe an interesting interaction between them, such as a hug, a kiss, a fight, giving an object, an emotional reaction / interaction. If there is more than one background in the images, pick the most appropriate one. Your output is only the caption itself, no comments or extra formatting. The caption is in a single long paragraph. If you feel the images are inappropriate, invent a new scene / characters inspired by these. Additionally, incorporate a specific movie director's visual style and describe the lighting setup in detail, including the type, color, and placement of light sources to create the desired mood and atmosphere. Always frame the scene, including details about the film grain, color grading, and any artifacts or characteristics specific.""" default_simple_prompt = """Create a brief, straightforward caption for this description, suitable for a text-to-image AI system. Focus on the main elements, key characters, and overall scene without elaborate details. Provide a clear and concise description in one or two sentences.""" poster_prompt = """Analyze the provided description and extract key information to create a movie poster style description. Format the output as follows: Title: A catchy, intriguing title that captures the essence of the scene, place the title in "". Main character: Give a description of the main character. Background: Describe the background in detail. Supporting characters: Describe the supporting characters Branding type: Describe the branding type Tagline: Include a tagline that captures the essence of the movie. Visual style: Ensure that the visual style fits the branding type and tagline. You are allowed to make up film and branding names, and do them like 80's, 90's or modern movie posters.""" if poster: base_prompt = poster_prompt elif custom_base_prompt.strip(): base_prompt = custom_base_prompt else: base_prompt = default_happy_prompt if happy_talk else default_simple_prompt if compress and not poster: compression_chars = { "soft": 600 if happy_talk else 300, "medium": 400 if happy_talk else 200, "hard": 200 if happy_talk else 100 } char_limit = compression_chars[compression_level] base_prompt += f" Compress the output to be concise while retaining key visual details. MAX OUTPUT SIZE no more than {char_limit} characters." # 올바른 OpenAI API 호출 형식 response = self.client.chat.completions.create( model=openai_model, messages=[ { "role": "system", "content": "You are a helpful assistant. Try your best to give best response possible to user." }, { "role": "user", "content": f"{base_prompt}\nDescription: {input_text}" } ], temperature=1, max_tokens=2048, top_p=1 ) # 응답 추출 output = response.choices[0].message.content # 출력 정리 if ": " in output: output = output.split(": ", 1)[1].strip() elif output.lower().startswith("here"): sentences = output.split(". ") if len(sentences) > 1: output = ". ".join(sentences[1:]).strip() # 프롬프트 저장 self.save_prompt(output) return output except Exception as e: print(f"An error occurred: {e}") return f"Error occurred while processing the request: {str(e)}" # 개선된 CSS css = """ footer { visibility: hidden; } .main-title { text-align: center; margin: 20px 0; } .main-title h1 { color: #2c3e50; font-size: 2.5em; } .badge-container { display: flex; justify-content: center; gap: 15px; margin: 20px 0; } .prompt-generator-container { background: #f8f9fa; border-radius: 10px; padding: 20px; margin: 10px 0; } .settings-container { background: #ffffff; border-radius: 8px; padding: 15px; margin: 10px 0; box-shadow: 0 2px 4px rgba(0,0,0,0.1); } .output-container { background: #e8f4f8; border-radius: 8px; padding: 15px; margin: 10px 0; } .generate-button { background: #3498db !important; color: white !important; font-size: 1.1em !important; padding: 12px 24px !important; margin: 10px 0 !important; } .openai-button { background: #27ae60 !important; color: white !important; font-size: 1.1em !important; padding: 12px 24px !important; } """ def create_interface(): prompt_generator = PromptGenerator() openai_node = OpenAIGenerationNode() with gr.Blocks(theme="soft", css=css) as demo: # 헤더 with gr.Row(elem_classes="main-title"): gr.HTML("""

🎨 Flux Prompt Generator

한글 입력 지원 | AI 기반 프롬프트 생성기

""") # 배지 with gr.Row(elem_classes="badge-container"): gr.HTML(""" OpenFree badge Discord badge """) # 메인 컨테이너 with gr.Row(): # 왼쪽 패널 - 설정 with gr.Column(scale=2): with gr.Box(elem_classes="settings-container"): gr.Markdown("### 🎯 기본 설정") seed = gr.Slider(0, 30000, label='시드값', step=1, value=random.randint(0,30000)) custom = gr.Textbox(label="✏️ 커스텀 프롬프트 (한글 가능)", placeholder="원하는 설명을 입력하세요...") subject = gr.Textbox(label="🎭 주제 (선택사항)", placeholder="예: 아름다운 여성, 귀여운 고양이 등...") # 전역 옵션 설정 gr.Markdown("### ⚡ 빠른 설정") global_option = gr.Radio( ["Disabled", "Random", "No Figure Rand"], label="모든 옵션 일괄 설정:", value="Disabled", info="한 번에 모든 설정을 변경합니다" ) # 상세 설정들 with gr.Accordion("🎨 아트폼 & 사진 타입", open=False): with gr.Box(elem_classes="settings-container"): artform = gr.Dropdown(["disabled", "random"] + ARTFORM, label="아트폼", value="disabled") photo_type = gr.Dropdown(["disabled", "random"] + PHOTO_TYPE, label="사진 타입", value="disabled") with gr.Accordion("👤 캐릭터 설정", open=False): with gr.Box(elem_classes="settings-container"): body_types = gr.Dropdown(["disabled", "random"] + BODY_TYPES, label="체형", value="disabled") default_tags = gr.Dropdown(["disabled", "random"] + DEFAULT_TAGS, label="기본 태그", value="disabled") roles = gr.Dropdown(["disabled", "random"] + ROLES, label="역할", value="disabled") hairstyles = gr.Dropdown(["disabled", "random"] + HAIRSTYLES, label="헤어스타일", value="disabled") clothing = gr.Dropdown(["disabled", "random"] + CLOTHING, label="의상", value="disabled") with gr.Accordion("🏞️ 장면 설정", open=False): with gr.Box(elem_classes="settings-container"): place = gr.Dropdown(["disabled", "random"] + PLACE, label="장소", value="disabled") lighting = gr.Dropdown(["disabled", "random"] + LIGHTING, label="조명", value="disabled") composition = gr.Dropdown(["disabled", "random"] + COMPOSITION, label="구도", value="disabled") pose = gr.Dropdown(["disabled", "random"] + POSE, label="포즈", value="disabled") background = gr.Dropdown(["disabled", "random"] + BACKGROUND, label="배경", value="disabled") with gr.Accordion("🎭 스타일 & 아티스트", open=False): with gr.Box(elem_classes="settings-container"): additional_details = gr.Dropdown(["disabled", "random"] + ADDITIONAL_DETAILS, label="추가 디테일", value="disabled") photography_styles = gr.Dropdown(["disabled", "random"] + PHOTOGRAPHY_STYLES, label="사진 스타일", value="disabled") device = gr.Dropdown(["disabled", "random"] + DEVICE, label="촬영 장비", value="disabled") photographer = gr.Dropdown(["disabled", "random"] + PHOTOGRAPHER, label="사진작가", value="disabled") artist = gr.Dropdown(["disabled", "random"] + ARTIST, label="아티스트", value="disabled") digital_artform = gr.Dropdown(["disabled", "random"] + DIGITAL_ARTFORM, label="디지털 아트폼", value="disabled") generate_button = gr.Button("🚀 프롬프트 생성", variant="primary", elem_classes="generate-button") # 중간 패널 - 이미지와 출력 with gr.Column(scale=2): with gr.Accordion("🖼️ 이미지 캡션 생성", open=False): with gr.Box(elem_classes="settings-container"): input_image = gr.Image(label="이미지 업로드 (선택사항)", type="pil") caption_output = gr.Textbox(label="생성된 캡션", lines=3) with gr.Row(): create_caption_button = gr.Button("📝 캡션 생성", variant="secondary") add_caption_button = gr.Button("➕ 프롬프트에 추가", variant="secondary") with gr.Box(elem_classes="output-container"): gr.Markdown("### 📋 생성된 프롬프트") output = gr.Textbox(label="메인 프롬프트", lines=4) with gr.Accordion("고급 출력 옵션", open=False): t5xxl_output = gr.Textbox(label="T5XXL", lines=2) clip_l_output = gr.Textbox(label="CLIP L", lines=2) clip_g_output = gr.Textbox(label="CLIP G", lines=2) # 오른쪽 패널 - OpenAI with gr.Column(scale=2): with gr.Box(elem_classes="prompt-generator-container"): gr.Markdown("### 🤖 OpenAI 프롬프트 향상") gr.HTML("

사용 모델: gpt-4.1-mini

") with gr.Box(elem_classes="settings-container"): with gr.Row(): happy_talk = gr.Checkbox(label="😊 Happy Talk", value=True, info="더 상세한 설명") compress = gr.Checkbox(label="🗜️ 압축", value=True, info="출력 길이 제한") compression_level = gr.Radio( ["soft", "medium", "hard"], label="압축 강도", value="hard", visible=True ) poster = gr.Checkbox(label="🎬 영화 포스터 스타일", value=False) custom_base_prompt = gr.Textbox( label="🛠️ 커스텀 베이스 프롬프트", lines=5, placeholder="OpenAI에게 특별한 지시사항을 입력하세요..." ) generate_text_button = gr.Button("✨ AI로 프롬프트 향상", variant="primary", elem_classes="openai-button") text_output = gr.Textbox( label="🎯 AI 향상 결과", lines=10, elem_classes="output-container" ) # 이벤트 핸들러들 def create_caption(image): if image is not None: return florence_caption(image) return "" create_caption_button.click( create_caption, inputs=[input_image], outputs=[caption_output] ) generate_button.click( prompt_generator.generate_prompt, inputs=[seed, custom, subject, artform, photo_type, body_types, default_tags, roles, hairstyles, additional_details, photography_styles, device, photographer, artist, digital_artform, place, lighting, clothing, composition, pose, background, input_image], outputs=[output, gr.Number(visible=False), t5xxl_output, clip_l_output, clip_g_output] ) add_caption_button.click( prompt_generator.add_caption_to_prompt, inputs=[output, caption_output], outputs=[output] ) generate_text_button.click( lambda input_text, happy_talk, compress, compression_level, poster, custom_base_prompt: openai_node.generate(input_text, happy_talk, compress, compression_level, poster, custom_base_prompt), inputs=[output, happy_talk, compress, compression_level, poster, custom_base_prompt], outputs=text_output ) # 압축 체크박스에 따른 압축 강도 표시/숨기기 compress.change( lambda x: gr.update(visible=x), inputs=[compress], outputs=[compression_level] ) # 전역 옵션 변경 함수 def update_all_options(choice): updates = {} dropdown_list = [ artform, photo_type, body_types, default_tags, roles, hairstyles, clothing, place, lighting, composition, pose, background, additional_details, photography_styles, device, photographer, artist, digital_artform ] if choice == "Disabled": for dropdown in dropdown_list: updates[dropdown] = gr.update(value="disabled") elif choice == "Random": for dropdown in dropdown_list: updates[dropdown] = gr.update(value="random") else: # No Figure Random # 캐릭터 관련 설정은 disabled for dropdown in [photo_type, body_types, default_tags, roles, hairstyles, clothing, pose, additional_details]: updates[dropdown] = gr.update(value="disabled") # 나머지는 random for dropdown in [artform, place, lighting, composition, background, photography_styles, device, photographer, artist, digital_artform]: updates[dropdown] = gr.update(value="random") return list(updates.values()) global_option.change( update_all_options, inputs=[global_option], outputs=[ artform, photo_type, body_types, default_tags, roles, hairstyles, clothing, place, lighting, composition, pose, background, additional_details, photography_styles, device, photographer, artist, digital_artform ] ) return demo if __name__ == "__main__": demo = create_interface() demo.launch()