import gradio as gr import torch from transformers import Qwen3VLForConditionalGeneration, AutoProcessor from qwen_vl_utils import process_vision_info from PIL import Image import random import os import warnings from typing import List, Tuple, Optional # Suppress specific warnings warnings.filterwarnings('ignore', message='.*meta device.*') # Multi-language support TRANSLATIONS = { "en": { "title": "Qwen VL Image Description Generator", "header": "🖼️ Image Description Generator based on Qwen Vision Language Models", "subtitle": "Upload an image and enter a prompt to generate a description using Qwen VL models.", "language": "Language", "language_info": "Select language", "model_selection": "Model Selection", "model_info": "Select a model for generating descriptions", "advanced_params": "⚙️ Advanced Parameters", "max_tokens": "Max New Tokens", "max_tokens_info": "Maximum number of tokens to generate", "temperature": "Temperature", "temperature_info": "Controls randomness of generation", "top_p": "Top-p (nucleus sampling)", "top_p_info": "Probability threshold for token sampling", "top_k": "Top-k", "top_k_info": "Number of most probable tokens to consider", "seed": "Seed", "seed_info": "Seed for reproducibility (-1 for random)", "random_seed_btn": "🎲 Random Seed", "single_processing": "📄 Single Processing", "batch_processing": "📚 Batch Processing", "upload_image": "Upload Image", "prompt": "Prompt", "prompt_placeholder": "For example: Create a product description for online store", "generate_btn": "🚀 Generate Description", "result": "Result", "upload_images": "Upload Images", "prompts_multiline": "Prompts (one per line)", "prompts_placeholder": "Create a product description for online store\nCreate SEO Description for product\n...", "prompts_info": "Specify one prompt for all images or one prompt per image", "process_batch_btn": "🚀 Process Batch", "results": "Results", "examples_title": "💡 Example Prompts:", "example_1": "Create a product description for online store", "example_2": "Create an SEO description for a product with a maximum of 160 characters.", "example_3": "Create an attractive product description for marketplace", "example_4": "Describe image in detail for product catalog", "error_no_image": "Please upload an image", "error_no_prompt": "Please enter a prompt", "error_no_images": "Please upload images", "error_no_prompts": "Please enter prompts (one per line)", "error_prompt_mismatch": "Number of prompts ({}) does not match number of images ({}). Specify either one prompt for all images or one prompt per image.", "error_generation": "Error generating description: {}", "loading_model": "Loading model: {}", "model_loaded": "Model {} successfully loaded on {}", "image_label": "=== Image {}: {} ===", "prompt_label": "Prompt: {}", "result_label": "Result: {}", "model_size_warning": "⚠️ Note: Large models (8B+) may use CPU offloading if GPU memory is insufficient, which can slow down generation." }, "ru": { "title": "Генератор описаний изображений Qwen VL", "header": "🖼️ Генератор описаний изображений на основе Qwen Vision Language Models", "subtitle": "Загрузите изображение и введите промт для генерации описания с помощью моделей Qwen VL.", "language": "Язык", "language_info": "Выберите язык", "model_selection": "Выбор модели", "model_info": "Выберите модель для генерации описаний", "advanced_params": "⚙️ Расширенные параметры", "max_tokens": "Макс. количество новых токенов", "max_tokens_info": "Максимальное количество токенов для генерации", "temperature": "Температура", "temperature_info": "Контролирует случайность генерации", "top_p": "Top-p (nucleus sampling)", "top_p_info": "Вероятностный порог для выборки токенов", "top_k": "Top-k", "top_k_info": "Количество наиболее вероятных токенов для рассмотрения", "seed": "Seed", "seed_info": "Seed для воспроизводимости (-1 для случайного)", "random_seed_btn": "🎲 Случайный seed", "single_processing": "📄 Одиночная обработка", "batch_processing": "📚 Пакетная обработка", "upload_image": "Загрузите изображение", "prompt": "Промт", "prompt_placeholder": "Например: Создать описание товара для онлайн магазина", "generate_btn": "🚀 Генерировать описание", "result": "Результат", "upload_images": "Загрузите изображения", "prompts_multiline": "Промты (по одному на строку)", "prompts_placeholder": "Создать описание товара для онлайн магазина\nСоздать SEO Description для товара\n...", "prompts_info": "Укажите один промт для всех изображений или по одному промту на каждое изображение", "process_batch_btn": "🚀 Обработать пакет", "results": "Результаты", "examples_title": "💡 Примеры промтов:", "example_1": "Создать описание товара '' на русском языке", "example_2": "Создать SEO Description для товара максимум 160 символов на русском языке", "example_3": "Создать привлекательное описание продукта для маркетплейса на русском языке", "example_4": "Детально описать изображение для каталога товаров на русском языке", "error_no_image": "Пожалуйста, загрузите изображение", "error_no_prompt": "Пожалуйста, введите промт", "error_no_images": "Пожалуйста, загрузите изображения", "error_no_prompts": "Пожалуйста, введите промты (по одному на строку)", "error_prompt_mismatch": "Количество промтов ({}) не совпадает с количеством изображений ({}). Укажите либо один промт для всех изображений, либо по одному промту на каждое изображение.", "error_generation": "Ошибка при генерации описания: {}", "loading_model": "Загрузка модели: {}", "model_loaded": "Модель {} успешно загружена на {}", "image_label": "=== Изображение {}: {} ===", "prompt_label": "Промт: {}", "result_label": "Результат: {}", "model_size_warning": "⚠️ Примечание: Большие модели (8B+) могут использовать выгрузку на CPU при недостатке памяти GPU, что может замедлить генерацию." }, "zh": { "title": "Qwen VL 图像描述生成器", "header": "🖼️ 基于 Qwen Vision Language Models 的图像描述生成器", "subtitle": "上传图像并输入提示词，使用 Qwen VL 模型生成描述。", "language": "语言", "language_info": "选择语言", "model_selection": "模型选择", "model_info": "选择用于生成描述的模型", "advanced_params": "⚙️ 高级参数", "max_tokens": "最大新令牌数", "max_tokens_info": "生成的最大令牌数", "temperature": "温度", "temperature_info": "控制生成的随机性", "top_p": "Top-p（核采样）", "top_p_info": "令牌采样的概率阈值", "top_k": "Top-k", "top_k_info": "考虑的最可能令牌数", "seed": "随机种子", "seed_info": "用于可重现性的种子（-1 表示随机）", "random_seed_btn": "🎲 随机种子", "single_processing": "📄 单张处理", "batch_processing": "📚 批量处理", "upload_image": "上传图像", "prompt": "提示词", "prompt_placeholder": "例如：为在线商店创建产品描述", "generate_btn": "🚀 生成描述", "result": "结果", "upload_images": "上传图像", "prompts_multiline": "提示词（每行一个）", "prompts_placeholder": "为在线商店创建产品描述\n为产品创建SEO描述\n...", "prompts_info": "为所有图像指定一个提示词，或为每个图像指定一个提示词", "process_batch_btn": "🚀 处理批次", "results": "结果", "examples_title": "💡 示例提示词：", "example_1": "为在线商店创建产品描述", "example_2": "为产品创建SEO描述最多 160 个字符", "example_3": "为市场创建有吸引力的产品描述", "example_4": "详细描述产品目录的图像", "error_no_image": "请上传图像", "error_no_prompt": "请输入提示词", "error_no_images": "请上传图像", "error_no_prompts": "请输入提示词（每行一个）", "error_prompt_mismatch": "提示词数量（{}）与图像数量（{}）不匹配。请为所有图像指定一个提示词，或为每个图像指定一个提示词。", "error_generation": "生成描述时出错：{}", "loading_model": "正在加载模型：{}", "model_loaded": "模型 {} 已成功加载到 {}", "image_label": "=== 图像 {}: {} ===", "prompt_label": "提示词：{}", "result_label": "结果：{}", "model_size_warning": "⚠️ 注意：如果 GPU 内存不足，大型模型（8B+）可能会使用 CPU 卸载，这可能会减慢生成速度。" } } # Default language current_language = "en" def get_text(key: str) -> str: """Get translated text for the current language""" return TRANSLATIONS[current_language].get(key, key) class ImageDescriptionGenerator: def __init__(self): self.model = None self.processor = None self.current_model_name = None self.device = "cuda" if torch.cuda.is_available() else "cpu" def load_model(self, model_name: str): """Загрузка модели только если она еще не загружена или изменилась""" if self.current_model_name == model_name and self.model is not None: return print(get_text("loading_model").format(model_name)) # Предупреждение о больших моделях if "8B" in model_name or "4B" in model_name: print(get_text("model_size_warning")) # Освобождаем память от предыдущей модели if self.model is not None: del self.model del self.processor torch.cuda.empty_cache() if torch.cuda.is_available() else None # Загружаем новую модель с подавлением предупреждений with warnings.catch_warnings(): warnings.filterwarnings('ignore') self.model = Qwen3VLForConditionalGeneration.from_pretrained( model_name, dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32, device_map="auto" ) self.processor = AutoProcessor.from_pretrained(model_name) self.current_model_name = model_name print(get_text("model_loaded").format(model_name, self.device)) def generate_description( self, image_path: str, prompt: str, model_name: str, max_new_tokens: int = 1024, temperature: float = 0.6, top_p: float = 0.9, top_k: int = 50, seed: int = -1 ) -> str: """Генерация описания для одного изображения""" try: # Загружаем модель если необходимо self.load_model(model_name) # Устанавливаем seed если указан if seed != -1: torch.manual_seed(seed) if torch.cuda.is_available(): torch.cuda.manual_seed(seed) # Подготавливаем сообщения для модели messages = [ { "role": "user", "content": [ { "type": "image", "image": image_path, }, {"type": "text", "text": prompt}, ], } ] # Подготавливаем текст для модели text = self.processor.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) # Обрабатываем изображение и текст image_inputs, video_inputs = process_vision_info(messages) inputs = self.processor( text=[text], images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt", ) inputs = inputs.to(self.device) # Генерируем ответ with torch.no_grad(): generated_ids = self.model.generate( **inputs, max_new_tokens=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k, do_sample=True if temperature > 0 else False ) # Декодируем результат generated_ids_trimmed = [ out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids) ] output_text = self.processor.batch_decode( generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False ) return output_text[0] except Exception as e: return get_text("error_generation").format(str(e)) # Создаем глобальный экземпляр генератора generator = ImageDescriptionGenerator() def process_single_image( image, prompt: str, model_name: str, max_new_tokens: int, temperature: float, top_p: float, top_k: int, seed: int ) -> str: """Обработка одного изображения""" if image is None: return get_text("error_no_image") if not prompt.strip(): return get_text("error_no_prompt") # Сохраняем временное изображение если это numpy array if hasattr(image, 'shape'): temp_path = "temp_image.jpg" Image.fromarray(image).save(temp_path) image_path = temp_path else: image_path = image result = generator.generate_description( image_path=image_path, prompt=prompt, model_name=model_name, max_new_tokens=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k, seed=seed ) # Удаляем временный файл if hasattr(image, 'shape') and os.path.exists(temp_path): os.remove(temp_path) return result def process_batch_images( files: List, prompts_text: str, model_name: str, max_new_tokens: int, temperature: float, top_p: float, top_k: int, seed: int ) -> str: """Обработка пакета изображений""" if not files: return get_text("error_no_images") if not prompts_text.strip(): return get_text("error_no_prompts") # Разбиваем промты по строкам prompts = [p.strip() for p in prompts_text.split('\n') if p.strip()] if len(prompts) == 1: # Если один промт, используем его для всех изображений prompts = prompts * len(files) elif len(prompts) != len(files): return get_text("error_prompt_mismatch").format(len(prompts), len(files)) results = [] for idx, (file, prompt) in enumerate(zip(files, prompts), 1): image_path = file.name if hasattr(file, 'name') else file result = generator.generate_description( image_path=image_path, prompt=prompt, model_name=model_name, max_new_tokens=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k, seed=seed if seed == -1 else seed + idx - 1 # Разный seed для каждого изображения ) results.append(get_text("image_label").format(idx, os.path.basename(image_path)) + "\n") results.append(get_text("prompt_label").format(prompt) + "\n") results.append(get_text("result_label").format(result) + "\n\n") return "".join(results) def random_seed() -> int: """Генерация случайного seed""" return random.randint(0, 2**32 - 1) def update_examples(): return [ [get_text("example_1")], [get_text("example_2")], [get_text("example_3")], [get_text("example_4")] ] def create_interface(): """Create Gradio interface with current language""" with gr.Blocks(title=get_text("title"), theme=gr.themes.Soft()) as demo: # Header that will be updated header_md = gr.Markdown(f""" # {get_text("header")} {get_text("subtitle")} """) # Общие настройки - модель и язык в одном ряду with gr.Row(): model_dropdown = gr.Dropdown( choices=[ "Qwen/Qwen3-VL-2B-Instruct", "Qwen/Qwen3-VL-4B-Instruct", "Qwen/Qwen3-VL-8B-Instruct", "Qwen/Qwen3-VL-32B-Instruct", "Qwen/Qwen3-VL-2B-Thinking", "Qwen/Qwen3-VL-4B-Thinking", "Qwen/Qwen3-VL-8B-Thinking", "Qwen/Qwen3-VL-32B-Thinking", ], value="Qwen/Qwen3-VL-2B-Instruct", label=get_text("model_selection"), info=get_text("model_info"), scale=3 ) language_dropdown = gr.Dropdown( choices=[("English", "en"), ("Русский", "ru"), ("中文", "zh")], value=current_language, label=get_text("language"), info=get_text("language_info"), scale=1 ) # Расширенные параметры advanced_accordion = gr.Accordion(get_text("advanced_params"), open=False) with advanced_accordion: with gr.Row(): max_tokens_slider = gr.Slider( minimum=1, maximum=4096, value=1024, step=1, label=get_text("max_tokens"), info=get_text("max_tokens_info") ) temperature_slider = gr.Slider( minimum=0.1, maximum=2.0, value=0.6, step=0.1, label=get_text("temperature"), info=get_text("temperature_info") ) with gr.Row(): top_p_slider = gr.Slider( minimum=0.05, maximum=1.0, value=0.9, step=0.05, label=get_text("top_p"), info=get_text("top_p_info") ) top_k_slider = gr.Slider( minimum=1, maximum=1000, value=50, step=1, label=get_text("top_k"), info=get_text("top_k_info") ) with gr.Row(): seed_number = gr.Number( value=-1, label=get_text("seed"), info=get_text("seed_info"), precision=0 ) random_seed_btn = gr.Button(get_text("random_seed_btn"), size="sm") # Вкладки для одиночной и пакетной обработки tabs = gr.Tabs() with tabs: # Вкладка одиночной обработки single_tab = gr.TabItem(get_text("single_processing")) with single_tab: with gr.Row(): with gr.Column(scale=1): single_image = gr.Image( type="numpy", label=get_text("upload_image"), height=350 ) single_prompt = gr.Textbox( label=get_text("prompt"), placeholder=get_text("prompt_placeholder"), lines=3 ) single_submit_btn = gr.Button(get_text("generate_btn"), variant="primary") with gr.Column(scale=1): single_output = gr.Textbox( label=get_text("result"), lines=15, show_copy_button=True ) # Кликабельные примеры промтов examples_title = gr.Markdown(f"### {get_text('examples_title')}") single_examples = gr.Dataset( components=[single_prompt], samples=update_examples(), type="values" ) # Вкладка пакетной обработки batch_tab = gr.TabItem(get_text("batch_processing")) with batch_tab: with gr.Row(): with gr.Column(scale=1): batch_images = gr.File( file_count="multiple", label=get_text("upload_images"), file_types=["image"] ) batch_prompts = gr.Textbox( label=get_text("prompts_multiline"), placeholder=get_text("prompts_placeholder"), lines=5, info=get_text("prompts_info") ) batch_submit_btn = gr.Button(get_text("process_batch_btn"), variant="primary") with gr.Column(scale=1): batch_output = gr.Textbox( label=get_text("results"), lines=20, show_copy_button=True ) # Обработчики событий def change_language(lang): global current_language current_language = lang # Return updated text for all components return [ f""" # {get_text("header")} {get_text("subtitle")} """, # header_md gr.update(label=get_text("model_selection"), info=get_text("model_info")), # model_dropdown gr.update(label=get_text("language"), info=get_text("language_info")), # language_dropdown gr.update(label=get_text("advanced_params")), # advanced_accordion gr.update(label=get_text("max_tokens"), info=get_text("max_tokens_info")), # max_tokens_slider gr.update(label=get_text("temperature"), info=get_text("temperature_info")), # temperature_slider gr.update(label=get_text("top_p"), info=get_text("top_p_info")), # top_p_slider gr.update(label=get_text("top_k"), info=get_text("top_k_info")), # top_k_slider gr.update(label=get_text("seed"), info=get_text("seed_info")), # seed_number gr.update(value=get_text("random_seed_btn")), # random_seed_btn gr.update(label=get_text("single_processing")), # single_tab gr.update(label=get_text("upload_image")), # single_image gr.update(label=get_text("prompt"), placeholder=get_text("prompt_placeholder")), # single_prompt gr.update(value=get_text("generate_btn")), # single_submit_btn gr.update(label=get_text("result")), # single_output f"### {get_text('examples_title')}", # examples_title gr.update(samples=update_examples()), # single_examples gr.update(label=get_text("batch_processing")), # batch_tab gr.update(label=get_text("upload_images")), # batch_images gr.update(label=get_text("prompts_multiline"), placeholder=get_text("prompts_placeholder"), info=get_text("prompts_info")), # batch_prompts gr.update(value=get_text("process_batch_btn")), # batch_submit_btn gr.update(label=get_text("results")), # batch_output ] language_dropdown.change( fn=change_language, inputs=language_dropdown, outputs=[ header_md, model_dropdown, language_dropdown, advanced_accordion, max_tokens_slider, temperature_slider, top_p_slider, top_k_slider, seed_number, random_seed_btn, single_tab, single_image, single_prompt, single_submit_btn, single_output, examples_title, single_examples, batch_tab, batch_images, batch_prompts, batch_submit_btn, batch_output, ] ) single_examples.click( fn=lambda x: x[0] if x else "", inputs=[single_examples], outputs=[single_prompt] ) random_seed_btn.click( fn=random_seed, outputs=seed_number ) single_submit_btn.click( fn=process_single_image, inputs=[ single_image, single_prompt, model_dropdown, max_tokens_slider, temperature_slider, top_p_slider, top_k_slider, seed_number ], outputs=single_output ) batch_submit_btn.click( fn=process_batch_images, inputs=[ batch_images, batch_prompts, model_dropdown, max_tokens_slider, temperature_slider, top_p_slider, top_k_slider, seed_number ], outputs=batch_output ) return demo # Создаем интерфейс Gradio demo = create_interface() if __name__ == "__main__": demo.launch( server_name="0.0.0.0", server_port=7860, share=False, show_error=True )