PHP前端开发

使用稳定的扩散V上的笔记本电脑上的AI驱动图像处理 - 这比您想象的要容易!

百变鹏仔 3周前 (02-05) #Python
文章标签 笔记本电脑

这个脚本利用稳定的扩散v1.5从拥抱面孔的扩散器库来基于给定文本提示符生成图像变化。通过使用火炬和pil,它处理输入图像,应用ai驱动的转换并保存结果。

您可以克隆此回购以获取代码https://github.com/alexander-uspenskiy/image_variations> 源代码:

import torchfrom diffusers import StableDiffusionImg2ImgPipelinefrom PIL import Imageimport requestsfrom io import BytesIOdef load_image(image_path, target_size=(768, 768)):    """    Load and preprocess the input image    """    if image_path.startswith('http'):        response = requests.get(image_path)        image = Image.open(BytesIO(response.content))    else:        image = Image.open(image_path)    # Resize and preserve aspect ratio    image = image.convert("RGB")    image.thumbnail(target_size, Image.Resampling.LANCZOS)    # Create new image with padding to reach target size    new_image = Image.new("RGB", target_size, (255, 255, 255))    new_image.paste(image, ((target_size[0] - image.size[0]) // 2,                           (target_size[1] - image.size[1]) // 2))    return new_imagedef generate_image_variation(    input_image_path,    prompt,    model_id="stable-diffusion-v1-5/stable-diffusion-v1-5",    num_images=1,    strength=0.75,    guidance_scale=7.5,    seed=None):    """    Generate variations of an input image using a specified prompt    Parameters:    - input_image_path: Path or URL to the input image    - prompt: Text prompt to guide the image generation    - model_id: Hugging Face model ID    - num_images: Number of variations to generate    - strength: How much to transform the input image (0-1)    - guidance_scale: How closely to follow the prompt    - seed: Random seed for reproducibility    Returns:    - List of generated images    """    # Set random seed if provided    if seed is not None:        torch.manual_seed(seed)    # Load the model    device = "cuda" if torch.cuda.is_available() else "cpu"    pipe = StableDiffusionImg2ImgPipeline.from_pretrained(        model_id,        torch_dtype=torch.float16 if device == "cuda" else torch.float32    ).to(device)    # Load and preprocess the input image    init_image = load_image(input_image_path)    # Generate images    result = pipe(        prompt=prompt,        image=init_image,        num_images_per_prompt=num_images,        strength=strength,        guidance_scale=guidance_scale    )    return result.imagesdef save_generated_images(images, output_prefix="generated"):    """    Save the generated images with sequential numbering    """    for i, image in enumerate(images):        image.save(f"images-out/{output_prefix}_{i}.png")# Example usageif __name__ == "__main__":    # Example parameters    input_image = "images-in/Image_name.jpg"  # or URL    prompt = "Draw the image in modern art style, photorealistic and detailed."    # Generate variations    generated_images = generate_image_variation(        input_image,        prompt,        num_images=3,        strength=0.75,        seed=42  # Optional: for reproducibility    )    # Save the results    save_generated_images(generated_images)

它的工作原理:

>加载和预处理输入图像

接受本地文件路径和url。

>将图像转换为rgb格式,并将其调整为768×768,以维持纵横比。
添加填充以适合目标尺寸。
初始化稳定扩散v1.5

>将模型加载在cuda上(如果有)或落回cpu。使用stablediffusionimg2imgpipeline处理输入映像。生成ai修改的图像变化


添加文本提示来指导转换。
强度(0-1)和引导量表(更高=更严格的提示依从性)等参数允许自定义。

每个提示支持多个输出图像。

将结果保存到图像输出目录。

>输出带有顺序命名方案的生成图像(生成_0.png,生成_1.png等)。

示例用例

>您可以使用以下提示来将一个人的图像转换为中世纪的国王提示=“在中世纪的环境中,将这个人当作强大的国王,逼真的和详细的。

初始图像:


结果:

cons&pros

cons:

在某些硬件配置上可能会很慢。

小尺寸模型限制。

用于微调输出的可自定义参数。

可重现的可选随机种子。