📄 中文摘要
扩散模型通过模拟从清晰图像到噪声的逐步过程，并学习逆向去噪来生成图像。其核心机制在于将清晰图像逐步添加随机噪声和模糊，然后反复训练模型如何从噪声中恢复原始图像。这种方法使模型能够从简单的描述中生成全新且真实的图像。扩散模型的秘密在于其处理噪声的方式，它学习的是从混乱到有序的方向，而非简单地记忆现有图片。每一步训练都让神经网络更深入地理解真实图像的特征，从而使其能够从零开始构想图像。此外，通过提供额外的提示，可以引导生成过程，使结果偏向特定的对象或场景，例如猫、人脸或明亮的日落。
理解扩散模型：一个统一的视角
出处: Understanding Diffusion Models: A Unified Perspective
发布: 2026年2月10日
📄 English Summary
Understanding Diffusion Models: A Unified Perspective
Diffusion models generate images by simulating a process of gradually adding noise to a clear image and then learning to reverse this process. The core mechanism involves taking a pristine image, progressively introducing random dots and blur, and subsequently practicing how to meticulously clean it, repeatedly. This iterative training enables the models to generate novel and realistic images from textual prompts. The fundamental principle lies in their approach to noise: they learn the trajectory from disorder to order, rather than merely memorizing existing pictures. Each training step refines the neural network's understanding of what constitutes a real image, empowering it to imagine and construct one from scratch when prompted. Furthermore, the generation process can be subtly guided with additional hints, directing the outcome towards specific subjects or scenes, such as a cat, a human face, or a vibrant sunset, thereby offering a high degree of creative control over the generated content.