扩散模型
定义
扩散模型通过学习逆转逐渐加噪过程来生成数据。 They have become the dominant approach for image generation (例如 DALL·E 2, Stable Diffusion).
与…不同 GANs, 训练稳定 (无最小-最大博弈); 不同于 VAEs, 样本清晰且多样. 代价是 many denoising steps at inference (though distillation and fewer-step schedulers reduce this). Used for text-to-image, inpainting, and video; see case study: DALL-E.
工作原理
**正向过程：**从数据 x0 开始，在 T 步中添加高斯噪声得到 x1、…、xT（近似pure noise). Reverse process: Learn a network that predicts the noise (or x0) at each step so you can go from xT to x0 by iteratively denoising. Training: take a real sample, add noise to a random step t, train the network to predict the added noise. Sampling: Start from random xT, run the learned reverse process 逐步 to get x0. The diagram summarizes forward (data → noise) and reverse (noise → data).
应用场景
Diffusion models are 从噪声生成和编辑高质量图像、音频和视频的首选方法.
- Image generation (例如 DALL·E 2, Stable Diffusion, Midjourney)
- Image editing, inpainting, and super-resolution
- Audio and video generation