扩散模型怎么生成How Diffusion Models Generate
How Diffusion Models Generate
生图主流方法在原理层长什么样：加噪、去噪、预测噪声。What the mainstream image-generation method looks like at the principle level: add noise, remove noise, predict noise.
训练分两个过程。前向过程把真图一步步加噪，直到变成纯噪声。
反向过程学的是去噪：每一步预测当前图里混进了多少噪声。
生成时从一片随机噪声出发，按学到的去噪步骤一步步「擦」出图像。
DDPM（Ho 等，2020）把这套做法确立为主流方法。
潜扩散模型（LDM，2021）把扩散搬进 VAE 压缩后的潜空间。算力大幅下降，今天多数生图模型是这一支。
Training has two processes. The forward process adds noise to a real image step by step until it is pure noise.
The reverse process learns denoising: at each step it predicts how much noise the current image contains.
Generation starts from random noise and 'erases' an image by following the learned denoising steps.
DDPM (Ho et al., 2020) established this recipe as the mainstream method.
Latent diffusion (LDM, 2021) runs the diffusion inside a VAE-compressed latent space, cutting compute sharply. Most image models today are this branch.
第二站的「扩散模型」告诉你它是什么；这一条告诉你它为什么这样工作。生图要迭代多步、「步数」是个参数，都从这里来。The station-2 entry 'diffusion model' tells you what it is; this one tells you why it works that way. The many iterations and the 'steps' parameter both come from here.
回到名词地图看第4站的其他 35 个词 →Back to the map: the other 35 terms in station 4 →