概念百科
33 个概念 · 33 个已配图文详解
每个概念都是一份独立的图文解释:直觉 → 公式 → 图示 → 动画 → 例题 → 常见误区。
已详解
AR vs Diffusion
Comparison axis — autoregressive sequential decoding vs parallel iterative refinement; throughput, quality, and infilling trade-offs.
Module 1: 直觉与生成模型基础 已详解Classifier Guidance
Use gradients of a separately trained classifier to steer diffusion sampling toward a class.
Module 5: 采样、引导与应用 已详解Classifier-Free Guidance
Trick to trade diversity for sample quality by extrapolating between conditional and unconditional predictions.
Module 5: 采样、引导与应用 已详解Consistency Models
Distillation framework that learns a one-step generator consistent with the PF-ODE.
Module 4: Flow Matching 已详解Continuous Normalizing Flow
Generative model defined by integrating a learned vector field; trained via maximum likelihood or flow matching.
Module 4: Flow Matching 已详解DDIM
Deterministic non-Markovian sampler for diffusion models; fewer steps than ancestral sampling.
Module 5: 采样、引导与应用 已详解DDPM
Denoising Diffusion Probabilistic Models — Ho et al. 2020; discrete-time Gaussian diffusion trained with a simple denoising loss.
Module 2: DDPM — 加噪与去噪 已详解Diffusion Language Model
A language model whose decoder is a discrete diffusion process (LLaDA, MDLM, DiffuLLaMA, etc.).
Module 6: 离散与语言扩散 已详解Discrete Diffusion
Diffusion processes that act on discrete tokens (vs continuous vectors); foundation for language diffusion.
Module 6: 离散与语言扩散 已详解Epsilon Prediction
Parameterise the network to predict the noise epsilon added to x_0; the original DDPM choice.
Module 2: DDPM — 加噪与去噪 已详解Flow Matching
Training a vector field that transports noise to data along straight-ish paths; generalises diffusion.
Module 4: Flow Matching 已详解Forward Diffusion Process
Markov chain that progressively adds Gaussian noise to data; defines q(x_t | x_0).
Module 2: DDPM — 加噪与去噪 已详解Image Generation
Sampling images from a generative model; standard benchmark for diffusion methods.
已详解Langevin Dynamics
MCMC sampler that follows the score plus Gaussian noise; basis of annealed Langevin sampling.
Module 3: Score 视角 已详解Latent Diffusion
Run diffusion in a learned latent space (e.g., VAE latents) instead of pixel space; the recipe behind Stable Diffusion.
Module 5: 采样、引导与应用 已详解Masked Diffusion
Absorbing-state discrete diffusion where the forward process replaces tokens with a [MASK]; trained like masked-LM.
Module 6: 离散与语言扩散 已详解Noise Schedule
Choice of beta_t (or alpha_bar_t) that controls how much noise is added per step — linear, cosine, sigmoid variants.
Module 2: DDPM — 加噪与去噪 已详解Parallel Decoding
Decoding strategies (Jacobi, lookahead, diffusion-style) that produce multiple tokens per step.
Module 6: 离散与语言扩散 已详解Probability Flow ODE
Deterministic ODE whose marginals match the SDE; enables exact likelihood and fast sampling.
Module 3: Score 视角 已详解Rectified Flow
Flow-matching variant that straightens trajectories via reflow; enables one-step generation.
Module 4: Flow Matching 已详解Reparameterization Trick
Express stochastic sampling as a deterministic function of input + noise so gradients can flow.
Module 1: 直觉与生成模型基础 已详解Reverse Diffusion Process
Learned chain that denoises from pure noise back to data; defines p_theta(x_{t-1} | x_t).
Module 2: DDPM — 加噪与去噪 已详解Score Function
Gradient of the log density, ∇_x log p(x); central object of score-based generative models.
Module 3: Score 视角 已详解Score Matching
Training objective for learning the score function without knowing the normalising constant.
Module 3: Score 视角 已详解Score-Entropy Discrete Diffusion
SEDD — replaces score matching with a concrete score-entropy loss for discrete data; Lou & Ermon.
Module 6: 离散与语言扩散 已详解SDE Formulation
Continuous-time view of diffusion as a stochastic differential equation; unifies VE, VP, and sub-VP SDEs.
Module 3: Score 视角 已详解Simple Loss
The MSE-on-epsilon objective from DDPM that drops VLB weighting and works in practice.
Module 2: DDPM — 加噪与去噪 已详解Tweedie's Formula
Closed form for posterior mean E[x_0 | x_t] in terms of the score; ties denoising to score estimation.
Module 3: Score 视角 已详解v-Prediction
Predict the "velocity" v = alpha * eps - sigma * x; improves training near boundaries; used by progressive distillation.
已详解VAE Encoder
Variational autoencoder that compresses pixels to a latent space used by latent diffusion.
Module 1: 直觉与生成模型基础 已详解Variational Lower Bound
ELBO decomposition of the diffusion log-likelihood into prior, denoising, and reconstruction terms.
Module 1: 直觉与生成模型基础 已详解Velocity Field
Time-dependent vector field v_t(x) that defines the ODE/SDE trajectory.
Module 4: Flow Matching 已详解Video Diffusion
Diffusion models that generate video by extending the denoiser with temporal layers or 3D attention.
Module 5: 采样、引导与应用