所属模块：
M6 · 多模态与生成模型 (Multimodal & Generative Models)| 专题分类：扩散模型基础 (Diffusion Models Foundations (DDPM))| 难度等级：Hard
一、核心一句话结论 (One-Sentence Summary)
ε-prediction 在高噪声区难（信号弱）；x_0-prediction 在低噪声区难；v-prediction 是两者的平衡，各 SNR 区间难度更均匀。
Velocity v-prediction parameterizes diffusion networks to predict a balanced linear combination of clean signal and noise, preventing gradient explosion and numerical collapse in high-resolution, zero-terminal SNR regimes.
二、核心考点要义 (Key Insights)
- 📌 ε-prediction：预测噪声（高噪声区相对易、低噪声区难？）
- 📌 x_0-prediction：预测干净图像（低噪声区易、高噪声区极难）
- 📌 v-prediction：预测’速度’，各 SNR 区间难度均衡
English Insights:
– Epsilon-prediction breakdown: predicting noise $,epsilon,$ becomes ill-conditioned in extreme noise regimes ($t to T, text{SNR} to 0$), where small errors in $,epsilon,$ cause explosive swings in estimated $,x_0,$
– V-prediction formulation: parameterizes network to predict velocity $,v = alpha_t epsilon – sigma_t x_0,$, smoothly interpolating between noise at $t=0$ and data signal at $t=T$
– Zero terminal SNR stability: v-prediction is mandatory for training models with zero terminal SNR, enabling high-contrast generation and arbitrary resolution scaling
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$v=sqrt{baralpha_t},epsilon-sqrt{1-baralpha_t},x_0;qquad epsilon=sqrt{baralpha_t},v+sqrt{1-baralpha_t},x_0$$
数学机理：三种参数化——它们预测同一个量（去噪方向）的不同线性组合。设 x_t=√ᾱt·x_0+√(1−ᾱ_t)·ε：(1) ε-prediction——网络输出 εθ ≈ ε。(2) x_0-prediction——网络输出 x̂0 ≈ x_0。(3) v-prediction（Salimans & Ho 2022）——网络输出 v ≈ √ᾱ_t·ε−√(1−ᾱ_t)·x_0（称为’速度’，因为它是’从 x_0 到 ε 的方向’在 SNR 加权下的表示）。三者的转换关系——给定任一，可算出其他两个：x̂_0=(x_t−√(1−ᾱ_t)εθ)/√ᾱ_t；ε̂=(x_t−√ᾱ_t x̂_0)/√(1−ᾱ_t)；v=√ᾱ_t ε−√(1−ᾱ_t) x_0。各 SNR 区间的难度差异——这是选择参数化的关键：(a) 高噪声区（t 大、ᾱ_t≈0）——x_t≈ε（几乎纯噪声）；此时’预测 x_0′极难（因为 x_0 的信息几乎被噪声淹没，且 x_0 的尺度远大于 ε 的尺度），而’预测 ε’相对容易（ε 就是主要成分）。(b) 低噪声区（t 小、ᾱ_t≈1）——x_t≈x_0；此时’预测 ε’难（噪声很小、相对误差大），而’预测 x_0′容易。(c) v-prediction 的平衡——v 的尺度在各 SNR 区间更均匀（因为它对 x_0 与 ε 做了 SNR 加权），故各时间步的学习难度更均衡；这使训练更稳定、且对’零终端 SNR’等调度更鲁棒。实证——(a) DDPM 默认 ε-prediction；(b) Stable Diffusion 2.0 / 后续模型 用 v-prediction（因为配合’零终端 SNR’调度时更稳定）；(c) 在’高分辨率 + 少步采样’场景 v-prediction 优势更明显。与调度的耦合——(a) ε-prediction 在’ᾱ_T≈0’（零终端）时，最后几步的任务’过难’（几乎纯噪声下预测噪声仍难？——实际是’预测 ε 在高噪声区容易’）；(b) v-prediction 与’零终端 SNR’配合良好（因为 v 在极端 SNR 下仍有界）；(c) 故’调度 + 参数化’需联合选择。其他参数化——(a) score-prediction（与 ε 成正比）；(b) flow-prediction（Flow Matching 预测速度场，与 v 相关但定义不同）。实践建议——(a) 通用 → ε-prediction（最成熟）；(b) 高分辨率/少步/零终端 SNR → v-prediction；(c) 训练不稳定时可尝试切换参数化。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Mathematical Mechanism: 1. The Three Target Parameterizations: For latent state $x_t = alpha_t x_0 + sigma_t epsilon$, with $alpha_t^2 + sigma_t^2 = 1$ (where $alpha_t = sqrt{bar{alpha}_t}$ and $sigma_t = sqrt{1 – bar{alpha}_t}$): (a) $epsilon$-prediction: Network outputs $hat{epsilon} = epsilon_theta(x_t, t)$. Estimated data state is: $$hat{x}_0 = frac{x_t – sigma_t hat{epsilon}}{alpha_t}$$ (b) $x_0$-prediction: Network outputs $hat{x}_0 = f_theta(x_t, t)$. Estimated noise state is: $$hat{epsilon} = frac{x_t – alpha_t hat{x}_0}{sigma_t}$$ (c) $v$-prediction (Salimans & Ho, 2022): Network outputs velocity $v_theta(x_t, t)$, defined as the angular velocity along the rotation trajectory: $$v_t equiv alpha_t epsilon – sigma_t x_0$$ 2. Inversion Under v-Prediction: Given predicted velocity $hat{v} = v_theta(x_t, t)$, both $hat{x}_0$ and $hat{epsilon}$ are recovered symmetrically without dividing by near-zero coefficients: $$hat{x}_0 = alpha_t x_t – sigma_t hat{v}$$ $$hat{epsilon} = sigma_t x_t + alpha_t hat{v}$$ Because $alpha_t, sigma_t in [0, 1]$ and $alpha_t^2 + sigma_t^2 = 1$, the transformations are orthonormal rotations, guaranteeing bounded gradients across all $t in [0, T]$. 3. The Failure of $epsilon$-Prediction at Zero Terminal SNR: At $t=T$, zero terminal SNR sets $alpha_T = 0, sigma_T = 1$. Under $epsilon$-prediction: $$hat{x}_0 = lim_{alpha_t to 0} frac{x_t – sigma_t hat{epsilon}}{alpha_t} longrightarrow frac{0}{0} quad (text{Numerical Explosion})$$ Under $v$-prediction: $$hat{x}_0 = 0 cdot x_T – 1 cdot hat{v} = -hat{v}, quad hat{epsilon} = 1 cdot x_T + 0 cdot hat{v} = x_T$$ The formulation is perfectly stable and well-conditioned.
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
深度剖析与工程权衡：① ‘不同 SNR 区间难度不同’是理解参数化的钥匙——ε 在高噪声区易、x_0 在低噪声区易；v 平衡两者。② ‘v-prediction 与零终端 SNR 配合’——这是 Stable Diffusion 2.0 等采用 v 的原因；故’参数化 + 调度’需联合设计。③ ‘转换关系’的实用价值——三者可互相转换，故推理时可用任一参数化（甚至混合，如’对高噪声步用 ε、低噪声步用 x_0’）。④ ‘高噪声区的数值稳定性’——x_0-prediction 在高噪声区会输出’极端的 x_0 估计’（因为要除以极小的 √ᾱ_t），导致数值不稳；这是它的主要缺陷。⑤ ‘与感知质量的关系’——有研究表明不同参数化影响’不同噪声水平的生成质量’（如 v 在高噪声区更好）；故选择会影响细节表现。⑥ 面试要点——被问’v-prediction 是什么’，应给出’v=√ᾱ_t ε−√(1−ᾱ_t) x_0（速度）+ 各 SNR 区间难度更均衡 + 与零终端 SNR 配合好‘与’三种参数化可互相转换‘；能指出’x_0-prediction 在高噪声区极难且不稳’是深度理解的标志。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
Deep Dive & Engineering Trade-offs: ① Why Modern Diffusion Architectures Adopt v-Prediction: Early diffusion models (SD 1.5, standard DDPM) used $epsilon$-prediction because they avoided $t=T$ with non-zero terminal SNR schedules (e.g., $bar{alpha}_T approx 0.0047$). However, non-zero terminal SNR prevents the model from generating true pure black or pure white images. Modern foundation models (SDXL 2.1, Stable Diffusion 3, Flux) enforce zero terminal SNR and require $v$-prediction or Flow Matching to maintain training stability. ② Distillation Compatibility: Progressive distillation and consistency distillation collapse when applied to $epsilon$-prediction models due to accumulated step errors at high noise. Distillation algorithms converge with $2text{–}4times$ higher stability when operating on $v$-prediction checkpoints. ③ Flow Matching Equivalence: In continuous Flow Matching with linear interpolation $x_t = (1-t) x_0 + t x_1$, the target vector field is $u_t(x) = frac{d x_t}{dt} = x_1 – x_0$. Velocity $v$-prediction is the trigonometric spherical equivalent of linear flow matching velocity fields. ④ Interview Strategy: Formulate the three parameterizations ($,epsilon,$, $,x_0,$, $,v,$), write down the velocity definition $v = alpha_t epsilon – sigma_t x_0$, show how $hat{x}_0$ and $hat{epsilon}$ are recovered via orthonormal rotation without division, and explain why $epsilon$-prediction breaks down at zero terminal SNR.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 在高噪声区用 x_0-prediction（极难且数值不稳）
- ⚠️ 认为参数化只是’记法不同’（影响训练稳定性）
English Pitfalls:
– Attempting to implement zero terminal SNR noise schedules on $epsilon$-prediction models, causing training divergence due to division by zero
– Swapping a pre-trained $epsilon$-prediction model into a $v$-prediction inference pipeline without mathematical conversion
– Assuming $v$-prediction changes the model backbone architecture; it modifies only the target regression variable and loss formulation
六、高频深度面试追问与预测 (Follow-Up Questions)
- 为什么 x_0-prediction 在高噪声区难？
- Why does $epsilon$-prediction become mathematically ill-conditioned when enforcing zero terminal SNR ($,bar{alpha}_T = 0,$)?
- v-prediction 什么时候更有优势？
- How does velocity $v$-prediction establish a mathematical bridge between discrete DDPM and continuous Flow Matching?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
去噪扩散概率模型 (DDPM)：前向加噪马尔可夫链与变分下界 (ELBO) 推导(DDPM: Forward Markov Noise & ELBO Denoising Derivation) - 🗺️ 知识图谱模块：
多模态与扩散模型导图
🔬 算法科学家与机器学习深度考察全量题库 (Science Depth)
本题收录于 TalentMe 算法科学家深度考察真题库 (Science Depth)。全库共 856 道硬核考点，深度覆盖数学统计、经典ML、深度学习、Transformer、大语言模型、多模态、推荐系统与 MLOps。支持 Jev 面经智能匹配、一键离线单文件 HTML 手册导出并直连 Obsidian 本地记忆。