所属模块：
M6 · 多模态与生成模型 (Multimodal & Generative Models)| 专题分类：引导与采样 (Guidance & Fast Sampling (CFG / DDIM))| 难度等级：Easy
一、核心一句话结论 (One-Sentence Summary)
用’条件预测’与’无条件预测’的差放大条件影响：ε̂=ε(∅)+s(ε(c)−ε(∅))；s>1 增强条件遵循、降低多样性。
Classifier-Free Guidance steers diffusion generation toward prompt conditions by linearly extrapolating between conditional and unconditional score predictions, eliminating external classifier networks.
二、核心考点要义 (Key Insights)
- 📌 同时算’有条件’与’无条件’预测，用差作为’条件方向’
- 📌 s 是引导强度：s=1 无引导、s>1 增强条件、s=0 无条件
- 📌 s 越大越贴合 prompt，但多样性下降、可能过饱和
English Insights:
– Extrapolation formula: combines unconditional prediction $,epsilon_theta(x_t, emptyset),$ and conditional prediction $,epsilon_theta(x_t, c),$ via guidance scale $s$: $,hat{epsilon} = epsilon_theta(x_t, emptyset) + s cdot (epsilon_theta(x_t, c) – epsilon_theta(x_t, emptyset)),$
– Joint training protocol: randomly drops the conditioning prompt $c$ with probability $p_{text{uncond}} in [0.1, 0.2]$ during training, replacing it with empty token $emptyset$
– Score sharpening effect: setting $s > 1$ amplifies conditional modes and suppresses unconditional background density, boosting text alignment and visual clarity at the expense of sample diversity
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$hatepsilon=epsilon_theta(x_t,varnothing)+sleft(epsilon_theta(x_t,c)-epsilon_theta(x_t,varnothing)right)$$
数学机理：CFG（Classifier-Free Guidance，Ho & Salimans 2022）——(1) 两次预测——同一次前向中（或用 batch 并行）计算 (a) 条件预测 εθ(x_t,c)（c 为文本 prompt）；(b) 无条件预测 εθ(x_t,∅)（条件置空）。(2) 组合——ε̂=εθ(x_t,∅)+s·(εθ(x_t,c)−εθ(x_t,∅))，其中 s 是引导强度。(3) 采样——用 ε̂ 代替 εθ 做去噪步。为什么’差’是条件方向——从 score 视角：ε 与 score 成正比（∇log p=−ε/√(1−ᾱt)），故’条件预测 − 无条件预测’正比于’∇log p(x|c)−∇log p(x)’——即条件对分布的’影响方向’（’要让图像更像 c 应该往哪走’）。故 CFG 等价于：∇log p̂=∇log p(x)+s·(∇log p(x|c)−∇log p(x))=(1−s)∇log p(x)+s·∇log p(x|c)——即把’无条件’与’有条件’的 score 做线性组合（当 s>1 时，条件 score 被放大、无条件被’负向’计入）。(4) s 的效应——(a) s=1 → 退化为’纯条件预测’（无引导）；(b) s>1 → 增强条件遵循（更贴合 prompt）但多样性下降、可能’过饱和’（色彩过艳、结构僵硬）；(c) s=0 → 无条件生成（忽略 prompt）；(d) 常用 s=7~12（Stable Diffusion 默认 7.5）。(5) 训练方式——训练时随机丢弃条件（如 10% 概率把 c 置空）并让模型同时学’有条件’与’无条件’预测；这样推理时才有 εθ(x_t,∅) 可用（这是 CFG 的关键训练技巧）。为什么有效（直觉）——(a) 无条件预测提供’基础分布’（通用的、符合数据的图像）；(b) 条件预测提供’朝 prompt 的方向’；(c) 放大这个方向（s>1）使结果更强地满足 prompt——代价是偏离数据分布（过度放大导致不自然）。与 classifier guidance 的对比——classifier guidance 需额外训练一个分类器并对 score 求梯度（见下一题）；CFG 不需要额外模型（用同一个模型的条件/无条件预测）。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Mathematical Mechanism: 1. Implicit Classifier Bayes Decomposition: By Bayes’ rule, the conditional score decomposes into unconditional score plus classifier gradient: $$nabla_x log p(x mid c) = nabla_x log p(x) + nabla_x log p(c mid x)$$ Ho & Salimans (2021) observed that the implicit classifier gradient can be expressed purely using conditional and unconditional generative models: $$nabla_x log p(c mid x) = nabla_x log p(x mid c) – nabla_x log p(x)$$ 2. Guided Score Formulation: Substituting this implicit classifier gradient into guided score with temperature/scale multiplier $s$: $$tilde{nabla}_x log p_s(x mid c) = nabla_x log p(x) + s cdot nabla_x log p(c mid x)$$ $$= nabla_x log p(x) + s big[ nabla_x log p(x mid c) – nabla_x log p(x) big] = (1 – s) nabla_x log p(x) + s nabla_x log p(x mid c)$$ 3. Noise Predictor Parameterization: Since score relates to predicted noise via $nabla_{x_t} log p(x_t) = – frac{epsilon_theta}{sqrt{1-bar{alpha}_t}}$, the guided noise prediction is: $$hat{epsilon}_theta(x_t, c) = epsilon_theta(x_t, emptyset) + s cdot big( epsilon_theta(x_t, c) – epsilon_theta(x_t, emptyset) big) = (1 – s) epsilon_theta(x_t, emptyset) + s epsilon_theta(x_t, c)$$ When $s = 1$, generation reduces to standard conditional diffusion without guidance. When $s > 1$, the vector difference $(epsilon(c) – epsilon(emptyset))$ pushes the generation trajectory away from generic unconditional data toward the conditioned prompt manifold.
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 ‘随机丢弃条件训练’是 CFG 的前提——若不这样做，模型没有’无条件预测’的能力，CFG 无法使用；这是实现中最易漏的细节。② ‘score 线性组合’是 CFG 的理论解释——它把 CFG 从’经验技巧’提升为’有理论依据的方法’；面试中能给出这一解释是深度理解的标志。③ ‘s>1 的代价’——过度引导会 (a) 降低多样性（不同种子给出相似结果）、(b) 产生’过饱和/过锐’的伪影（因为放大了’超出数据分布的方向’）、(c) 损害’提示之外的合理性’（如生成不合物理的物体）。④ ‘CFG 的成本 ×2’——每步需两次前向；这是扩散推理的主要成本之一（见后续题）。⑤ ‘CFG 在 DiT 中的实现’——DiT 用 adaLN 注入条件，故’无条件’即把条件设为 null/零；实现上常用’batch 拼接’（条件与无条件合成一个 batch）一次前向完成（见 DiT 的 CFG 题）。⑥ ‘与负向提示的关系’——CFG 可推广为’多条件’（如正向 + 负向）；负向提示即’用负向文本作为’无条件’的位置’（见负向提示题）。⑦ 面试要点——被问’CFG 是什么’，应写出组合公式 + score 线性组合的解释 + s 的效应 + 训练时随机丢条件；能指出’CFG 的成本 ×2’与’过度引导的伪影’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 The Dual Forward Pass Tax: Standard CFG requires evaluating the neural network twice per sampling step: once for $epsilon_theta(x_t, c)$ and once for $epsilon_theta(x_t, emptyset)$. In production serving, the batch size is doubled ($2B$), doubling latency and FLOPs per request. ② The Conditioning Dropout Rate ($p_{text{uncond}}$): Setting $p_{text{uncond}}$ too low ($ 30%$) degrades conditional text adherence. Industry standard consensus fixes $p_{text{uncond}} = 0.10text{–}0.15$. ③ Negative Prompting as Dynamic Unconditional Steering: In standard CFG, the unconditional branch uses an empty string $emptyset$. In creative text-to-image tools (Automatic1111, Midjourney), users provide a ‘negative prompt’ $c_{text{neg}}$ (‘blurry, distorted, extra limbs’). Replacing $epsilon(emptyset)$ with $epsilon(c_{text{neg}})$ actively repels generated samples from the negative semantic manifold. ④ Interview Strategy: Formulate the Bayesian score decomposition, derive the linear combination $hat{epsilon} = epsilon(emptyset) + s(epsilon(c) – epsilon(emptyset))$, explain how training with condition dropout ($p=0.1$) enables single-model dual capability, and explain the negative prompting mechanism.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 训练时不随机丢条件（推理无 ε(∅) 可用）
- ⚠️ 把 s 调得过大（过饱和、多样性崩塌）
English Pitfalls:
– Setting guidance scale $s=1.0$ and expecting strong text prompt adherence; without guidance ($s > 1$), complex prompt composition collapses
– Forgetting to train with unconditional condition dropout ($p_{text{uncond}}$), making CFG extrapolation impossible at inference time
– Assuming CFG adds no serving compute; CFG doubles the effective forward-pass batch size at every sampling step
六、高频深度面试追问与预测 (Follow-Up Questions)
- 为什么’差’能代表条件方向？
- How does substituting a negative text prompt in place of the empty condition $emptyset$ in CFG mathematically repel unwanted visual features?
- CFG 与 classifier guidance 的关系？
- Why is conditioning dropout ($p_{text{uncond}} approx 0.1$) during training mandatory for enabling Classifier-Free Guidance at inference time?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
扩散引导与加速采样：Classifier-Free Guidance (CFG) 与 DDIM 确定性采样(Classifier-Free Guidance (CFG) & Accelerated DDIM Sampling) - 🗺️ 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。