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：Medium
一、核心一句话结论 (One-Sentence Summary)
s 越大越贴合 prompt 但多样性下降、易过饱和；s 太小则忽略 prompt；需按任务调（常用 5~12）。
The guidance scale $s$ balances sample diversity against prompt adherence, where excessive scaling induces high-frequency unnatural contrast, mode collapse, and numerical latent over-saturation.
二、核心考点要义 (Key Insights)
- 📌 s 小：忽略 prompt、结果多样但不对题
- 📌 s 大：贴合 prompt 但过饱和（色彩过艳、结构僵硬）、多样性低
- 📌 常用 5~12（SD 默认 7.5）；需按任务与模型调
English Insights:
– Guidance scale spectrum: $s=1.0$ (pure conditional, high diversity, weak prompt fidelity); $s=4.0text{–}7.5$ (optimal visual quality and prompt adherence); $s > 15$ (severe over-saturation, color burn, and anatomical distortion)
– Density ratio sharpening: CFG implicitly samples from an artificially sharpened conditional distribution $,p_s(x mid c) propto p(x) left( frac{p(x mid c)}{p(x)} right)^s,$, collapsing sample variance onto narrow mode peaks
– Mitigation mechanisms: CFG Rescaling (rescaling guided latents to match unconditional variance) and Dynamic Thresholding (clamping extreme percentile activations)
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$text{s}uparrowRightarrowtext{prompt adherence}uparrow, text{diversity}downarrow, text{artifacts}uparrow$$
数学机理：s 的双向效应——CFG 的组合为 ε̂=ε(∅)+s(ε(c)−ε(∅))。(1) s 太小（如 <3）——条件影响弱，生成结果’接近无条件分布’（多样但与 prompt 关联弱）；s=1 时完全无引导。(2) s 适中（如 7~12）——良好的 prompt 遵循与合理多样性（默认区间）。(3) s 太大（如 >15）——(a) 过饱和（oversaturation）——颜色过艳、对比度过高（因为放大了’超出数据分布的方向’，导致像素值饱和）；(b) 结构僵硬/重复——过度优化条件导致’模式坍缩’（不同种子给出相似结果）；(c) 不自然——违背物理/解剖（如多手指、扭曲）；(d) 多样性崩塌。为什么 s 过大产生伪影——从 score 视角：s>1 时组合为 (1−s)∇log p(x)+s∇log p(x|c)，其中 (1−s)<0 意味着减去无条件分布（’远离通用图像’）；s 越大’远离’越多，导致样本落在数据分布的低密度区域（不自然、过饱和）。如何选择 s——(a) 经验区间（SD 1.5/SDXL 常用 7~7.5；SD3/Flux 等新模型常用 3.5~5，因为它们训练更好、对引导依赖低）；(b) 按任务调——(i) 需要严格遵循 prompt（如产品图）→ s 大一些；(ii) 需要多样性/艺术性 → s 小一些；(c) 动态 s——不同时间步用不同 s（如前期大、后期小，避免过饱和）；(d) 自动选择——用’CFG rescale’（对引导后的结果做尺度校正，避免过饱和）或’自适应引导’。缓解过饱和的技巧——(a) CFG rescale（–cfg-rescale）——把引导后的预测的标准差重新缩放到’无条件预测的标准差’，避免过饱和；(b) 动态阈值（dynamic thresholding）（Imagen 的做法）——对预测的 x_0 做分位数截断；(c) 更小的 s + 更好的模型（新模型对引导的依赖更低）。与其他旋钮的交互——(a) 步数——s 大时可能需要更多步（因为每步’走得更远’）；(b) 调度——不同调度的’信噪比分布’影响 s 的效果；(c) 参数化（ε/v）——影响引导的数值行为。实践建议——(a) 从 7.5 起步（旧模型）或 4~5（新模型）；(b) 若过饱和则 (i) 降低 s、(ii) 开 CFG rescale、(iii) 用动态阈值；(c) 若忽略 prompt 则提高 s（但注意上限）；(d) 记录 s（因为它显著影响结果，评测时需报告）。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Mathematical Mechanism: 1. Effective Density Under CFG: The continuous probability distribution targeted by CFG with scale $s$ is: $$p_s(x mid c) propto p(x)^{1 – s} cdot p(x mid c)^s = p(x) left( frac{p(x mid c)}{p(x)} right)^s$$ When $s > 1$, probability mass is concentrated exponentially on regions where the likelihood ratio $frac{p(x mid c)}{p(x)}$ is maximized. As $s to infty$, $p_s(x mid c)$ collapses to a set of Dirac delta peaks (extreme mode collapse). 2. Latent Value Drift and Over-Saturation: In latent diffusion, latents $z_t$ are assumed to maintain standard normal statistics $z_t sim mathcal{N}(0, I)$. When $s$ is large (e.g., $s=15$), the vector difference $Delta = s(epsilon(c) – epsilon(emptyset))$ pushes latent values far outside the $[-3, 3]$ standard normal range into extreme values ($|z| > 10$). When decoded by the VAE decoder, these out-of-distribution latents clip against pixel boundaries $[0, 255]$, producing harsh contrast, plasticky skin textures, and bright halos. 3. CFG Rescale (Lin et al., 2024): Normalizes the guided prediction $hat{epsilon}$ to match the standard deviation of conditional prediction $epsilon(c)$: $$sigma_c = text{std}(epsilon(c)), quad sigma_{text{guided}} = text{std}(hat{epsilon})$$ $$hat{epsilon}_{text{rescaled}} = phi cdot left( hat{epsilon} cdot frac{sigma_c}{sigma_{text{guided}}} right) + (1 – phi) hat{epsilon}$$ with factor $phi approx 0.7$, eliminating over-saturation without sacrificing prompt adherence.
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 ‘s>1 意味着减去无条件分布’——这是理解过饱和的钥匙：s 越大’越远离通用分布’，故落到低密度区（不自然）。② ‘新模型对引导依赖更低’——因为训练更好、条件遵循更强；故 s 的默认值在下降（7.5 → 3.5~5）。③ ‘CFG rescale 是实用技巧’——它几乎零成本地缓解过饱和（把引导结果的尺度拉回无条件预测的尺度）；应在 s 较大时开启。④ ‘多样性 vs 遵循的权衡是根本——CFG 无法同时最大化两者；故需按应用选点（产品图重遵循、艺术创重多样）。⑤ ‘评测必须报告 s’——不同 s 的结果差异巨大；不报告 s 的评测不可比。⑥ 面试要点——被问’guidance scale 怎么调’，应给出’s 小忽略 prompt、s 大过饱和与多样性崩塌、常用 7~12（新模型 3.5~5）‘与’过饱和的机制（减去无条件分布 → 低密度区）+ 缓解（CFG rescale / 动态阈值）‘；能指出’新模型 s 默认值下降’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 Modern Trend of Declining Default Guidance Scales: Early models (SD 1.5) required aggressive guidance scales ($s=7.5text{–}9.0$) to overcome weak cross-attention text conditioning. Modern foundation models with massive text encoders (SD3 with T5-XXL, Flux.1) possess vastly superior conditional alignment, operating optimally at modest guidance scales ($s=3.0text{–}4.5$), or even relying on distilled guidance-free samplers ($s=1.0$). ② Dynamic Thresholding (Imagen): Saharia et al. demonstrated that when $|x_0| > 1.0$, instead of static clipping to $[-1, 1]$ (which creates flat blown-out highlights), dynamic thresholding computes the $99.5$th percentile value $s_p = text{percentile}(|x_0|, 99.5)$ and scales the tensor by $max(1.0, s_p)$, preserving edge gradients and structural contrast. ③ Diversity vs Adherence Frontier: For creative concept brainstorming or asset exploration, lower scale ($s=3.0text{–}5.0$) provides rich variations in camera angle, lighting, and composition. For strict instruction execution and typography generation, higher scale ($s=6.0text{–}8.0$) paired with CFG Rescale is optimal. ⑤ Interview Strategy: Formulate the sharpened density $p_s(x mid c) propto p(x) [p(x mid c)/p(x)]^s$, explain the mathematical origin of over-saturation from out-of-distribution latent variance, derive CFG Rescale, and contrast Imagen’s dynamic thresholding against static clipping.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 把 s 调得过大（过饱和、多样性崩塌）
- ⚠️ 评测时不报告 s（结果不可比）
English Pitfalls:
– Increasing guidance scale beyond 12 without enabling CFG Rescale or Dynamic Thresholding, causing severe color burn and distorted anatomy
– Using early SD 1.5 guidance scales ($s=8.0$) on modern DiT architectures (SD3, Flux), which are tuned for lower scales ($s=3.5$)
– Applying static hard clipping ($[-1, 1]$) to over-saturated latents, creating flat untextured specular blowouts
六、高频深度面试追问与预测 (Follow-Up Questions)
- 为什么 s 过大产生过饱和？
- Why does increasing the CFG scale s cause latent activations to drift outside the pre-trained distribution range of the VAE decoder?
- 如何自动选 s？
- How does Dynamic Thresholding adaptively scale extreme latent percentiles to prevent specular highlight clipping?
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。