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：Hard
一、核心一句话结论 (One-Sentence Summary)
SDE 采样（DDPM/Euler-a）不可复现；确定性采样（DDIM/DPM-Solver）给定种子可复现；需固定种子与确定性算子。
Deterministic diffusion ODE samplers guarantee mathematical reproducibility given identical noise seeds, while stochastic SDE samplers inject continuous random perturbations that produce varied outputs.
二、核心考点要义 (Key Insights)
- 📌 确定性采样：给定 x_T（种子）→ 唯一结果（可复现）
- 📌 随机采样：每步注入噪声 → 同种子不同结果
- 📌 可复现需：固定种子 + 确定性算子 + 固定采样器/步数/CFG
English Insights:
– Deterministic sampling (DDIM, DPM-Solver): ODE flow establishes a deterministic 1-to-1 bijection between initial noise $x_T$ and generated image $x_0$
– Stochastic sampling (DDPM, Euler-a): continuously injects Wiener noise perturbations at every step, creating high sample diversity but non-reproducible trajectories
– Hardware and kernel non-determinism: identical seeds can still produce diverging images across different GPU architectures, CUDA versions, or non-deterministic atomic operations
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$text{det}: x_Tto x_0 text{unique};qquad text{stoch}: x_{t-1}=f(x_t)+sigma_t z, zsimmathcal{N}(0,I)$$
数学机理：随机性的来源——(1) 起始噪声 x_T——从 N(0,I) 采样（种子决定）；(2) 采样过程中的噪声注入——DDPM/Euler-a 每步加 σ_t·z（见 DDIM vs DDPM 题）。确定性采样（DDIM/DPM-Solver）——σ_t=0，故 x_T→x_0 是确定映射：给定同一个 x_T（由种子决定），结果完全一致（可复现）。随机采样（DDPM/Euler-a）——每步注入噪声；即使固定起始种子，中间的噪声采样仍会引入随机性（除非也固定每个中间噪声的种子）；故’同种子不同结果’是常态。可复现的条件——(a) 固定起始种子（决定 x_T）；(b) 确定性采样器（或固定所有中间噪声的种子）；(c) 确定性算子（GPU 上的原子操作/某些 kernel 非确定，见 M3 的可复现性题）；(d) 固定的超参（步数、CFG s、调度、分辨率）；(e) 固定的模型版本与精度（不同版本/精度产生不同结果）。为什么可复现性重要——(a) 调试——无法复现就无法定位问题；(b) 产品——用户期望’同输入同输出’（尤其需要审计/合规的场景）；(c) A/B 测试——需控制变量（否则无法归因）；(d) 学术对比——需公平比较。不可复现的场景与应对——(a) 追求多样性 → 随机采样（不同种子给不同结果）；(b) 追求可复现 → 确定性采样 + 固定种子；(c) 折中 → 用’少量随机步’（如只在中间几步加噪）。与’编辑’的关系——确定性采样支持 inversion（把真实图像反推到 x_T，再修改 prompt 重新生成）；这是图像编辑（SDEdit、prompt-to-prompt、DDIM inversion）的基础；随机采样不可逆（难以精确反推）。工程实践——(a) 记录完整配置（种子、采样器、步数、s、模型版本）；(b) 用确定性采样做回归测试；(c) 多随机种子评估（因为单次结果有方差，见 M5 的统计显著性问题）。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Mathematical Mechanism: 1. Deterministic Mapping: In a Probability Flow ODE solver: $$x_0 = Phi(x_T; theta, c)$$ where $Phi$ is the numerical integration operator. Because the ODE vector field is Lipschitz continuous, the trajectory $x(t)$ is uniquely determined by initial value $x_T sim mathcal{N}(0, I)$ and model parameters $theta$. If the random seed generating $x_T$ is fixed, $x_0$ is mathematically invariant: $$text{Var}big(x_0 mid x_T, theta, cbig) = 0$$ 2. Stochastic SDE Trajectory: In ancestral or reverse SDE sampling: $$x_{t-1} = mu_theta(x_t, t) + sigma_t z_t, quad z_t sim mathcal{N}(0, I)$$ The total trajectory entropy is: $$H(x_0 mid x_T) = sum_{t=1}^T H(z_t) > 0$$ Even if initial seed $x_T$ is locked, intermediate perturbations $z_t$ generate a distribution of outputs. 3. Sources of Hardware Non-Determinism: Even with fixed seeds, floating-point non-associativity causes divergence: $$(a + b) + c neq a + (b + c) quad (text{IEEE 754})$$ (a) Atomic Add Operations: CUDA atomic additions in FlashAttention or cuBLAS accumulate tensor values in non-deterministic thread order. (b) Tensor Core Precision: TF32 or FP16 rounding differences between Ampere (A100) and Hopper (H100) architectures create minor numerical perturbations $epsilon approx 10^{-5}$ at $t=T$ that amplify exponentially along non-linear diffusion trajectories.
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 ‘确定性采样可复现、随机采样不可’——这是两者的核心工程差异；面试中能指出是深度理解的标志。② ‘可复现需多层固定’——不只是种子（还有算子确定性、超参、模型版本）；这是’可复现性’的完整清单。③ ‘确定性采样支持编辑’——inversion 需要可逆映射；这是 DDIM 在编辑任务中流行的原因。④ ‘多样性需靠不同种子’——确定性采样下，多样性来自 x_T 的随机性；故’同 prompt 生成多图’需换种子。⑤ ‘单次结果有方差’——即使确定性采样，’换种子’也会给出不同结果；故评估需多种子（避免’运气好’的结论）。⑥ 面试要点——被问’扩散采样可复现吗’，应给出’确定性采样可复现（固定种子 + 确定性算子）+ 随机采样不可 + 可复现的完整清单‘与’inversion/编辑依赖确定性‘；能指出’评估需多种子’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 Butterfly Effect in Diffusion Trajectories: Because the reverse trajectory unrolls through 30-50 non-linear deep network layers, a tiny numerical discrepancy of $10^{-6}$ at step 1 expands across steps: $|Delta x_t| propto e^{int L(t) dt}$. By step 0, two runs on different GPU types with the exact same seed may generate images with different hair configurations or background elements. Achieving true bitwise reproducibility requires setting `torch.use_deterministic_algorithms(True)` and locking identical GPU hardware. ② Seed Invariance in Creative Workflows: Deterministic ODE solvers are essential for user-facing creative tools: users expect that locking the seed allows them to tweak single words in the prompt (‘change blue jacket to red jacket’) without altering the subject’s face or composition. ③ Batch Size Invariance Trap: Generating 4 images in a single batch (`batch_size=4`) versus generating 4 images sequentially with sequential seeds (`batch_size=1`) can produce different images if the random number generator (RNG) strides differently across batch tensor dimensions. Using individual generator instances per sample (`torch.Generator(device=’cuda’).manual_seed(…)`) guarantees batch-size independence. ⑤ Interview Strategy: Contrast deterministic ODE mapping $Phi(x_T) to x_0$ against stochastic SDE entropy accumulation, explain floating-point atomic add non-determinism, describe the chaotic amplification of minor numerical shifts, and present per-sample generator seeding.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 用随机采样做精确的图像编辑（不可逆）
- ⚠️ 单种子评估（忽略结果的方差）
English Pitfalls:
– Assuming identical random seeds will produce bitwise identical images across different GPU architectures (e.g., RTX 3090 vs A100)
– Using stochastic samplers (Euler-a) for image editing workflows that require deterministic latent seed consistency
– Failing to instantiate independent per-sample PyTorch torch.Generator instances, causing batch-size-dependent seed drift
六、高频深度面试追问与预测 (Follow-Up Questions)
- 为什么’同种子不同结果’？
- Why does a minor floating-point discrepancy ($10^{-5}$) at early sampling steps amplify into major visual divergence in generated images?
- 可复现性对产品与调试的重要性？
- How does instantiating dedicated per-sample PyTorch CPU/CUDA generators guarantee batch-size independent reproducibility?
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。