概念

Velocity Field

— 时间索引的向量场 $u_t(x)$,告诉每个粒子此刻该往哪走。

很多人初接触 velocity field 会下意识把它想成"噪声方向"——"沿着这个方向走就能去噪"。这是被 score $\nabla \log p_t$ 的去噪直觉污染过的理解。velocity field 不是"噪声方向"——它是一个时间索引的向量场(TIME-INDEXED vector field),它告诉你"在时刻 $t$、位置 $x$ 上,该把粒子往哪推、推多快"。它和"去噪"无关,和"轨迹"无关,它只是一张会随时间变形的速度地图。粒子的轨迹是这张地图上的"被动产物"——你把一颗石子扔进河里,石子走的曲线是轨迹;河水本身在每个位置、每个时刻流动的方向和速度,才是 velocity field。

"想象一个盒子里有非常多粒子 $P_1, P_2, P_3, \dots$。出于某种原因这些粒子在盒子里运动——这是粒子 $P_1$ 的轨迹,那是 $P_2$ 的轨迹;类似地,盒子里每一个点都有它自己的轨迹。"
—— Vizuara · Principles of Diffusion · L10 @00:05

Vizuara 的 Dr. Raj 在 Lecture 9–10 选了一个连续介质力学的入口:把数据生成当成"概率流体"被风搬运。盒子里粒子各自跑各自的轨迹,可看似杂乱,但你只要写下"每个位置每个时刻的速度" $u_t(x)$——也就是速度场——所有粒子的运动就被完整刻画了。这套图像借自 Euler 描述流体的视角:你不追每个粒子(Lagrangian),而盯着空间每一点看"现在水从哪流来、往哪流去"(Eulerian)。velocity field 就是 Eulerian 速度,把它放进 ODE $\dot x = u_t(x)$ 一积分,就得到任意起点出发的粒子轨迹。

"想象风把太平洋的水搬运到东海岸——速度场告诉你每个位置的风往哪吹、吹得多快。你只要顺着它积分,水就会从西海岸流到东海岸。这就是 ODE 把噪声变成图像的方式。"
—— CMU 10-799 S26 · L5 @08:10(改述)

CMU 的 Tianqi 在 Lecture 5 用了另一个更具体的"风搬水"比喻——风把太平洋的水搬到东海岸;velocity field 就是这阵会随时间变化的风。两条比喻指向同一件事:velocity field 是搬运概率质量的"风"或"水流",粒子是被动跟随的微粒。在机器学习的上下文里,这件事对应一个具体得不能再具体的目标:把 prior $z_0 \sim p_0 = \mathcal{N}(0, I)$ 当作"西海岸的水",把数据 $x = z_1 \sim p_1 = p_{\text{data}}$ 当作"东海岸要送达的水",神经网络 $v_\theta(x, t)$ 拟合的就是这阵风——训完之后任意噪声样本顺风一积分就变成图像。

注意 velocity field 与三个相邻概念的边界。它不是 score——score 是对数密度的空间梯度 $\nabla \log p_t(x)$,量纲 $1/[x]$;velocity 是位移除以时间,量纲 $[x]/[t]$;只有在高斯扩散路径下两者可以仿射互换。它不是 drift——SDE $dx = f(x,t) dt + g(t) dW$ 里的 $f$ 是"漂移项",只是 velocity 的一部分,PF-ODE 把 drift 和 score 组合才得到完整 velocity。它不是噪声预测 $\epsilon_\theta$——$\epsilon$ 是单步残差,velocity 是连续时间瞬时变化率;前者依附于离散时间步、后者属于连续时间 ODE。把这三组概念混作一团,是初学 Flow Matching / Rectified Flow / EDM 时最常见的卡点。

还有一组语义陷阱值得提前打掉:velocity field 是欧拉视角(Eulerian)的,不是拉格朗日视角(Lagrangian)的。拉格朗日视角追踪每个粒子的运动 $\phi_t(z_0)$,问"$t = 0.5$ 时第 $i$ 号粒子跑到了哪里"——这是轨迹。欧拉视角则盯着空间的固定坐标,问"$t = 0.5$ 时位置 $(0.3, 0.7)$ 上的水流向哪流"——这是速度场。同一物理事实有两种表达,但速度场这个对象天然以欧拉视角定义,因为它要回答"任意位置任意时刻该往哪推"而非"具体某粒子的命运"。这层视角差异在工程上很重要:神经网络 $v_\theta(x, t)$ 训练时不知道输入的 $x$ 是从哪个 $z_0$ 出发的——它只能根据当前位置 + 当前时间输出速度,所以网络架构和损失函数都必须假定欧拉视角。这也解释了为什么不能"对每个训练样本单独学一条轨迹"——你训练集里只有 $(x_t, t)$ 对,没有完整轨迹标签。

第一层定义:从 ODE 出发

给定时间 $[0, 1]$ 上的一族向量场 $u_t: \mathbb{R}^d \to \mathbb{R}^d$,每个 $u_t(x)$ 是空间一点 $x$ 上的速度向量。把它放进常微分方程

$$\frac{dx}{dt} = u_t(x), \qquad x(0) = z_0 \sim p_0,$$

沿时间积分就得到一条flow $\phi_t(z_0)$。生成式建模要做的事情是:选一个易采样的初始分布 $p_0$(一般是 $\mathcal{N}(0, I)$),让 ODE 把 $p_0$ 推到 $t=1$ 时的目标分布 $p_1 = p_{\text{data}}$。神经网络要学的就是这个 $u_t$。Tianqi 在 L5 把它点得很清楚:

"基本上就是 Ricky 提的 continuous normalizing flow——构造一个一般模型,把初始分布(比如 Gaussian)的样本通过一个 ODE 积分变成目标分布(比如图像分布)的样本。这个 ODE 字面上就是:数据的无穷小变化应该跟随 velocity。"
—— CMU 10-799 S26 · L5 @08:05
Why:为什么用 ODE 而不是 SDE?SDE 每一步都加噪声,多次采样轨迹不同——不可复现、需要随机数生成器;ODE 完全确定,给定 $z_0$ 唯一对应一个 $z_1$,可复现、可反向、可做精确密度估计(CNF 用换元公式算 likelihood)。velocity field 是把 diffusion 从 SDE 视角搬到 ODE 视角的关键桥梁——这就是 PF-ODE(Probability Flow ODE)这个名字的来历:一个 ODE,它的边际分布和原 SDE 完全一样,但每个样本走的是确定性轨迹。还有一层工程理由:ODE 系统允许使用 RK4、Heun、Dormand-Prince 等高阶积分器,把全局误差从 $O(\Delta t)$ 降到 $O(\Delta t^4)$,这是 DPM-Solver、UniPC 等"少步采样器"得以存在的数学根基;SDE 离散误差受随机扰动主导,高阶展开收益很小。

第二层定义:连续性方程(质量守恒)

velocity field 不是随便挑的——它必须和概率路径 $p_t$ 相容,满足连续性方程(continuity equation)

$$\frac{\partial p_t(x)}{\partial t} + \nabla \cdot \bigl(p_t(x)\, u_t(x)\bigr) = 0.$$

翻译成大白话:单位时间内某区域概率密度的下降速率 = 流出该区域的概率通量(散度)。物理上等价于"质量不会凭空出现也不会凭空消失"——所有概率质量被严格地从一处搬运到另一处。CMU L13 把这件事用一句话概括:

"$Q_t$ 就是 flow——本质上就是 velocity。它告诉你从这个状态跳到那个状态的可能性有多大。所有的 flow 加起来必须为零——这就是质量守恒。"
—— CMU 10-799 S26 · L13 @06:09
Why:连续性方程是 velocity field 和概率路径之间的"硬约束"——给定 $p_t$,并非任意 $u_t$ 都对应这个 $p_t$;反过来,给定 $u_t$,它确定的 $p_t$ 也是唯一的(在 well-posed 条件下)。但要注意——同一个概率路径 $p_t$ 可以由无数个 velocity field 驱动(差一个"无散度"的旋转项 $w_t$,满足 $\nabla \cdot (p_t w_t) = 0$)。Flow Matching 训练里 L2 回归隐含地挑出其中"L2 模长最小"的那一个——对应物理上的最小动能传输。这就是为什么实践中 Flow Matching 的轨迹"比理论上更直、更短"。

第三层定义:Flow Matching 里的 marginal 速度

在 Flow Matching 框架下,marginal velocity field $u_t(x)$ 的精确表达是

$$u_t(x) = \mathbb{E}_{x_0, x_1 \sim p(x_0, x_1 \mid x_t = x)}\bigl[x_1 - x_0\bigr],$$

即"所有可能造出 $x_t = x$ 的端点对 $(x_0, x_1)$ 上的位移期望"。这个期望写不出闭式——你不知道后验 $p(x_0, x_1 \mid x_t = x)$。Lipman et al. 2023 用条件 Flow Matching 绕开:固定 $x_1$ 时条件路径取直线 $x_t = (1-t) x_0 + t x_1$,条件速度

$$u_t(x \mid x_1) = \frac{d x_t}{dt} = x_1 - x_0$$

是闭式的、与 $t$ 和 $x_t$ 无关,仅依赖端点之差。Tianqi 在 L5 演示过具体推导:

"当 $\Delta t \to 0$,你就得到这个 velocity——它告诉你无穷小步长下该往哪走。不论你在哪个 $t$,你永远是同一速度同一方向——就是 data 减 noise,$x_1 - x_0$。"
—— CMU 10-799 S26 · L5 @03:28
Why:L2 回归 $\|v_\theta(x,t) - (x_1 - x_0)\|^2$ 在最优解处自然收敛到上面那个条件期望——这是 CFM ≡ FM 等价定理的核心。结果是:你训练时永远在拟合"端点差"这个简单 target,但训完模型自动逼近 marginal velocity;marginal 在数据空间的轨迹交叉点会被"平均化",因此采样轨迹略有弯曲(这就是 Rectified Flow 要再做一次 reflow 拉直的原因)。这里有一个深层观察:训练 target $x_1 - x_0$ 本身有高方差——对同一个 $(x_t, t)$,不同 $(x_0, x_1)$ 组合给出的速度差异极大;网络学到的只是这些速度的条件均值,方差信息被丢掉了。所以 velocity field 是"对所有可能轨迹平均"的产物,并非单条轨迹的几何切线——这一点几乎所有学生第一次看 CFM 都会卡住,不点破就一直误以为"网络在学一条具体轨迹"。

第四层定义:PF-ODE 里的 velocity

如果你从 score-based diffusion 出发——前向 SDE $dx = f(x,t) dt + g(t) dW$——则其对应的 Probability Flow ODE(Song et al. 2021)使用的 velocity field 是

$$u_t(x) = f(x, t) - \tfrac{1}{2}\, g(t)^2\, \nabla \log p_t(x).$$

它是"漂移 drift" 和"分数 score" 的线性组合。从这个公式可以看出:velocity 在高斯扩散路径下确实是 score 的仿射函数——所以你才能在 DDPM 训好之后直接换 ODE 采样器(DDIM、DPM-Solver 都基于 PF-ODE),无需重训。但反之不成立:换成线性插值路径或 optimal transport coupling,PF-ODE 公式就不再有效,velocity 不再能写成 score 的仿射函数;这正是 Flow Matching 比 score matching"路径自由度更大"的根源。

Why:PF-ODE 这个组合 $f - \tfrac{1}{2} g^2 \nabla \log p_t$ 看似复杂,物理直觉其实很简单——SDE 的 drift $f$ 是确定性"推力"、扩散项 $g \, dW$ 是随机"抖动";把随机抖动用"score 指向高密度区"的等价漂移代替(这是 Fokker-Planck 方程的恒等变形),就把 SDE 改写成等价边际分布的 ODE。score 在此扮演的角色是"用确定性漂移替代随机扩散的等价物"。这件事的奇妙之处在于:原 SDE 单条样本的轨迹是布朗运动(处处不可微),PF-ODE 单条样本的轨迹是光滑曲线——但两者的边际分布 $p_t$ 完全相同。这是 velocity field 视角与 score 视角能互译的根本原因,也是为什么 EDM(Karras et al. 2022)能用 ODE 采样器拿到和 SDE 采样器一样的 FID 但少一个数量级的步数。
$t=0$ · 先验 $p_0$ (Gaussian) $t=0.5$ · 中间路径 $p_{0.5}$ $t=1$ · 数据 $p_1$ (MoG) 空间每一点的箭头 = $u_t(x)$;箭头随 $t$ 变化(左到右) $t=0$ 时箭头几乎平行向右;中段开始分叉指向不同 mode;$t=1$ 时局部收敛到 MoG 中心

这张图传达的核心信号:velocity field 是一张"会变形的箭头地图"。$t = 0$ 时它把所有粒子大致向右推(因为先验是各向同性 Gaussian,没有方向偏好);$t$ 增大时箭头逐渐"分叉"——靠上半空间的箭头指向上方的 mode、靠下半空间的箭头指向下方的 mode;$t \to 1$ 时所有箭头长度趋零、粒子收敛到数据 mode 上。这就是"velocity field 把 prior 搬向 MoG data" 在几何上的样子。

🟡 教材外补充:

实践中神经网络 $v_\theta(x, t)$ 的输入设计有几个隐藏细节,源讲师没展开但工程上重要。第一,$t$ 必须通过 sinusoidal embedding 或 Fourier features 嵌入而非直接拼接——直接拼一个 $[0, 1]$ 的标量会让网络在初期完全忽略时间信息(梯度被 $x$ 那 $\sim 10^3$ 个维度淹没)。SD3 用 16-dim Fourier features、Flux 用 256-dim sinusoidal——这是把 velocity field 训成"会随时间变形"的必要条件,否则你训出来的网络就退化成一个静态向量场。第二,$v_\theta$ 输出的量纲要和 $x$ 一致——常见 bug 是在 latent diffusion 里 latent 标度大约 $\pm 4$、velocity target $x_1 - x_0$ 标度也是 $\pm 5$,但如果你忘记取消 VAE 的 0.18215 缩放,速度幅度会差一个数量级,采样直接发散。第三,inference 时的 $t$ 调度可以与训练分离——训练用 $t \sim \mathcal{U}[0, 1]$ 或 logit-normal,但采样可以用 cosine spacing 或 Karras schedule,因为 velocity field 是连续的,对采样网格不挑剔。

把 velocity field 放到更大的概念地图上:它是连续标准化流(CNF, Chen et al. 2018)的核心对象,CNF 之前的离散标准化流(NICE、RealNVP)只能堆叠固定数量的可逆变换,CNF 用 ODE 让"层数"变成连续时间——velocity field 就是这个 ODE 的右端。Tianqi 在 L5 提到 "Ricky 提的 continuous normalizing flow" 指的就是这件事。但 CNF 原版用极大似然 + adjoint method 训练,需要在前向 + 反向各积一次 ODE,代价奇高、训不起来大模型。Flow Matching 的真正贡献是跳过似然:直接 L2 回归条件 velocity,把训练复杂度降到和 DDPM 一致——这是 velocity field 从"理论上漂亮的对象"变成"SD3 / Flux 工业默认骨架"的关键转折。从这个角度看,velocity field 自 2018 年起就在那里等着被正确使用,2023 年才被 Flow Matching 用对方式。

拖动滑块改变时间 $t$。不变量:箭头位置不变(它们锚定在空间网格上),但箭头方向和长度随 $t$ 改变——$t = 0$ 附近所有箭头几乎平行(先验各向同性),$t \to 1$ 时箭头按粒子所在位置"分叉"指向不同数据 mode。这正是 "velocity field is time-indexed" 的精确视觉含义:同一个 $x$ 在不同 $t$ 上速度完全不同。

t = 0.00
prior $p_0$ data $p_1$ (MoG, 2 modes) 每个箭头 = 该点 $u_t(x)$;同位置在不同 $t$ 速度不同

手算:从 velocity 到一步采样

设 $x_0 = -1$(噪声样本),$x_1 = 4$(数据样本),路径取线性插值。条件 velocity

$$u_t(x_t \mid x_1) = \frac{d}{dt}\bigl[(1-t)\cdot(-1) + t \cdot 4\bigr] = -(-1) + 4 = 5.$$

注意它不依赖 $t$ 也不依赖 $x_t$——线性插值下整条轨迹速度恒为 5。验证插值点:$t = 0.3$ 时 $x_t = 0.7 \cdot (-1) + 0.3 \cdot 4 = -0.7 + 1.2 = 0.5$;$t = 0.6$ 时 $x_t = 0.4 \cdot (-1) + 0.6 \cdot 4 = -0.4 + 2.4 = 2.0$。两点差 $2.0 - 0.5 = 1.5$,时间差 $0.6 - 0.3 = 0.3$,平均速度 $1.5 / 0.3 = 5$——和 $u_t = 5$ 一致。Euler 5 步采样($\Delta t = 0.2$)从 $x_0 = -1$ 出发:$-1 \to 0 \to 1 \to 2 \to 3 \to 4$,恰好命中 $x_1 = 4$。

对比 PF-ODE 视角:若用 variance-preserving 路径 $\alpha_t = \cos(\pi t / 2),\, \sigma_t = \sin(\pi t / 2)$,则 $\dot \alpha_t = -\tfrac{\pi}{2}\sin(\pi t / 2),\, \dot \sigma_t = \tfrac{\pi}{2}\cos(\pi t / 2)$,$t = 0.3$ 时 $\alpha = 0.891,\, \sigma = 0.454,\, \dot \alpha = -0.713,\, \dot \sigma = 1.399$,velocity $u_t(x \mid x_1) = \dot \alpha_t x_1 + \dot \sigma_t x_0 = -0.713 \cdot 4 + 1.399 \cdot (-1) = -2.85 - 1.40 = -4.25$——和线性路径的 $5$ 完全不同!这正是"velocity 依赖路径设计"的具体体现。

常见误区

  1. "velocity field 就是去噪方向 / 噪声的负梯度"。反例:在线性插值路径下 $u_t(x_t \mid x_1) = x_1 - x_0$,与"去噪"无关——它是从噪声指向数据的向量(升级方向,不是降级);而 DDPM 里 $\epsilon_\theta$ 预测的"噪声"是从数据指向噪声(要被减掉的)。两者方向相反、量纲不同——velocity 量纲是 $[x]/[t]$(位移除时间),$\epsilon$ 量纲是 $[x]$(位移);把它们混作"去噪方向"是 Flow Matching 初学者最常见的错。
  2. "velocity field 是一条粒子的轨迹切线"。反例:velocity field 是整个空间每一点的速度向量构成的地图,不是某条特定轨迹的属性。同一时刻 $t = 0.5$,位置 $x = (0.5, 0.5)$ 的速度 $u_{0.5}(0.5, 0.5) = (3, 2)$,位置 $x = (2, 2)$ 的速度 $u_{0.5}(2, 2) = (-1, 4)$——这俩同时存在、互不相干,是场的两个采样点。Vizuara 的盒子粒子图(L10 @00:05)就是这个意思:"每一个点都有它自己的轨迹"——即每一点都对应一个速度向量,整体是场而非曲线。
  3. "velocity field 在不同时刻不变(stationary)"。反例:取 MoG 数据分布(两个 mode 在 $(1, 0)$ 和 $(-1, 0)$)、Gaussian 先验。$t = 0$ 时 $u_0(0, 0) \approx (0, 0)$(原点是先验中心、整体没偏好);$t = 0.5$ 时 $u_{0.5}(0.01, 0) \approx (2, 0)$(轻微偏右就被右 mode 吸引);$t = 0.5$ 时 $u_{0.5}(-0.01, 0) \approx (-2, 0)$(轻微偏左就被左 mode 吸引);$t = 0.99$ 时 $u_{0.99}(1, 0) \approx (0, 0)$(已到达 mode 中心)。同一空间点在不同 $t$ 速度截然不同——所以 $u_t$ 必须显式带 $t$ 索引,神经网络必须输入 time embedding。把 velocity field 当作静态场是 Helmholtz 力学遗留下来的偏见,在概率传输里不成立。

相关概念

来源

  1. cmu-10799-diffusion-flow · L5 · 从 SDE 到 ODE 与 velocity 的引入 · (0:02)
  2. cmu-10799-diffusion-flow · L5 · "$\Delta t \to 0$ 得到 velocity" · (3:28)
  3. cmu-10799-diffusion-flow · L5 · Continuous Normalizing Flow / Ricky 提出 velocity 视角 · (7:10)
  4. cmu-10799-diffusion-flow · L5 · "风把太平洋的水搬到东海岸" 比喻 · (8:05)
  5. diffusion-principles-vizuara · L10 · 盒子粒子轨迹引入 velocity field · (0:05)
查看 120 条课程原文出处
CourseLectureTimeExcerpt
cmu-10799-diffusion-flow L10: CMU 10799 S26: Lecture 11 - Guest Lecture Linqi (Alex) Zhou from Luma AI - Diffusion & Flow Matching 32: 04 and G GMMN can also be seen as a special and G GMMN can also be seen as a special case of Imm. Uh case of Imm. Uh case of Imm. Uh you can easily see that when if you set you can easily see that when i…
cmu-10799-diffusion-flow L10: CMU 10799 S26: Lecture 11 - Guest Lecture Linqi (Alex) Zhou from Luma AI - Diffusion & Flow Matching 32: 53 essentially in this paper the drift field for their uh for their model is field for their uh for their model is field for their uh for their model is also parameterized via uh via MMD also parameteriz…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 00: 02 So last time we learned about how So last time we learned about how diffusion turn noise into data by diffusion turn noise into data by diffusion turn noise into data by gradually adding noise and gra…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 00: 52 by an SDE with this score function and then we can just do score matching to then we can just do score matching to then we can just do score matching to learn uh how to solve the uh the learn uh how t…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 01: 41 That's exactly correct. Okay. Why don't we just do linear interpolation of the we just do linear interpolation of the we just do linear interpolation of the two things, right? Um so exact how two thin…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 02: 26 this trajectory in this way. Um yeah so remember the same noise same data. remember the same noise same data. remember the same noise same data. Then learning the transformation along Then learning th…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 03: 28 progress. So like 0.25 time um times uh progress. So like 0.25 time um times uh data minus noise, right? Uh so basically data minus noise, right? Uh so basically data minus noise, right? Uh so basical…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 04: 20 this is literally what we need to learn, this is literally what we need to learn, right? This is like literally the only right? This is like literally the only right? This is like literally the only t…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 05: 17 Right? How do you progress through the chain? You just compute the displacement chain? You just compute the displacement chain? You just compute the displacement for x and then directly apply the for …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 06: 26 full matching. Class is over. Hey, no, not really. Okay, fine. Uh but no, not really. Okay, fine. Uh but no, not really. Okay, fine. Uh but basically the the question that we're basically the the ques…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 07: 10 people start to de develop this thing. Okay. So basically back in 2018 there Okay. So basically back in 2018 there Okay. So basically back in 2018 there this like pretty seminal paper um uh this like …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 08: 05 can just solve this OD by integrating can just solve this OD by integrating through uh the velocity and then you through uh the velocity and then you through uh the velocity and then you you'll be abl…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 08: 55 Anyone else I haven't talked to today? Nobody remember what is normalizing Nobody remember what is normalizing Nobody remember what is normalizing flow? >> You can reconstruct the process. >> You can …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 09: 50 invertible transformation because invertible transformation because everything is deterministic. everything is deterministic. everything is deterministic. Okay, any question? Okay, any question? Okay,…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 11: 01 no new mass uh is is going to get added and no mass is going to disappear and no mass is going to disappear and no mass is going to disappear getting dropped. So like translate it getting dropped. So …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 11: 50 probability flux. So imagine a flux is like the amount of flow the amount of like the amount of flow the amount of like the amount of flow the amount of water uh per unit per time that's going water u…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 12: 55 in a container. It's in like a contained in a container. It's in like a contained tank. So, the probability mass never tank. So, the probability mass never tank. So, the probability mass never change.…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 14: 00 in math in this way. So this is like in math in this way. So this is like basically like how much probability gets basically like how much probability gets basically like how much probability gets cha…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 15: 12 two assumptions. >> Uh two assumptions. One is that >> Uh two assumptions. One is that >> Uh two assumptions. One is that add up to one. The second one is that add up to one. The second one is that ad…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 16: 04 thing that like flow out >> as physically inclined. >> as physically inclined. >> as physically inclined. >> It is okay. >> It is okay. >> It is okay. >> I'm just a civil engineer. >> I'm just a civil…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 17: 08 the divergence connection again the divergence connection again >> yes I I think I missed a P here actually >> yes I I think I missed a P here actually >> yes I I think I missed a P here actually by t…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 18: 21 that like so this is the correct that like so this is the correct divergence. Sorry that that that one I divergence. Sorry that that that one I divergence. Sorry that that that one I missed P. Um but …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 19: 13 the log probability of of the log density. Okay. And this OD is actually density. Okay. And this OD is actually density. Okay. And this OD is actually equal to so the derivative of the total equal to …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 20: 52 So exactly so so how do you train your continuous how do you train your continuous how do you train your continuous normalizing flow from here? What do we normalizing flow from here? What do we normal…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 22: 04 we actually need to do numerical integration right. But this is at integration right. But this is at integration right. But this is at training time. So that just means that training time. So that jus…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 23: 07 numerical integration, right? So it still has a lot of steps. Can we do still has a lot of steps. Can we do still has a lot of steps. Can we do something that doesn't require a lot of something that d…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 24: 11 just need to make sure that the velocity actually match with the ground truth actually match with the ground truth actually match with the ground truth velocity. velocity. velocity. So you just do L2 …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 25: 12 changing every day be like actually I think that the the storm is moving north think that the the storm is moving north think that the the storm is moving north or maybe not. We don't know, right? So …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 25: 58 Yeah. That's right. That's right. You just go straight. just go straight. just go straight. Yeah, exactly. Uh so basically this is Yeah, exactly. Uh so basically this is Yeah, exactly. Uh so basically…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 26: 57 just start from some prior which we can just start from some prior which we can choose. So it doesn't matter we can choose. So it doesn't matter we can choose. So it doesn't matter we can choose the p…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 27: 43 the the data point then you can get a marginal marginal marginal uh yeah and also uh just the boundary uh yeah and also uh just the boundary uh yeah and also uh just the boundary condition essentially…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 29: 03 probability path and also just like how probability path and also just like how like and reweed by how likely in general like and reweed by how likely in general like and reweed by how likely in gener…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 30: 38 this average is like averaging by like this average is like averaging by like how likely is the data point given your how likely is the data point given your how likely is the data point given your cu…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 31: 44 regularization condition. This can can can be done. Uh and then uh basically can be done. Uh and then uh basically can be done. Uh and then uh basically you uh input your um because we know you uh inp…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 32: 46 can construct can generate the marginal property path which is what we want. property path which is what we want. property path which is what we want. Oh, I saw so many confusion. Oh, I saw so many co…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 33: 50 Think about that. Isn't that pretty Think about that. Isn't that pretty cool? Okay. Why? Why? Why is that? Okay. cool? Okay. Why? Why? Why is that? Okay. cool? Okay. Why? Why? Why is that? Okay. So, t…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 34: 46 Like basically you if you swap uh the Like basically you if you swap uh the dx1 term inside so you swap the order of dx1 term inside so you swap the order of dx1 term inside so you swap the order of t…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 35: 50 we're compressing the bullet. >> It's literally the same. It's literally >> It's literally the same. It's literally >> It's literally the same. It's literally the same. It's equivalent. It's not even …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 36: 53 Well we first have the the the endpoint distribution and because we're distribution and because we're distribution and because we're transforming it straight to the point transforming it straight to t…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 37: 52 >> Yeah. No no this is what we chose. This is a design choice. So that so this is a design choice. So that so this is a design choice. So that so this thing is generalizable to any kind of thing is ge…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 39: 09 we can directly sample from >> therefor you >> like previous lectures like gausian >> like previous lectures like gausian assumption is used like this closure on assumption is used like this closure o…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 40: 22 we talked about stuff when we u when we actually uh decrease the the velocity at actually uh decrease the the velocity at actually uh decrease the the velocity at which we adding noise. which we addin…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 41: 26 this particular path. Only on this this particular path. Only on this particular x0 to x1 path. But it's not particular x0 to x1 path. But it's not particular x0 to x1 path. But it's not in general. i…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 42: 24 >> because if you look at the if you look >> because if you look at the if you look at the this the the left hand side right at the this the the left hand side right at the this the the left hand side…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 43: 16 literally the same conference and they all got into the same conference. all got into the same conference. all got into the same conference. So yeah, so the first group is a group So yeah, so the firs…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 44: 00 but you may have this question, right? I feel like some people are having this feel like some people are having this feel like some people are having this question but they didn't ask. So like questio…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 44: 58 do is now let's say you you after training with the linear interpolation training with the linear interpolation training with the linear interpolation you already got this coupling. So the you already…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 45: 50 right? So you can map from one point in right? So you can map from one point in P 0 to another point in P1 and then you P 0 to another point in P1 and then you P 0 to another point in P1 and then you …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 46: 34 because you don't know like basically if you if you only do one step there can be you if you only do one step there can be you if you only do one step there can be like a lot of um uh uh like like lik…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 47: 37 >> The third image is the linear >> The third image is the linear interpolation of the new coupling. interpolation of the new coupling. interpolation of the new coupling. >> That what is that doing? >…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 48: 17 connected this side to that side and connected this side to that side and that side to this side. There's no that side to this side. There's no that side to this side. There's no there's no this anymo…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 48: 56 perfectly trained into this process, >> then you do it again. Yeah. Yeah. So, >> then you do it again. Yeah. Yeah. So, >> then you do it again. Yeah. Yeah. So, this is just you you get it this is just…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 49: 44 well. you're just like kind of random random walking a little bit. Whereas for random walking a little bit. Whereas for random walking a little bit. Whereas for flow matching, you don't really do flow…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 50: 27 question is is flow matching also the question is is flow matching also the same thing because if you think about it same thing because if you think about it same thing because if you think about it y…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 51: 12 can write everything as if it is flow can write everything as if it is flow matching as well and then you can matching as well and then you can matching as well and then you can actually just recover …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 52: 03 still need to take care of that part. But it turns out that we actually do But it turns out that we actually do But it turns out that we actually do know how to take care of that part. And know how to…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 52: 44 OD, right? So this is the just um say we OD, right? So this is the just um say we have this particular OD and it turns out have this particular OD and it turns out have this particular OD and it turns…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 53: 30 change for the probability means that the marginalss are all the same. So this the marginalss are all the same. So this the marginalss are all the same. So this just means that now if you have your ju…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 54: 36 >> That's a great question. Um so for >> That's a great question. Um so for example, I don't know if you guys example, I don't know if you guys example, I don't know if you guys familiar with the RL. …
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 55: 40 Use the use the corresponding PF OD of Use the use the corresponding PF OD of the SDE. Right? So you just plug it in the SDE. Right? So you just plug it in the SDE. Right? So you just plug it in and t…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 57: 07 >> A lot of uh uh lot of PD getting converted to OD. Is it are we just doing converted to OD. Is it are we just doing converted to OD. Is it are we just doing it because it's simpler or do we you it b…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 57: 53 diversity but but yes so like essentially um if you're sampling from a essentially um if you're sampling from a essentially um if you're sampling from a SDE then it's going to be yeah because SDE then…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 59: 06 sampling chain through your by by by sampling chain through your by by by backward integrating your velocity and backward integrating your velocity and backward integrating your velocity and then duri…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 60: 12 you know just like you know you just add something divide something you it's the something divide something you it's the something divide something you it's the same thing. Okay. same thing. Okay. sam…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 61: 01 next week we're gonna be able to see like like like uh different options to further uh different options to further uh different options to further improve what we have learned. Um so for improve what…
cmu-10799-diffusion-flow L5: CMU 10799 S26: Lecture 5 - Flow Matching - Diffusion & Flow Matching 61: 41 the generation faster? And we're going to learn uh to do it with or without to learn uh to do it with or without to learn uh to do it with or without training retraining of the model. And training ret…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 00: 04 Hello everyone and welcome to the next Hello everyone and welcome to the next lecture of the course principles of lecture of the course principles of lecture of the course principles of diffusion mode…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 01: 27 Similarly, every single point in the box is following a specific trajectory is following a specific trajectory is following a specific trajectory and the particles are moving as and the particles are …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 02: 49 with not just space but it is changing with not just space but it is changing with time also. [snorts] Why it is with time also. [snorts] Why it is with time also. [snorts] Why it is changing with spa…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 03: 44 Now we approach another person. Let's Now we approach another person. Let's call that person uh detective number call that person uh detective number call that person uh detective number two. And we a…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 04: 36 from trajectories and then calculating from trajectories and then calculating the velocity field. [snorts] Now they the velocity field. [snorts] Now they the velocity field. [snorts] Now they are thin…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 06: 04 which is a specific point in space. which is a specific point in space. [snorts] Let's say this is x0. Now what is x1? X1 is equal to Now what is x1? X1 is equal to X0 + delta X X0 + delta X X0 + delt…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 07: 11 particles. We have looked at their particles. We have looked at their trajectories. We have looked at the trajectories. We have looked at the trajectories. We have looked at the mysterious force that …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 08: 08 their trajectories and you collect all these trajectories and you collect all these trajectories and you collect all these trajectories and you call them as a flow map or just and you call them as a f…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 09: 01 can get a fair idea about what the flow can get a fair idea about what the flow is representing is representing is representing by by by having this square grid and having this square grid and having …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 09: 50 flow models in the first place. So far whatever we have discussed can be whatever we have discussed can be whatever we have discussed can be understood by a physics student easily understood by a phys…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 10: 47 okay generate a new image of a cat. Now the way you solve this problem is Now the way you solve this problem is Now the way you solve this problem is you try to find a distribution you try to find a d…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 11: 50 a gshian distribution that is noise and you end up with the actual distribution. you end up with the actual distribution. you end up with the actual distribution. So the path that we are going to foll…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 12: 55 in a way which is similar to how the in a way which is similar to how the particles are moving in space and time. particles are moving in space and time. particles are moving in space and time. What d…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 14: 00 So we are casting the deep generative modeling problem modeling problem modeling problem as a problem where we are trying to as a problem where we are trying to as a problem where we are trying to pre…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 15: 05 [snorts] and it depends on time as well. So this is exactly the same equation So this is exactly the same equation So this is exactly the same equation that we looked at before. that we looked at befo…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 15: 54 to predict this. So this is approximated by a neural network. by a neural network. by a neural network. Okay. And these are the conditions we Okay. And these are the conditions we Okay. And these are …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 16: 56 distribution that we want to predict in the first place. the first place. the first place. And it turns out that this is at the end And it turns out that this is at the end And it turns out that this …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 17: 52 But what if you start with a gshian But what if you start with a gshian distribution and you end up with distribution and you end up with distribution and you end up with something that does not repre…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 18: 45 is the price of the house. In this case, is the price of the house. In this case, we want to minimize the loss function so we want to minimize the loss function so we want to minimize the loss functio…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 19: 42 this target is not very easy. In fact we have absolutely no idea what this target have absolutely no idea what this target have absolutely no idea what this target is. is. is. And that is going to be …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 20: 46 uh how the training target is predicted uh how the training target is predicted in the first place. in the first place. in the first place. And whatever I'm going to discuss next And whatever I'm goin…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 21: 51 breed. So we have different So we have different So we have different we have a probability distribution that we have a probability distribution that we have a probability distribution that you know t…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 22: 46 going to follow next and uh we are going to go through a little bit of to go through a little bit of to go through a little bit of mathematics right now so please don't be mathematics right now so ple…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 23: 38 Okay. So, first concept that we will Okay. So, first concept that we will discuss is called as So let's let's try to understand what is So let's let's try to understand what is the meaning of [snorts]…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 24: 51 to something I do know even if it is like just one simple single data point like just one simple single data point like just one simple single data point [snorts] and that is exactly what [snorts] and…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 25: 51 know the data which is given to us. We know this initial distribution also. know this initial distribution also. know this initial distribution also. So it is very rational to think that we So it is v…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 27: 05 go to the point Z. So it is conditioned on the point Z. on the point Z. So what can I do so that each individual So what can I do so that each individual So what can I do so that each individual distr…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 28: 46 such that at every time step I will have such that at every time step I will have a gshian and a gshian and a gshian and the probability distribution will change the probability distribution will chan…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 29: 39 probability path which is linear in probability path which is linear in time. time. time. So the mean increases from zero to the So the mean increases from zero to the So the mean increases from zero …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 30: 39 going to generate those trajectories. uh and this is exactly what we are going uh and this is exactly what we are going uh and this is exactly what we are going to do next. But before we move, I want …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 31: 23 anyways uh for now just think about this as being a conditional probability path as being a conditional probability path as being a conditional probability path which is conditioned on the point Z whi…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 32: 15 single point that is the probability path and we want to know at every single path and we want to know at every single path and we want to know at every single time step in this transition how does ti…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 33: 17 trajectories the reverse problem trajectories the reverse problem we simply sample the initial point which we simply sample the initial point which we simply sample the initial point which was x0 in o…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 34: 06 this transformation? There is a vector field which is responsible for this field which is responsible for this field which is responsible for this transformation transformation and [snorts] and [snort…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 35: 06 every single point the velocity vector is pointed towards them. And And for points which are far away the for points which are far away the for points which are far away the vectors are longer in magn…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 36: 05 this is because as as as time proceeds this is because as as as time proceeds we probably want to uh move the velocity we probably want to uh move the velocity we probably want to uh move the velocity…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 36: 56 we don't need anything else using this formula we can simply calculate the formula we can simply calculate the formula we can simply calculate the trajectories of each and every point so trajectories …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 37: 46 Now this u of t is is called as Now this u of t is is called as conditional So you you might look at this and think So you you might look at this and think uh well if you look at the training uh well …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 38: 40 our target and train our neural network to predict it for every single point to predict it for every single point to predict it for every single point zed. zed. zed. It turns out that what we will do …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 39: 37 velocity vector field which is pointed from our data to the data point where we from our data to the data point where we from our data to the data point where we want to go. So it's it's a it's arrow …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 40: 32 final flow that you get the the final trajectories that you get those are the trajectories that you get those are the trajectories that you get those are the marginal probability paths. marginal proba…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 41: 24 to us. So here if you see to us. So here if you see the final distribution that we get which the final distribution that we get which the final distribution that we get which is which are these um is …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 42: 50 want to reach which is a sample of eight washians washians washians and initially we start with a random and initially we start with a random and initially we start with a random noise. noise. noise. …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 43: 45 Now uh here let's play this piece of code to understand this piece of code to understand this piece of code to understand this u and you'll probably get something like and you'll probably get somethin…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 44: 39 in our target. But uh the the issue is that we do not have access to this that we do not have access to this that we do not have access to this because we only have access to our because we only have …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 45: 39 by z minus x divided by 1 minus t. by z minus x divided by 1 minus t. [snorts] So these these these field [snorts] So these these these field [snorts] So these these these field lines are almost linea…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 46: 30 we do not have access to infinite samples. we have access to finite samples. we have access to finite samples. we have access to finite samples samples samples and uh this this uh credit goes to uh an…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 47: 18 understand what we are doing here. Essentially we are treating data Essentially we are treating data Essentially we are treating data generation as a flow problem. We know generation as a flow problem…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 48: 17 field is impossibly complex. You cannot write down an equation that You cannot write down an equation that You cannot write down an equation that defines what a Giblly movie looks like. defines what a…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 49: 06 be true and it does actually work. We we be true and it does actually work. We we have seen this in other uh aspects of have seen this in other uh aspects of have seen this in other uh aspects of deep…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 49: 54 uh I I hope this lecture was interesting for all of you. I haven't yet given the for all of you. I haven't yet given the for all of you. I haven't yet given the final reveal. Uh I'm I'm delaying it fi…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 50: 45 this this came out last year this this came out last year and this is based on uh flow matching. and this is based on uh flow matching. and this is based on uh flow matching. So here essentially what …
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 51: 32 is which is the action expert here which is a it's based on flow matching and uh is a it's based on flow matching and uh is a it's based on flow matching and uh the training target that you see here i…
diffusion-principles-vizuara L1: Lecture 10 - Constructing Training Target for Flow Models | Principles of Diffusion Models 52: 09 lecture which is going to be like a magician's reveal. Thank you.