Probability Flow ODE

A probability flow ODE is a deterministic trajectory whose marginal distributions match those of a corresponding diffusion SDE. It is useful because a score-based model can be sampled by integrating an ODE rather than simulating stochastic reverse diffusion.

For a forward SDE:

$$dx=f(x,t)dt+g(t)dw$$

the probability flow ODE is:

$$\frac{dx}{dt}=f(x,t)-\frac{1}{2}g(t{)}^2{\nabla }_x\log p_t(x)$$

$f$ is the drift, $g$ is the diffusion coefficient, $w$ is Brownian motion, and ${\nabla }_x\log p_t(x)$ is the score at time $t$.

The deterministic velocity is:

$$v(x,t)=f(x,t)-\frac{1}{2}g(t{)}^2{s}_{\theta }(x,t)$$

where $s_{\theta }(x,t)\approx {\nabla }_x\log p_t(x)$. This makes probability flow ODEs a bridge between score models and flow-like deterministic transport.

Why It Matters

• Turns stochastic diffusion sampling into deterministic ODE integration.
• Connects score-based diffusion to continuous normalizing flow views.
• Enables likelihood computation through instantaneous change of variables when the divergence is tractable or estimated.
• Makes solver choice part of the sampling contract.

Change of Variables

For an ODE:

$$\frac{dx}{dt}=v_{\theta }(x,t)$$

the log density evolves as:

$$\frac{d}{dt}\log p_t(x_t)=-{\nabla }_x\cdot v_{\theta }(x_t,t)$$

where ${\nabla }_x\cdot v$ is the divergence of the velocity field. Integrating this term gives a likelihood route:

$$\log p_0(x_0)=\log p_T(x_T)+{\int }_0^T{\nabla }_x\cdot v_{\theta }(x_t,t)dt$$

depending on time direction convention. The sign should be checked in each paper’s notation.

Sampling Contract

Sampling starts from noise $x_T\sim p_T$ and integrates backward to $x_0$:

$$x_0=\mathrm{ODESolve}\left(x_T,v_{\theta }(x,t),T\to 0\right)$$

where $v_{\theta }$ is the learned or score-derived velocity field.

The ODE can also be read as deterministic transport:

$$\frac{dx}{dt}=v_{\theta }(x,t)$$

where $v_{\theta }$ may be parameterized directly, as in Flow matching, or derived from a score field:

$$v_{\theta }(x,t)=f(x,t)-\frac{1}{2}g(t{)}^2{s}_{\theta }(x,t)$$

This makes the learned object explicit: some papers train a score, others train a velocity, and others train a denoiser that is converted to one of these.

ODE Solver Boundary

The sampler is not only the learned model:

$$\text{sample}=\mathrm{ODESolve}(v_{\theta },x_T,\text{solver},\text{tolerance},\text{NFE})$$

where NFE is the number of function evaluations. Quality, speed, and diversity comparisons are weak unless solver settings are fixed or reported.

Solver Choice	Claim Affected
NFE / step count	speed-quality tradeoff
adaptive tolerance	wall time and numerical error
time discretization	endpoint detail and stability
stochastic vs deterministic	sample diversity
divergence estimator	likelihood accuracy

Checks

• Which ODE solver, tolerance, and number of function evaluations are used?
• Is the score accurate enough at the integration endpoints?
• Does deterministic sampling reduce diversity for the task?
• Are coordinates, graphs, or sequences integrated in a representation that preserves validity?
• Is the comparison fair against stochastic samplers with different compute budgets?
• Is the model trained as score prediction, noise prediction, denoising, or velocity prediction?
• Is likelihood estimation claimed, or only deterministic sampling?
• Is the sign convention for forward/backward integration stated?
• Is divergence computed exactly, approximated, or not used?

Related

• Score matching
• Score-based model
• Diffusion model
• Flow matching
• Normalizing flow
• Energy-based model
• Sampling
• Dynamical systems