Point-Mass Schedule in Generative Modeling

• Point-mass schedules are interpolation methods that initiate from a point mass, redefining the coupling between noise and data in generative models.
• They modify traditional SDE/ODE dynamics by enforcing zero initial drift and leveraging statistically optimal diffusion, resulting in enhanced numerical stability.
• Retrofitting pretrained flow and diffusion models with these schedules reduces integration steps while maintaining high sample quality.

A point-mass schedule is a class of interpolation schedules in generative modeling via stochastic interpolants, where the conventional Gaussian base measure collapses to a point mass at the initial time, fundamentally altering the induced SDE/ODE dynamics and allowing for numerically efficient sampling. This schedule generalizes the way base and target distributions are interpolated, leading to improved sampling complexity and convergence properties, particularly under statistically optimal SDE sampling. Point-mass schedules enable retrofitting of pretrained flow and diffusion models to yield high-quality generative samples in fewer integration steps, often without retraining (Damsholt et al., 3 Feb 2026).

1. Formal Definition and Construction

In the stochastic interpolant framework, the coupling of a standard Gaussian $Z\sim N(0,I)$ and target data $X\sim \rho_X$ is achieved via a spatially linear interpolant $I_t = \alpha_t Z + \beta_t X$, $t\in[0,1]$, where the schedule $(\alpha, \beta)$ traditionally satisfies $\alpha_0 = 1$, $\beta_0=0$, $\alpha_1=0$, $\beta_1=1$, and monotonicity conditions $\dot\alpha_t < 0, \dot\beta_t > 0$. The point-mass schedule is characterized by relaxing the initial boundary: both $\alpha_0=0$ and $\beta_0=0$. More precisely ((Damsholt et al., 3 Feb 2026), Definition 4.2):

• $\alpha, \beta \in C^1([0,1])$ with $\alpha_0=\alpha_1=\beta_0=0$, $\beta_1=1$,
• $\alpha_t>0$, $\dot\beta_t>0$ for $t\in(0,1)$,
• $\beta_t=o(\alpha_t)$ as $t\to 0^+$, ensuring $I_0=0$ almost surely,
• $\alpha_t^2=O(\beta_t)$ as $t\to 0^+$ (regularity),
• $$u_0 \equiv \lim_{t\to 0^+} \frac{\beta_t}{\alpha_t+\beta_t} = 0$$ (finite derivative),
• $\frac{d}{dt}(\beta_t/\alpha_t)>0$ so $t\mapsto u_t$ is strictly increasing.

These properties enforce initial sampling from a point mass, with interpolation governed by $\beta = o(\alpha)$ so the initial evolution is dominated by drift.

2. SDE/ODE Induced by Schedule

For any schedule $(\alpha, \beta)$, the marginal law of $I_t$ is $\rho(t, x) = \mathrm{Law}(I_t)$. The following mean fields are defined:

• $\eta_Z(t, x) = \mathbb{E}[Z|I_t = x]$,
• $\eta_X(t, x) = \mathbb{E}[X|I_t = x]$,
• $s(t, x) = \nabla_x \log \rho(t, x)$ (score).

For any $C^1$ non-negative diffusion-scale $\epsilon_t$, the driven drift is

$$b^\epsilon(t, x) = \dot\alpha_t \eta_Z(t, x) + \dot\beta_t \eta_X(t, x) + \epsilon_t s(t, x).$$

The unique strong solution to

$$dX_t^\epsilon = b^\epsilon(t, X_t^\epsilon)dt + \sqrt{2\epsilon_t} dW_t, \quad X_0^\epsilon \sim N(0,I)$$

satisfies $\mathrm{Law}(X_t^\epsilon) = \mathrm{Law}(I_t)$ and $X_1^\epsilon \sim \rho_X$. The deterministic case $\epsilon \equiv 0$ gives the probability-flow ODE.

3. Statistical-Optimal Diffusion and Lazy Schedules

Closed-form identities ((Damsholt et al., 3 Feb 2026), Proposition 3.2) relate mean fields and drift:

• $\eta_Z(t, x) = -\alpha_t s(t, x)$,
• $\eta_X(t, x) = (x + \alpha_t^2 s(t, x))/\beta_t$,
• $$b^\epsilon(t, x) = (\epsilon_t^* + \epsilon_t)s(t, x) + (\dot\beta_t/\beta_t)x$$, with the statistically-optimal diffusion scale

$$\epsilon_t^* = \alpha_t^2\frac{\dot\beta_t}{\beta_t} - \alpha_t\dot\alpha_t.$$

Theorem 3.3 shows $\epsilon^*$ uniquely minimizes the path-space KL divergence between the true SDE and its plug-in approximation.

For $X\sim N(0, I)$, the lazy schedule is defined by requiring the drift to vanish identically:

• ODE lazy ($\epsilon \equiv 0$): $b(t,x)=0$ iff $\alpha_t^2 + \beta_t^2 = 1$ for all $t$ (variance-preserving schedule).
• SDE lazy ($\epsilon = \epsilon^*$): $b^*(t,x) = 0$ iff $\alpha_t^2 + \beta_t^2 = \beta_t$, which forces $\alpha_0=0$, $\beta_0=0$, i.e., a point-mass schedule.

It follows that $2\epsilon_t^* = \dot\beta_t$, and thus $$\int_0^t 2\epsilon_u^* du = \beta_t - \beta_0 = \beta_t$$.

4. Path-Wise Conversion to Point-Mass Schedule

Theorem 4.6 provides a path transform allowing translation between arbitrary schedules and the point-mass schedule. Define $c_t = \alpha_t + \beta_t$, $u_t = \beta_t / (\alpha_t + \beta_t)$. Given an SDE under $(\alpha, \beta, \epsilon)$, consider the linear schedule $(\bar{\alpha}_t=1-t,\, \bar{\beta}_t=t)$ and set

• $$\bar{\epsilon}_{u_t} = \alpha_t \epsilon_t / (\beta_t \epsilon_t^*)$$,
• rescale the Wiener process: $\bar{W}_u = \int_0^{t_u} \sqrt{\dot{u}_s} dW_s$.

The solution $\overline{X}_{u_t}$ to the rescaled linear schedule satisfies

$X_t^\epsilon = c_t\, \overline{X}_{u_t}$

for the original schedule. Therefore, any pretrained flow or diffusion model under any schedule may be converted to a point-mass schedule via successive schedule transformations.

5. Practical Sampling Algorithms

For the "equal-logit-time" point-mass schedule $u_t = t$ ((Damsholt et al., 3 Feb 2026), Example 5.5), explicit pseudocode is given for both ODE lazy and SDE lazy sampling, using a pretrained linear-flow velocity $v^{flow}(t, x) = \bar{b}(t, x)$:

Lazy ODE ($\epsilon\equiv 0$):

input: pretrained linear-flow velocity v̄(t,·), stepsize Δt initialize t=0, x∼N(0,I) loop n=0…N−1: d = (1−t)^2 + t^2 b = ((1−2t)/d)⋅x + (1/√d)·v̄(t,√d x) x ← x + Δt·b t ← t+Δt return x

Lazy SDE ($\epsilon = \epsilon^*$):

input: pretrained linear-flow velocity v̄(t,·), stepsize Δt initialize t=Δt, x = 0 (since point-mass initial X₀=0) loop n=1…N−1: d = (1−t)^2 + t^2 b* = (2/d)[(1−2t)x + t·v̄(t,(d/t)x)] ΔW ∼ N(0,(β_{t+Δt}−β_t)·I), where β_t=t^2/d x ← x + Δt b* + ΔW t ← t+Δt return x

This construction demonstrates that point-mass schedule sampling can leverage existing pretrained models and simple update rules for efficient sampling.

6. Theoretical Properties and Numerical Considerations

Point-mass schedules under statistically optimal SDE sampling exhibit several key properties:

• The initial state is a point mass, and dynamics have zero drift under the lazy SDE schedule.
• For deterministic (ODE) lazy sampling, the variance-preserving condition $\alpha^2 + \beta^2=1$ recovers the standard diffusion schedule.
• The path-space KL divergence is invariant to the schedule (Proposition 4.9), so numerical stability, rather than statistical optimality, drives optimal schedule selection.
• Bounded time-derivatives $\dot\alpha, \dot\beta$ improve numerical integration, especially near $t=0,1$, reducing required step counts and stabilization overhead.

A systematic implication is that, for Gaussian data and SDE lazy schedules, optimal sampling starts from a point mass and proceeds with minimal stochastic "stiffness," contributing to practical acceleration.

7. Empirical Performance and Applications

Experiments using a 1.3B-parameter PRX latent flow model for text-to-image generation ((Damsholt et al., 3 Feb 2026), Section 7) demonstrate that point-mass schedules enable meaningful reductions in integration steps:

• With ODE sampling, converting to the lazy-ODE schedule yields up to 3 fewer solver steps (of 64–172).
• With statistically optimal SDE sampling, the reduction is more substantial: 128 lazy-SDE steps achieve parity with approximately 171 linear-SDE steps (roughly 25% fewer steps).
• The predictor–corrector integrator is robust across a variety of prompts, with observed improvements in RMSE and convergence consistency.

This suggests that retrofitting existing generative models to point-mass schedules, especially under the lazy SDE construction, provides a principled avenue for accelerating sample generation with little or no loss in output quality.