Technical Summary: Midpoint Generative Models (MGM)

Problem Statement

Diffusion and flow models have become central to generative AI, enabling applications in image, video, and audio synthesis. However, a primary bottleneck for practical deployment is inference speed; standard diffusion models require hundreds or thousands of function evaluations (NFE) to generate high-quality samples. While several approaches exist to accelerate generation—such as developing better numerical solvers, distilling pre-trained models into one-step generators, or training one-step models from scratch using structural properties of diffusion—there remains a need for principled, effective methods to train one-step generators directly from data without relying on a pre-trained "teacher" model.

Methodology

The paper introduces Midpoint Generative Models (MGM), a framework for training one-step generative models based on a specific symmetry property of Flow Matching with linear interpolation.

Core Insight: Midpoint Symmetry

The authors observe that in Flow Matching, when interpolating between two identical endpoint distributions ($p_0 = p_1$) using a linear path $X_t = (1-t)X_0 + tX_1$, the velocity field $v_t(x)$ vanishes exactly at the midpoint time $t=1/2$. Specifically, $v_{1/2}(x) = \mathbb{E}[X_1 - X_0 | X_{1/2} = x] = 0$. Conversely, if $p_0 \neq p_1$, this field is generally non-zero. This property allows the norm of the velocity field at the midpoint to serve as a measure of discrepancy between distributions.

The Midpoint Divergence

To formalize this, the authors define the Midpoint Divergence:
$$D_{\text{mid}}(p_0, p_1) := \mathbb{E}_{X_{1/2}} \left[ \| \mathbb{E}_{X_0, X_1}[X_1 - X_0 | X_{1/2}] \|_2^2 \right]$$
They prove that this divergence is definite: it is non-negative and equals zero if and only if $p_0 = p_1$.

Extending Beyond the Midpoint

A naive extension of this concept to all time steps $t \in [0, 1]$ fails because the velocity field is non-zero at non-midpoint times even when $p_0 = p_1$, due to an inherent orientation bias (it is easier to reconstruct the "past" endpoint than the "future" one). To resolve this, MGM introduces randomly flipped interpolations:

1. A Bernoulli variable $B \sim \text{Bernoulli}(1/2)$ determines whether the interpolation proceeds forward ($X_t$) or backward ($X_{1-t}$).
2. The observation is the flipped variable $\tilde{X}_t$, but the flip $B$ is unobserved.
3. This symmetrization ensures that for any $t$, if $p_0 = p_1$, the conditional displacement field $\tilde{v}_t(x) = \mathbb{E}[X_1 - X_0 | \tilde{X}_t = x]$ vanishes.

This leads to the Time-integrated Midpoint Divergence, defined by integrating the squared norm of this symmetrized displacement field over $t \in [0, 1/2]$.

Generalization to Stochastic Interpolants

The framework is further generalized to symmetric stochastic interpolants of the form $X_t = I_t(X_0, X_1) + \sigma_t \epsilon$, where the interpolation function $I_t$ and noise schedule $\sigma_t$ satisfy symmetry conditions ($I_t(x_0, x_1) = I_{1-t}(x_1, x_0)$ and $\sigma_t = \sigma_{1-t}$). This yields a Generalized Midpoint Divergence that remains a valid discrepancy measure.

Variational Training Objective

Directly minimizing the divergence is intractable because the true conditional displacement field is unknown. The authors derive a variational formulation:
$$D(p_\theta, p_{\text{data}}) = \max_{f} \int_0^{1/2} \mathbb{E} \left[ 2\langle f(\tilde{X}_t), X_1 - X_0 \rangle - \|f(\tilde{X}_t)\|_2^2 \right] dt$$
where $f$ is a learnable vector-valued function (a "displacement field" or critic). This results in a minimax objective:

1. Update $f_\psi$ (Critic): Regress $f_\psi$ to approximate the symmetrized displacement field $X_1 - X_0$.
2. Update $G_\theta$ (Generator): Minimize the variational objective using the fixed $f_\psi$.

The generator $G_\theta$ maps a latent prior $z$ to $X_0$, which is then paired with real data $X_1$ to compute the loss.

Key Contributions

1. Midpoint Divergence: The introduction of a new distributional discrepancy derived from the midpoint symmetry of linear interpolant Flow Matching. The authors prove its definiteness (vanishing iff distributions are equal).
2. Symmetrization via Time Flip: The extension of the divergence to all time steps via random time-flipping, removing orientation bias and enabling time-integrated objectives.
3. Generalized Midpoint Divergence: The generalization of the concept to symmetric stochastic interpolants, providing a flexible framework for various noise schedules.
4. MGM Algorithm: A practical training framework for one-step generators that:
   • Trains directly from data without a pre-trained diffusion/flow teacher.
   • Uses a learned displacement field to approximate the divergence.
   • Achieves competitive performance against existing one-step methods.

Experimental Results

The authors evaluate MGM on unconditional CIFAR-10 (32x32) generation, reporting the Frechet Inception Distance (FID) for 1-NFE sampling.

• Performance: MGM achieves an FID of 2.27, which is the best result among methods that do not use a teacher-model training signal (i.e., methods trained from scratch without distilling a pre-trained model). It outperforms other teacher-free approaches like DiffRatio-DiJS (2.39) and iCT-deep (2.51).
• Ablation Studies:
  • Time Integration: Using only the midpoint ($t=1/2$) yields valid but noisier samples (FID ~13.55 with critic noise), whereas integrating over time significantly improves quality (FID 2.27).
  • Random Flipping: The naive unflipped time-integrated objective fails (FID > 100) due to orientation bias, confirming the necessity of the random flip.
  • Stochasticity: A small amount of noise in the interpolant ($\sigma > 0$) is crucial for stable training; deterministic bridges perform poorly.
  • Critic Stochasticity: Decoupling the noise levels for the generator and the critic (using noise only for the critic) yields the best results.

Significance and Claims

The paper claims that MGM offers an effective and theoretically grounded approach to generative modeling. Its significance lies in:

• Theoretical Foundation: It reframes symmetric stochastic interpolants not just as generative paths, but as a basis for comparing endpoint distributions via a definite divergence.
• Training Paradigm: It provides a principled way to train one-step generators directly from data, avoiding the complexity and potential limitations of distilling pre-trained models.
• Competitiveness: It demonstrates that a method based on this specific symmetry can achieve state-of-the-art results among teacher-free one-step generators, challenging the notion that distillation is strictly necessary for high-quality, fast generation.

The authors acknowledge limitations, noting that the minimax training can introduce optimization instability similar to GANs, and that performance depends on design choices like the noise schedule and interpolant type. They do not claim to surpass all teacher-based distillation methods (e.g., CTM achieves 1.87 FID) but position MGM as the leading alternative for training from scratch.