Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models

TL;DR

Equilibrium Matching (EqM) replaces Flow Matching’s time-conditional velocity field with a single time-invariant gradient field that is compatible with an underlying (implicit) energy landscape. Ground-truth samples are stationary points of that landscape — the gradient vanishes on the data manifold and grows away from it — so sampling becomes gradient descent on the learned energy, with adjustable step sizes, adaptive optimizers, and adaptive compute, instead of integration along a prescribed time horizon. EqM reports faster convergence than Flow Matching at matched compute and an FID of 1.90 on ImageNet-256, and naturally handles partially-noised denoising, OOD detection, and image composition as side-effects of the energy view.

Key claims

EqM learns the equilibrium gradient of an implicit energy landscape rather than a non-equilibrium time-conditional velocity, with the target gradient defined as zero on ground-truth samples and pointing from noise to data on intermediate corrupted samples [Project page §Method].
The training objective is a regression on this target gradient at samples drawn from a noise→data corruption scheme, with no time embedding fed to the network [Project page §Training].
Sampling is reformulated as gradient descent on the learned landscape and supports adjustable step sizes, adaptive optimizers (e.g. Adam-style), and adaptive compute — decoupling sample quality from a fixed integration horizon [Project page §Sampling; Abstract].
Under identical step sizes and step counts on a 2-mode toy distribution, EqM converges much faster than Flow Matching, whose time-conditional velocity is only guaranteed to converge to ground truth at the final timestep $t=1$ [Project page §Conceptual visualization].
EqM surpasses diffusion / flow models empirically, reaching FID 1.90 on ImageNet-256×256 [Abstract; arXiv:2510.02300].
The framework is theoretically justified to learn and sample from the data manifold [Abstract].
Because the landscape is energy-shaped, EqM handles tasks that are awkward for flow models out-of-the-box: partially noised image denoising, OOD detection (energy as anomaly score), and image composition (sum of energies) [Abstract].

Method

EqM defines a corruption scheme that linearly transitions between data and noise (analogous to a Flow Matching schedule) and then trains a single time-invariant network $g_\theta(x)$ to regress a target gradient $\nabla_x \mathcal{E}(x)$ that is zero at ground-truth samples and points from noise back toward data on intermediate samples. Because $g_\theta$ has no time input, the learned field is the gradient of an implicit scalar energy: the network parameterizes $\nabla_x \mathcal{E}$ directly rather than $\mathcal{E}$ itself.

At inference, samples are drawn by gradient descent on this landscape from a noise initialization, $x_{k+1} = x_k - \eta_k g_\theta(x_k)$ . Because the gradient field is autonomous, the step size $\eta_k$ and the optimizer choice are free parameters at inference rather than fixed by the training-time integration grid — Adam-style momentum, line search, or adaptive compute budgets all apply directly. Compared against Flow Matching, EqM substitutes a time-varying velocity ( $v_\theta(x, t)$ , only zero at $t=1$ near data) for a time-invariant gradient ( $g_\theta(x)$ , zero on the data manifold) — the conceptual 2D dynamics show EqM trajectories converging to ground-truth modes under any sufficiently small step size, where FM trajectories only converge at the prescribed final time.

Results

The headline number is FID 1.90 on ImageNet-256×256, claimed to surpass diffusion and flow baselines at comparable scale [Abstract]. The toy 2D-two-mode comparison shows EqM converging in markedly fewer steps than Flow Matching at identical step size and step count, with FM trajectories ending short of the ground-truth modes at intermediate horizons [Project page]. The paper also reports applications to partially-noised denoising, OOD detection, and image composition as native byproducts of the energy view, but the quantitative details on those tasks are in the arXiv PDF rather than the project page.

Why it’s interesting

EqM lands directly in two threads the wiki has been tracking. First, the parameterization-stability question raised by The Geometry of Noise: Why Diffusion Models Don't Need Noise Conditioning: Sahraee-Ardakan et al. prove that velocity-style parameterizations are the structurally stable family when noise conditioning is removed, and they cite EqM as one of the velocity-based blind generators their analysis explains. EqM is the constructive counterpart to that theory paper — it gives the energy landscape the theory invokes (the Marginal Energy) a concrete training objective and sampling rule. Second, EqM is the closest cousin yet to Generative Modeling via Drifting (Drifting Models, Kaiming He’s group), which also reframes generation around a static field whose equilibria are at the data — EqM makes the equilibrium-at-data property explicit via an energy formulation, while Drifting Models reach a similar regime through a different construction. The “adjustable step size / optimizer at inference” property also overlaps with the few-step generation thread (SoFlow, MeanFlow): EqM gets adaptive compute essentially for free because the field is autonomous.

The training-efficiency claim (“much faster convergence than Flow Matching” on the project page) is consistent with the Diffusion training efficiency page’s running tally of alternatives to the standard “DiT + flow matching + VAE latents” recipe — but the comparison points are project-page toy figures and an ImageNet FID rather than a head-to-head against PixelGen / RAE / Self-Flow at matched compute, so the headline number should be read in that context.