---
title: Iterative Proportional Flow-Matching (IPFM)
url: https://www.emergentmind.com/topics/iterative-proportional-flow-matching-ipfm
type: topic
---

# Iterative Proportional Flow-Matching (IPFM)

Searching arXiv for the cited and related papers to ground the article.
Iterative Proportional Flow-Matching (IPFM) denotes an iterative class of generative transport procedures in which a learned flow is repeatedly corrected toward a target distribution. In the current literature, the term has two related but distinct usages. In "Iterative Flow Matching -- Path Correction and Gradual Refinement for Enhanced Generative Modeling" [2502.16445], the term is an alignment of terminology for an iterative flow matching framework built from end-path correction and gradual refinement; the paper explicitly states that it did not use the specific term "Iterative Proportional Flow-Matching." In "Incorporating Pre-trained Diffusion Models in Solving the Schrödinger Bridge Problem" [2508.18095], IPFM is an explicit reparameterization for Schrödinger bridge training that matches terminus-directed flow targets scaled proportionally by the remaining schedule. In both settings, the central mechanism is iterative re-anchoring of transport dynamics to reduce mismatch between the distributions seen during training and those induced by actual integration.

## 1. Terminology and scope

The term IPFM is not yet fully standardized. One line of work uses it as a convenient label for an iterative enhancement of continuous-time flow matching, while another introduces it as a discrete-time Schrödinger bridge reparameterization within a unified SB–SGM framework [2502.16445] [2508.18095]. This suggests that IPFM should presently be understood as a family of closely related iterative correction schemes rather than a single canonical algorithm.

| Usage of IPFM | Problem setting | Meaning of “proportional” |
|---|---|---|
| Iterative flow matching | Image generative modeling via probability flow ODEs | Time-rescaling factor $\frac{1}{1-t_j}$ in gradual refinement; optional discrepancy-based loss reweighting was not used in the reported experiments |
| SB reparameterization | Discrete-time Schrödinger bridge training with Gaussian conditionals | Schedule-dependent factors $\frac{1}{\bar{\gamma}_{k+1}}$ and $\frac{1}{1-\bar{\gamma}_k}$ in terminus-directed flow targets |

The first formulation is organized around correcting the pushforward distribution produced by a flow-matching model. The second is organized around alternating forward and backward bridge training, with flow targets derived from endpoint-conditioned Gaussian bridge identities. The overlap is structural: both repeatedly replace idealized training paths with targets anchored to actually generated states.

## 2. Mathematical setting and transport viewpoint

In the iterative flow-matching formulation, a base distribution $\pi_0$ is transported to a target distribution $\pi_T$ by learning a velocity field $\mathbf{v}_\theta(\mathbf{x},t)$ in the probability flow ODE
$$
\frac{d\mathbf{x}_t}{dt} = \mathbf{v}_\theta(\mathbf{x}_t, t),\quad \mathbf{x}_0\sim \pi_0,
$$
with induced density path $\{\rho_t\}$ satisfying the continuity equation
$$
\partial_t \rho_t(\mathbf{x}) + \nabla \cdot \big(\rho_t(\mathbf{x})\,\mathbf{v}_t(\mathbf{x})\big) = 0.
$$
The default path parameterization is the linear homotopy
$$
\mathbf{x}_t = t\,\mathbf{x}_T + (1-t)\,\mathbf{x}_0,
$$
with target velocity
$$
\mathbf{v}^*(\mathbf{x}_t, t) = \mathbf{x}_T - \mathbf{x}_0.
$$
The standard flow-matching loss is least-squares regression on homotopy samples:
$$
\mathcal{L}_{\text{FM}}(\theta) = \mathbb{E}_{t,\mathbf{x}_0,\mathbf{x}_T}\big[\|\mathbf{v}_\theta(\mathbf{x}_t, t) - \mathbf{v}^*(\mathbf{x}_t, t)\|^2\big].
$$
In low-dimensional latent settings, the same paper also instantiates $\mathbf{v}_\theta$ as an RBF interpolant with Tikhonov regularization [2502.16445].

In the Schrödinger bridge formulation, the starting point is the path-space KL problem
$$
\pi^* = \arg\min_{\pi \in \mathscr{P}_{N+1}} \Big\{ \mathrm{KL}(\pi \,\|\, p_{\mathrm{ref}}) : \pi_0 = p_{\mathrm{data}},\;\pi_N = p_{\mathrm{prior}} \Big\},
$$
where the reference path measure factorizes as a discrete forward Markov chain with Gaussian transitions
$$
p_{k+1|k}(x_{k+1}\mid x_k) = \mathcal{N}\!\big(x_{k+1};\,x_k+\gamma_{k+1} f_k(x_k),\,2\gamma_{k+1} I\big).
$$
That paper uses the continuity equation only as intuition for “flow” learning; the implemented objectives are discrete-time and are tied to Gaussian bridge updates rather than an explicit PDE formulation [2508.18095].

Both lines of work connect IPFM to transport. The iterative flow-matching paper relates the continuity equation and kinetic energy integral to the Benamou–Brenier formulation of $W_2$, while the Schrödinger bridge paper places IPFM inside entropic optimal transport over path measures [2502.16445] [2508.18095].

## 3. Iterative correction in flow matching

A central claim of the iterative flow-matching formulation is that hallucinations arise because training and inference follow different intermediate distributions. Training samples lie on straight homotopy lines between paired $(\mathbf{x}_0,\mathbf{x}_T)$, but inference integrates the learned field globally. As a result, the ODE can produce bent trajectories whose intermediate states follow $\widehat{\pi}_t \neq \pi_t$, generating covariate shift: the model is accurate on $\pi_t$ but must generalize off-path to $\widehat{\pi}_t$. Endpoint deviations $\widehat{\pi}_1 \neq \pi_T$ then produce out-of-distribution samples or “hallucinations” [2502.16445].

This mismatch can be written through the trajectory residual
$$
\mathbf{r}(t) := \mathbf{x}_t^{\text{ode}} - \mathbf{x}_t^{\text{lin}},
$$
with dynamics
$$
\dot{\mathbf{r}}(t) = \mathbf{v}_\theta(\mathbf{x}_t^{\text{ode}}, t) - \mathbf{v}^*(\mathbf{x}_t^{\text{lin}}, t).
$$
The paper argues that this residual accumulates interpolation error and off-path extrapolation, especially when homotopy lines intersect or nearly intersect near $t=0,1$ [2502.16445].

Two correction mechanisms are then defined.

**End-path correction** retrains a flow from the current pushforward distribution to the target. Given $\mathbf{x}_k \sim \widehat{\pi}_k$, the iteration trains
$$
\mathcal{L}^{(k)}(\theta) = \mathbb{E}_{t,\mathbf{x}_k,\mathbf{x}_T}\big[\|\mathbf{v}^{(k)}_\theta(\mathbf{x}_t^{(k)}, t) - (\mathbf{x}_T-\mathbf{x}_k)\|^2\big],
$$
where
$$
\mathbf{x}_t^{(k)} = t\,\mathbf{x}_T + (1-t)\,\mathbf{x}_k,
$$
and then integrates
$$
\frac{d\mathbf{x}_t}{dt} = \mathbf{v}^{(k)}_\theta(\mathbf{x}_t, t),\quad \mathbf{x}(0)=\mathbf{x}_k,\quad \mathbf{x}_{k+1} = \mathbf{x}(1).
$$
The resulting fixed-point viewpoint is
$$
\widehat{\pi}_{k+1} = \mathsf{FM}(\widehat{\pi}_k,\pi_T),
$$
with ideal contractive behavior
$$
\mathcal{D}(\widehat{\pi}_{k+1},\pi_T) \le \gamma\,\mathcal{D}(\widehat{\pi}_k,\pi_T),\quad 0<\gamma<1.
$$

**Gradual refinement** partitions time into checkpoints $0=t_0<t_1<\dots<t_n=1$ and retrains on short corrected segments. On segment $j$, the homotopy is re-anchored at the actual integrated state:
$$
\mathbf{x}_t^{(j)} = \frac{t-t_j}{1-t_j}\,\mathbf{x}_T + \Big(1-\frac{t-t_j}{1-t_j}\Big)\,\widehat{\mathbf{x}(t_j)},
$$
with scaled target velocity
$$
\mathbf{v}^{*,(j)}(\mathbf{x}_t^{(j)}, t) = \frac{1}{1-t_j}\big(\mathbf{x}_T - \widehat{\mathbf{x}(t_j)}\big).
$$
Its objective is
$$
\mathcal{L}^{(j)}(\theta) = \mathbb{E}_{t\in[t_j,t_{j+1}]}\big[\|\mathbf{v}_\theta(\mathbf{x}_t^{(j)}, t) - \mathbf{v}^{*,(j)}(\mathbf{x}_t^{(j)}, t)\|^2\big].
$$
Here the term “proportional” refers explicitly to the factor $\frac{1}{1-t_j}$, which rescales the segment velocity by the remaining time to the terminal point. The paper also allows an optional reweighting
$$
w_k(t,\mathbf{x}) \propto \psi\big(\mathcal{D}(\widehat{\pi}_k,\pi_T)\big),
$$
but states that such adaptive weights were not used in the reported experiments [2502.16445].

The theoretical motivation combines three ingredients: a contractivity assumption for the flow-matching update, Talagrand-based control of $W_2$ by KL when $\pi_0$ is log-concave, and an RBF approximation estimate
$$
N = \mathcal{O}\Big(\big(\tfrac{c}{\epsilon}\big)^{d/s}\Big),
$$
showing that smaller transport velocity norms lower approximation complexity. The intended consequence is that iterative correction both shrinks the required velocity magnitude and aligns training with the actual marginals encountered during integration [2502.16445].

## 4. IPFM within Schrödinger bridge training

In the Schrödinger bridge paper, IPFM is defined directly as an iterative reparameterization of the bridge objectives. The learned quantities are discrete-time flow predictors pointing from an intermediate state toward the relevant terminus, scaled by the remaining fraction of the schedule. Specifically, at time index $k$ the targets are
$$
\frac{x_0 - x_{k+1}}{\bar{\gamma}_{k+1}}
\quad \text{and} \quad
\frac{x_N - x_k}{1 - \bar{\gamma}_k},
$$
where
$$
\bar{\gamma}_k = \sum_{i=0}^k \gamma_i.
$$
The corresponding losses are
$$
\mathcal{L}_{\tilde{b}_{k+1}^n}
=
\mathbb{E}_{\,p_{0,k+1}^n}\!\left[\left\|\tilde{b}_{k+1}^n(x_{k+1}) - \frac{x_0 - x_{k+1}}{\bar{\gamma}_{k+1}}\right\|^2\right],
$$
and
$$
\mathcal{L}_{\tilde{f}_{k}^{n+1}}
=
\mathbb{E}_{\,q_{k,N}^n}\!\left[\left\|\tilde{f}_{k}^{n+1}(x_k) - \frac{x_N - x_k}{1 - \bar{\gamma}_k}\right\|^2\right].
$$
These are the “proportional” targets of this formulation [2508.18095].

IPFM is related to two companion reparameterizations introduced in the same paper. IPMM predicts next-step means directly, and IPTM predicts the termini themselves. IPFM connects to IPMM by
$$
B_{k+1}^n(x_{k+1}) = x_{k+1} + \gamma_{k+1}\, \tilde{b}_{k+1}^n(x_{k+1}),
$$
$$
F_k^{n+1}(x_k) = x_k + \gamma_{k+1}\, \tilde{f}_k^{n+1}(x_k).
$$
Thus, the IPFM velocity multiplied by the local step size yields the mean update for the Gaussian bridge conditional [2508.18095].

The paper’s alternating scheme uses current backward trajectories to train the forward model and current forward trajectories to train the backward model. Proposition 1 states that the IPMM mean-matching losses are approximately equivalent to the original DSB losses under Gaussian conditional models. Proposition 2 derives proportional Gaussian bridge means of the form
$$
\mu_{k+1}^n(x_{k+1},x_0) \approx x_{k+1} + \frac{\gamma_{k+1}}{\bar{\gamma}_{k+1}}(x_0 - x_{k+1}),
$$
and
$$
\tilde{\mu}_k^n(x_k',x_N') \approx x_k' + \frac{\gamma_{k+1}}{1-\bar{\gamma}_k}(x_N' - x_k'),
$$
which justify the IPTM and IPFM targets [2508.18095].

Conceptually, this version of IPFM is the one closest to a bridge between SB and continuous-time flow matching. The paper explicitly describes it as “flow-matching-like” within the SB framework, but still discrete-time, alternating, and tied to Gaussian bridge transitions rather than standard rectified-flow or continuous FM training [2508.18095].

## 5. Initialization, algorithms, and implementation regimes

The two IPFM lineages differ sharply in how they are instantiated.

In iterative flow matching, the algorithmic structure is stagewise correction. End-path correction initializes with samples $\{\mathbf{x}_0^{(i)}\}\sim\pi_0$, trains a standard FM model, integrates to obtain $\{\mathbf{x}_{1}^{(i)}\}$, computes a discrepancy $\varepsilon_k=\mathcal{D}(\{\mathbf{x}_{k+1}^{(i)}\},\{\mathbf{x}_T^{(i)}\})$, and repeats until $\varepsilon_k \le \text{tol}$. Gradual refinement instead iterates over temporal checkpoints, training and integrating on each corrected segment. The paper recommends choosing a discrepancy metric aligned with the data space, such as FID/IS in image space or ICP/transport-cost in latent space, and stopping when the discrepancy plateaus near the target’s internal self-similarity level [2502.16445].

That paper reports two main implementation regimes. One uses neural $\mathbf{v}_\theta$ in pixel or latent space for high-dimensional settings. The other uses an RBF interpolant in moderate-dimensional latent spaces with Gaussian kernel smoothing and ridge parameter $\beta$. For images, the reported settings are MNIST with an autoencoder to 32-D latent space and CIFAR-10 with an autoencoder to 64-D latent space. End-path correction uses the linear schedule $s(t)=t$, while gradual refinement uses uniform partitions such as $n=6$ equal segments [2502.16445].

In the Schrödinger bridge formulation, the algorithm is an alternating forward/backward training loop over $2L$ epochs and $N$ timesteps with noise schedule $\{\gamma_k\}$ satisfying $\sum_k \gamma_k = 1$. A key practical contribution is plug-and-play initialization from pre-trained SGMs. For a pre-trained flow-matching-style model $m_{\theta_1}$, the backward bridge mean predictor is initialized as
$$
B_{\beta^0}(k,x) := x + \frac{1}{N}\, m_{\theta_1}(k,x).
$$
For high-resolution bidirectional tasks, the paper also uses dual initialization
$$
B_{\beta^1}(k,x) := x + \frac{1}{N}\, m_{\theta_1}(k,x),\qquad
F_{\alpha^1}(k,x) := x + \frac{1}{N}\, m_{\theta_2}(N-k,x).
$$
The stated rationale is that the IPMM, IPTM, and IPFM targets align with common SGM targets, whereas naïve initialization of the original DSB objective suffers from time-state misalignment [2508.18095].

The reported hyperparameters for the SB experiments are Adam with learning rate $1\mathrm{e}{-4}$, $\beta=(0.9, 0.99)$, no weight decay, and batch size $256$. The paper uses $50$ timesteps for CelebA $64\times64$, a symmetric $\gamma$ schedule that grows then decays, and ADM or LDM backbones depending on the image setting [2508.18095].

## 6. Empirical behavior, limitations, and current interpretation

The iterative flow-matching paper reports that, in a Gaussian mixture example, after $3$ end-path corrections the transport-cost metric dropped by roughly two orders of magnitude. For MNIST, FID decreased monotonically with iteration in latent space, approaching the dataset’s internal self-similarity. For CIFAR-10, the similarity metric in latent space decreased across iterations, with qualitative sample quality improving. The same paper states that end-path correction consistently improved fidelity and reduced out-of-distribution artifacts, while gradual refinement achieved similar endpoints but was less robust near $t=0$; it attributes this sensitivity to the first segment, where many homotopy lines intersect and the target velocity points toward a center-of-mass direction [2502.16445].

The Schrödinger bridge paper reports stronger image-quality numbers for its own IPFM formulation. On Gaussian bridges, averaged $\mathrm{KL}(\pi_t \,\|\, \pi_t^{\mathrm{SB}})$ across times shows that from scratch IPFM, IPTM, and IPMM outperform DSB and several other baselines; for example, at $d=50$, IPFM is reported as approximately $9.93$ versus approximately $32.91$ for DSB, and with SGM initialization improves further to approximately $7.97$. On unconditional CelebA $64\times64$, IPFM achieves FID approximately $14.63$, compared with approximately $23.29$ for SGM. On AFHQ $512\times512$ cat$\to$dog translation, IPFM achieves FID approximately $9.96$, compared with approximately $15.46$ for the pre-trained SGM and approximately $20$–$55$ for SB baselines. The paper also states that IPFM consistently improves FID on Horse$\leftrightarrow$Zebra and Selfie$\leftrightarrow$Anime tasks [2508.18095].

Several limitations recur across both formulations. In iterative flow matching, computational overhead scales roughly linearly with the number of IPFM iterations $K$, and sampling cost scales with the number of stages and ODE steps. Extremely high-dimensional raw spaces can challenge velocity interpolation, especially for RBFs, and the paper recommends latent-space training for images [2502.16445]. In the Schrödinger bridge setting, convergence depends strongly on initialization quality; poor initialization yields inaccurate trajectories and slow convergence, and the paper explicitly notes that its analysis provides equivalence propositions and empirical evidence rather than full continuous-time convergence proofs [2508.18095].

A common misconception is to treat IPFM as a single fully settled method. The current literature does not support that reading. One formulation is an iterative correction scheme for standard flow matching, and the other is a proportional reparameterization of discrete Schrödinger bridge training. Another misconception is that “proportional” always means adaptive discrepancy weighting. In the iterative flow-matching paper, the reported experiments did not use adaptive weighting and used “proportional” only through the time-rescaling factor in gradual refinement. In the Schrödinger bridge paper, “proportional” refers to schedule-normalized terminus-directed targets [2502.16445] [2508.18095].

Taken together, these works define IPFM as a technically coherent but terminologically bifurcated research direction. Its shared principle is iterative correction of transport targets using the actually induced distributions or bridge trajectories, with explicit proportional scaling that makes the learned direction field compatible with the remaining time or noise schedule.

Source: https://www.emergentmind.com/topics/iterative-proportional-flow-matching-ipfm