Asymmetric Flow Modeling (AsymFlow)

• AsymFlow is a rank-asymmetric velocity parameterization that restricts noise estimation to a low-dimensional subspace while preserving full data predictions.
• It recovers the complete velocity analytically without modifying the underlying transformer architecture or sampling procedures.
• Experimental results demonstrate improved ImageNet FID scores and faster convergence, and it enables finetuning latent models into high-fidelity pixel-space generators.

Searching arXiv for the primary paper and closely related flow-matching generation work to ground the article. Asymmetric Flow Modeling, commonly abbreviated AsymFlow, denotes a rank-asymmetric velocity parameterization for flow-based generation in high-dimensional spaces. In this formulation, noise prediction is restricted to a low-rank subspace while data prediction remains full-dimensional; from this asymmetric prediction, the full-dimensional velocity is recovered analytically without changing the network architecture or training/sampling procedures (Chen et al., 13 May 2026). The formulation was introduced to address the difficulty of velocity prediction in pixel-space settings where the ambient dimension is large but the data exhibit strong low-rank structure. The same work reports a 1.57 FID on ImageNet 256×256 and describes the first-ever route for finetuning pretrained latent flow models into pixel-space models by aligning a low-rank pixel subspace to the latent space (Chen et al., 13 May 2026).

1. Terminological scope

The label “AsymFlow” is not unique to generative modeling. It has also been used for a symmetric-asymmetric collision comparison framework for small-system flow in heavy-ion physics (Huang et al., 22 Jul 2025), for numerical simulation of asymmetric merging flow in a rectangular channel (Siddiqui, 2013), and for $q$-heat flow, $q$-Laplacian, and Sobolev spaces on asymmetric metric measure spaces (Kristály et al., 3 Sep 2025). Related asymmetric-flow constructions also appear in studies of shark-inspired helical pipes (Levin et al., 2024), asymmetric longitudinal flow decorrelations in proton–nucleus collisions (Wu et al., 2021), phoretic flow induced by asymmetric confinement (Lisicki et al., 2016), and spatio-temporal flow propagation with conservation laws (Feng et al., 5 Nov 2025).

In current generative-model usage, however, Asymmetric Flow Modeling most specifically refers to the rank-asymmetric formulation introduced in “Asymmetric Flow Models” (Chen et al., 13 May 2026). In that context, “asymmetry” refers not to geometric directionality or forward–backward transport, but to the unequal treatment of the noise and data components of the flow velocity.

2. Motivation and problem setting

In flow-matching diffusion models, the interpolation between data $x_0 \in \mathbb{R}^D$ and noise is written as

$$x_t = \alpha_t x_0 + \sigma_t \epsilon,\quad \epsilon \sim \mathcal{N}(0,I_D),\quad \alpha_t = 1-t,\ \sigma_t=t.$$

The reverse-time ODE has velocity

$$v(x_t,t)=\mathbb{E}\bigl[(x_t-x_0)/\sigma_t \mid x_t\bigr] = \mathbb{E}[\epsilon - x_0 \mid x_t] = \mathbb{E}[\epsilon \mid x_t] - x_0.$$

This decomposition is the starting point for AsymFlow (Chen et al., 13 May 2026).

The central motivation is that modern large plain transformers, including DiT/JiT-like models, struggle in high-dimensional pixel-space because predicting the full noise term $\epsilon \in \mathbb{R}^D$ inflates internal activations with unstructured Gaussian noise. The comparison drawn in the paper is between two established parameterizations. In “$\epsilon$-prediction”, the model directly regresses full Gaussian noise, which requires full-rank noise. In “$x_0$-prediction”, the model regresses clean data and then recovers velocity through $(x_t-\hat x_0)/\sigma_t$, which is described as numerically unstable at low noise $\sigma_t \to 0$ (Chen et al., 13 May 2026).

AsymFlow is introduced as an intermediate construction. It predicts a full-rank data term but restricts the noise term to a low-dimensional subspace of rank $q$0. This suggests that the formulation is designed to preserve the semantic richness of full-dimensional data prediction while reducing the burden of modeling high-dimensional noise.

3. Rank-asymmetric velocity decomposition

The formalism begins with a matrix $q$1 with orthonormal columns, $q$2. The associated low-rank projector is

$q$3

with complementary projector

$q$4

The standard full velocity target is

$q$5

AsymFlow replaces this with the asymmetric target

$q$6

The asymmetry is explicit: the noise term is projected into $q$7, but the data term remains full-dimensional (Chen et al., 13 May 2026).

Projecting $q$8 into the two complementary subspaces yields

$q$9

The paper characterizes this as “$x_0 \in \mathbb{R}^D$0-like” behavior in $x_0 \in \mathbb{R}^D$1 and “$x_0 \in \mathbb{R}^D$2-like” behavior in $x_0 \in \mathbb{R}^D$3.

The full velocity required for the loss and sampler is then recovered analytically: $x_0 \in \mathbb{R}^D$4 This recovery formula is the defining mechanism of the method. It allows a network trained on the asymmetric target $x_0 \in \mathbb{R}^D$5 to supply the true flow velocity used in optimization and sampling, while leaving the backbone unchanged (Chen et al., 13 May 2026).

4. Training objective, architecture, and sampling

A notable implementation feature is that AsymFlow requires no changes to the transformer architecture, including JiT / DiT backbones. The modifications occur only in the training target and in the computations immediately before the loss and during sampling (Chen et al., 13 May 2026).

The method uses patch-wise tokenization, and each image patch of dimension $x_0 \in \mathbb{R}^D$6 uses the same projector $x_0 \in \mathbb{R}^D$7. The rank $x_0 \in \mathbb{R}^D$8 is a hyperparameter; the ImageNet experiments use $x_0 \in \mathbb{R}^D$9. Two procedures are given for constructing $$x_t = \alpha_t x_0 + \sigma_t \epsilon,\quad \epsilon \sim \mathcal{N}(0,I_D),\quad \alpha_t = 1-t,\ \sigma_t=t.$$0. For training from scratch, one collects many patches, runs PCA, and takes the top-$$x_t = \alpha_t x_0 + \sigma_t \epsilon,\quad \epsilon \sim \mathcal{N}(0,I_D),\quad \alpha_t = 1-t,\ \sigma_t=t.$$1 singular vectors. For finetuning from a latent model, $$x_t = \alpha_t x_0 + \sigma_t \epsilon,\quad \epsilon \sim \mathcal{N}(0,I_D),\quad \alpha_t = 1-t,\ \sigma_t=t.$$2 is obtained by solving an orthogonal Procrustes problem between latent tokens and decoded pixel patches (Chen et al., 13 May 2026).

Training minimizes the standard flow-matching loss after recovering the full velocity: $$x_t = \alpha_t x_0 + \sigma_t \epsilon,\quad \epsilon \sim \mathcal{N}(0,I_D),\quad \alpha_t = 1-t,\ \sigma_t=t.$$3 The workflow is specified in four steps: sample $$x_t = \alpha_t x_0 + \sigma_t \epsilon,\quad \epsilon \sim \mathcal{N}(0,I_D),\quad \alpha_t = 1-t,\ \sigma_t=t.$$4, $$x_t = \alpha_t x_0 + \sigma_t \epsilon,\quad \epsilon \sim \mathcal{N}(0,I_D),\quad \alpha_t = 1-t,\ \sigma_t=t.$$5, and $$x_t = \alpha_t x_0 + \sigma_t \epsilon,\quad \epsilon \sim \mathcal{N}(0,I_D),\quad \alpha_t = 1-t,\ \sigma_t=t.$$6; form $$x_t = \alpha_t x_0 + \sigma_t \epsilon,\quad \epsilon \sim \mathcal{N}(0,I_D),\quad \alpha_t = 1-t,\ \sigma_t=t.$$7; compute $$x_t = \alpha_t x_0 + \sigma_t \epsilon,\quad \epsilon \sim \mathcal{N}(0,I_D),\quad \alpha_t = 1-t,\ \sigma_t=t.$$8; recover the full $$x_t = \alpha_t x_0 + \sigma_t \epsilon,\quad \epsilon \sim \mathcal{N}(0,I_D),\quad \alpha_t = 1-t,\ \sigma_t=t.$$9 through Eq. (3), then compute $$v(x_t,t)=\mathbb{E}\bigl[(x_t-x_0)/\sigma_t \mid x_t\bigr] = \mathbb{E}[\epsilon - x_0 \mid x_t] = \mathbb{E}[\epsilon \mid x_t] - x_0.$$0 (Chen et al., 13 May 2026).

At inference time, the network outputs $$v(x_t,t)=\mathbb{E}\bigl[(x_t-x_0)/\sigma_t \mid x_t\bigr] = \mathbb{E}[\epsilon - x_0 \mid x_t] = \mathbb{E}[\epsilon \mid x_t] - x_0.$$1 at each step; the method then applies the same recovery formula to obtain $$v(x_t,t)=\mathbb{E}\bigl[(x_t-x_0)/\sigma_t \mid x_t\bigr] = \mathbb{E}[\epsilon - x_0 \mid x_t] = \mathbb{E}[\epsilon \mid x_t] - x_0.$$2. The paper states that everything else, including the ODE solver and guidance, remains unchanged. In the supplied pseudocode, the low-rank branch is $$v(x_t,t)=\mathbb{E}\bigl[(x_t-x_0)/\sigma_t \mid x_t\bigr] = \mathbb{E}[\epsilon - x_0 \mid x_t] = \mathbb{E}[\epsilon \mid x_t] - x_0.$$3, the orthogonal branch is $$v(x_t,t)=\mathbb{E}\bigl[(x_t-x_0)/\sigma_t \mid x_t\bigr] = \mathbb{E}[\epsilon - x_0 \mid x_t] = \mathbb{E}[\epsilon \mid x_t] - x_0.$$4, and the full velocity is their sum (Chen et al., 13 May 2026).

5. Latent-to-pixel finetuning

One of the distinctive claims of AsymFlow is that it provides the first practical route to lift large pretrained latent flows into high-fidelity pixel-space generators (Chen et al., 13 May 2026). The setup assumes a pretrained latent model $$v(x_t,t)=\mathbb{E}\bigl[(x_t-x_0)/\sigma_t \mid x_t\bigr] = \mathbb{E}[\epsilon - x_0 \mid x_t] = \mathbb{E}[\epsilon \mid x_t] - x_0.$$5 operating on latent tokens $$v(x_t,t)=\mathbb{E}\bigl[(x_t-x_0)/\sigma_t \mid x_t\bigr] = \mathbb{E}[\epsilon - x_0 \mid x_t] = \mathbb{E}[\epsilon \mid x_t] - x_0.$$6 with latent-space velocity $$v(x_t,t)=\mathbb{E}\bigl[(x_t-x_0)/\sigma_t \mid x_t\bigr] = \mathbb{E}[\epsilon - x_0 \mid x_t] = \mathbb{E}[\epsilon \mid x_t] - x_0.$$7.

A lift matrix $$v(x_t,t)=\mathbb{E}\bigl[(x_t-x_0)/\sigma_t \mid x_t\bigr] = \mathbb{E}[\epsilon - x_0 \mid x_t] = \mathbb{E}[\epsilon \mid x_t] - x_0.$$8 with orthonormal columns is chosen by orthogonal Procrustes between decoded latent representations and pixel patches. The low-rank pixel variables are then defined as

$$v(x_t,t)=\mathbb{E}\bigl[(x_t-x_0)/\sigma_t \mid x_t\bigr] = \mathbb{E}[\epsilon - x_0 \mid x_t] = \mathbb{E}[\epsilon \mid x_t] - x_0.$$9

The paper states that if both ODEs start from paired noise, the pixel trajectory

$\epsilon \in \mathbb{R}^D$0

exactly follows the lifted latent model plus a known orthogonal noise drift (Chen et al., 13 May 2026).

This leads to a three-part initialization scheme. The input layer maps pixel patches $\epsilon \in \mathbb{R}^D$1 to latent $\epsilon \in \mathbb{R}^D$2, together with time-mapping and scale calibration. The backbone is the frozen pretrained latent flow network $\epsilon \in \mathbb{R}^D$3. The output layer maps the latent prediction back to the pixel low-rank branch $\epsilon \in \mathbb{R}^D$4 and combines it with the orthogonal branch $\epsilon \in \mathbb{R}^D$5 (Chen et al., 13 May 2026).

Because orthogonal Procrustes determines directions but not scale, the formulation adds a scalar $\epsilon \in \mathbb{R}^D$6 and a time reparameterization

$\epsilon \in \mathbb{R}^D$7

so that signal–noise ratios align exactly with the latent schedule.

For finetuning, the paper defines a variance-reduced loss

$\epsilon \in \mathbb{R}^D$8

with

$\epsilon \in \mathbb{R}^D$9

A further LPIPS loss is introduced as a perceptual correction, interpolated by a schedule $\epsilon$0, producing the final finetune objective

$\epsilon$1

The stated rationale is that, at initialization, the pixel model predicts $\epsilon$2, so the only gap to the true AsymFlow target $\epsilon$3 is the low-level difference $\epsilon$4; finetuning then needs only to correct that projection error (Chen et al., 13 May 2026).

6. Empirical results and reported performance

For ImageNet 256×256, AsymFlow is trained from scratch using a JiT-H/16 plain transformer, patch size 16, and rank $\epsilon$5 PCA subspace. Under guided-sampling FIDs under ADM evaluation, the reported baseline is JiT-H/16 full $\epsilon$6 prediction with FID $\epsilon$7, while AsymFlow ($\epsilon$8) + REPA loss achieves FID $\epsilon$9 (Chen et al., 13 May 2026).

Ablations reported in the same source vary the rank from $x_0$0 and identify the best result at $x_0$1. The paper further states that a random subspace yields no gain, and that convergence at $x_0$2 is $x_0$3 faster. This suggests that the benefit depends not only on reduced rank but also on a subspace aligned to data structure.

For text-to-image generation, the method finetunes FLUX.2 klein (9B) into a 1024×1024 pixel AsymFlow model, referred to as “AsymFLUX.2 klein,” on 3M LAION-Aesthetics images. The procedure freezes the main transformer and finetunes only the input/output projections together with rank-256 LoRA adapters. Sampling uses 32 steps UniPC + orthogonal-projection guidance (Chen et al., 13 May 2026).

The reported benchmark changes are:

• HPSv3: 9.50 $x_0$4 10.66
• DPG-Bench: 85.2 $x_0$5 86.8
• GenEval: 0.80 $x_0$6 0.82

The same source states that, against PixelDiT-T2I (1024×1024 pixel diffusion), AsymFLUX improves on all three metrics by a clear margin, and that the generated images qualitatively show sharper textures, richer styles, and higher realism (Chen et al., 13 May 2026).

Within the scope of the reported results, AsymFlow is therefore positioned as a method that addresses a specific bottleneck of plain diffusion transformers: the requirement to model full-dimensional Gaussian noise in pixel space. Its asymmetry is not architectural asymmetry in the usual network-design sense, but an asymmetry in the target decomposition of the velocity field, combined with an exact analytical reconstruction of the full flow.