Disentangled Latent-CFM: Methods & Applications

• The paper introduces an auxiliary-variable-guided extension to latent conditional flow matching that enforces semantic alignment between latent dimensions and physical factors, improving generation fidelity.
• DL-CFM partitions the latent space into guided and residual components, enabling interpretable generative control and recovery of established physical relations such as the halo mass-concentration trend.
• The method integrates flow matching, conditional KL divergence, and correlation-based penalties to maintain high generative quality while facilitating robust disentangled representation learning.

Searching arXiv for the cited papers and closely related work to ground the article. Search query: arXiv (Ganguli et al., 26 Feb 2026) Disentangled Latent-CFM auxiliary-variable-guided generative models halo structures Disentangled Latent-CFM (DL-CFM) denotes a class of latent conditional flow-matching models in which the latent representation is explicitly structured so that selected coordinates align with interpretable factors rather than remaining entangled. In the astrophysical formulation, DL-CFM is introduced as an auxiliary-variable-guided extension of latent conditional flow matching for thermal Sunyaev–Zel’dovich halo maps, where halo mass and concentration are used to shape the latent space while preserving generative fidelity (Ganguli et al., 26 Feb 2026). The same acronym has also been used for a more general flow-matching framework that treats disentanglement as factor-conditioned transport in a compact latent space with an orthogonality-based non-overlap constraint (Chi et al., 5 Feb 2026). Across these uses, the common idea is that generative flow in latent space should be decomposed or regularized so that latent directions acquire a sharper semantic interpretation.

1. Genealogy from latent conditional flow matching

DL-CFM is most naturally understood relative to Latent-CFM, which augments conditional flow matching with a latent variable $f$ capturing global factors, modes, or semantic structure of the target distribution (Samaddar et al., 7 May 2025). The motivation is that standard flow matching usually starts from $p_0(x)=\mathcal N(0,I)$, whereas many datasets lie on a lower-dimensional, multimodal manifold. Latent-CFM therefore conditions the vector field on a latent code so that the model learns transport within a latent-selected subpopulation rather than learning all modes at once.

In that formulation, the endpoint coupling is factorized as

$$q(x_0,x_1)=\int q(f)\,q(x_0,x_1\mid f)\,df, \qquad q(x_0,x_1\mid f)=p_0(x_0)\,p_1(x_1\mid f),$$

and the latent-conditioned objective becomes

$$\mathcal{L}_{\text{Latent-CFM}} = \mathbb{E}_{t,q(x_0,x_1),q(f\mid x_0,x_1),p_t(x\mid x_0,x_1)} \left[\|v_\theta(x,f,t)-u_t(x\mid x_0,x_1)\|_2^2\right].$$

The paper considers a pretrained VAE encoder or, in synthetic settings, a pretrained Gaussian mixture model as the latent structure model. It reports improved generation quality with significantly less training, up to $\sim 50\%$ less, and computation than state-of-the-art flow matching models, while also giving lower PDE residuals on a 2d Darcy flow dataset and enabling latent-feature-conditioned generation (Samaddar et al., 7 May 2025).

This suggests a direct lineage: DL-CFM preserves the latent-conditioned transport perspective of Latent-CFM, but adds explicit mechanisms for disentanglement, semantic alignment, or both. In the astrophysical variant, this addition takes the form of auxiliary-informed priors and correlation-based penalties; in the representation-learning variant, it takes the form of factor-conditioned routing and orthogonality regularization.

2. Auxiliary-guided DL-CFM for astrophysical image generation

In the astrophysical paper, DL-CFM is designed for simulated Compton-$y$ maps of galaxy clusters from CRK-HACC hydrodynamic simulations, with the explicit goal of making the latent space both generative and physically interpretable (Ganguli et al., 26 Feb 2026). The starting point is a conditional flow matching formulation based on the ODE

$$\frac{d \phi_t(x)}{dt} = u_t(\phi_t(x)), \qquad \phi_0(x)=x_0,$$

together with the conditional path

$$p_t(\cdot\mid x_0,x)=\mathcal{N}\big(\,\cdot;\, t x + (1-t)x_0,\ \sigma^2 I_p\big),$$

where $p_0=\mathcal{N}(0,I)$, typically $\sigma=0$ in their experiments, and

$p_0(x)=\mathcal N(0,I)$0

DL-CFM lifts this construction into a latent-variable setting by conditioning the neural vector field on a latent code inferred from the image,

$p_0(x)=\mathcal N(0,I)$1

The latent code is partitioned as

$p_0(x)=\mathcal N(0,I)$2

where $p_0(x)=\mathcal N(0,I)$3 is the auxiliary-guided block and $p_0(x)=\mathcal N(0,I)$4 is the residual or reconstruction-focused block. In the reported application, $p_0(x)=\mathcal N(0,I)$5 and the auxiliary variables are

$p_0(x)=\mathcal N(0,I)$6

corresponding to $p_0(x)=\mathcal N(0,I)$7 after normalization to $p_0(x)=\mathcal N(0,I)$8.

The latent prior is conditioned on the auxiliary variables: $p_0(x)=\mathcal N(0,I)$9 with

$$q(x_0,x_1)=\int q(f)\,q(x_0,x_1\mid f)\,df, \qquad q(x_0,x_1\mid f)=p_0(x_0)\,p_1(x_1\mid f),$$0

Because $$q(x_0,x_1)=\int q(f)\,q(x_0,x_1\mid f)\,df, \qquad q(x_0,x_1\mid f)=p_0(x_0)\,p_1(x_1\mid f),$$1, the guided coordinates are softly tethered to the supplied physical quantities, while the residual coordinates remain roughly standard normal. The intended division is that the guided block should encode the known physical factors one-to-one, whereas the residual block should capture remaining morphology, including asymmetries, mergers, or other structure not explained by mass and concentration alone (Ganguli et al., 26 Feb 2026).

3. Objective function and training mechanics

The full DL-CFM objective combines flow matching, a conditional KL term, and correlation-based disentanglement penalties (Ganguli et al., 26 Feb 2026): $$q(x_0,x_1)=\int q(f)\,q(x_0,x_1\mid f)\,df, \qquad q(x_0,x_1\mid f)=p_0(x_0)\,p_1(x_1\mid f),$$2

Its terms are assigned distinct roles. The flow-matching term preserves the transport objective. The KL term keeps the inferred posterior close to the auxiliary-informed prior. The alignment term enforces explicitness, meaning latent coordinate $$q(x_0,x_1)=\int q(f)\,q(x_0,x_1\mid f)\,df, \qquad q(x_0,x_1\mid f)=p_0(x_0)\,p_1(x_1\mid f),$$3 should track $$q(x_0,x_1)=\int q(f)\,q(x_0,x_1\mid f)\,df, \qquad q(x_0,x_1\mid f)=p_0(x_0)\,p_1(x_1\mid f),$$4. The intra-independence penalty suppresses leakage from one auxiliary variable into the other guided coordinates, and the inter-independence penalty suppresses correlation between the auxiliary variables and the residual block.

The penalties are implemented with minibatch correlation statistics using polynomial feature lifts up to degree $$q(x_0,x_1)=\int q(f)\,q(x_0,x_1\mid f)\,df, \qquad q(x_0,x_1\mid f)=p_0(x_0)\,p_1(x_1\mid f),$$5: $$q(x_0,x_1)=\int q(f)\,q(x_0,x_1\mid f)\,df, \qquad q(x_0,x_1\mid f)=p_0(x_0)\,p_1(x_1\mid f),$$6

$$q(x_0,x_1)=\int q(f)\,q(x_0,x_1\mid f)\,df, \qquad q(x_0,x_1\mid f)=p_0(x_0)\,p_1(x_1\mid f),$$7

with

$$q(x_0,x_1)=\int q(f)\,q(x_0,x_1\mid f)\,df, \qquad q(x_0,x_1\mid f)=p_0(x_0)\,p_1(x_1\mid f),$$8

The paper describes these as lightweight, scale-free, and computable from minibatches.

Architecturally, DL-CFM keeps the U-Net backbone from ICFM for the vector field model but adds an encoder branch: a deep convolutional network with four downsampling convolution blocks, each followed by batch normalization and leaky ReLU, then two dilated convolution blocks, then flattening and a linear head producing latent mean and log-variance. Latent sampling uses the reparameterization trick, the sampled latent is projected through a trainable MLP into an embedding space, and the latent embedding is added to the time embedding and fed into the same U-Net used by ICFM. Training jointly optimizes encoder and vector field parameters for 240k steps with Adam, learning rate $$q(x_0,x_1)=\int q(f)\,q(x_0,x_1\mid f)\,df, \qquad q(x_0,x_1\mid f)=p_0(x_0)\,p_1(x_1\mid f),$$9, batch size 128, latent dimension 256, and regularization weights

$$\mathcal{L}_{\text{Latent-CFM}} = \mathbb{E}_{t,q(x_0,x_1),q(f\mid x_0,x_1),p_t(x\mid x_0,x_1)} \left[\|v_\theta(x,f,t)-u_t(x\mid x_0,x_1)\|_2^2\right].$$0

Inference samples a latent code once from the encoder and then solves the ODE with an adaptive solver, dopri5 (Ganguli et al., 26 Feb 2026).

4. Empirical behavior and scientific interpretation

The reported empirical claim is not that DL-CFM merely decorrelates latent coordinates, but that it preserves generation quality while making latent directions physically legible (Ganguli et al., 26 Feb 2026). Relative to the baseline ICFM model, the distance metrics between real and generated samples are described as similar overall: Sinkhorn distance is slightly better for ICFM, Energy distance is slightly better for DL-CFM, and Gaussian and Laplacian distances are essentially matched. The paper therefore treats interpretability gains as not coming at the cost of a substantial fidelity loss.

The most prominent result concerns the structure of the latent space. The first two latent dimensions, corresponding to $$\mathcal{L}_{\text{Latent-CFM}} = \mathbb{E}_{t,q(x_0,x_1),q(f\mid x_0,x_1),p_t(x\mid x_0,x_1)} \left[\|v_\theta(x,f,t)-u_t(x\mid x_0,x_1)\|_2^2\right].$$1, show near one-to-one, monotonic relationships with $$\mathcal{L}_{\text{Latent-CFM}} = \mathbb{E}_{t,q(x_0,x_1),q(f\mid x_0,x_1),p_t(x\mid x_0,x_1)} \left[\|v_\theta(x,f,t)-u_t(x\mid x_0,x_1)\|_2^2\right].$$2 and $$\mathcal{L}_{\text{Latent-CFM}} = \mathbb{E}_{t,q(x_0,x_1),q(f\mid x_0,x_1),p_t(x\mid x_0,x_1)} \left[\|v_\theta(x,f,t)-u_t(x\mid x_0,x_1)\|_2^2\right].$$3. The generated latent scatter also reproduces the known mass-concentration relation from the simulation catalog. This is the central scientific claim of the method: an auxiliary-guided generative model can recover an established physical scaling relation while still retaining a residual latent subspace for morphology not summarized by those two variables.

Latent traversals are used to demonstrate controllability. Varying the guided coordinates while keeping $$\mathcal{L}_{\text{Latent-CFM}} = \mathbb{E}_{t,q(x_0,x_1),q(f\mid x_0,x_1),p_t(x\mid x_0,x_1)} \left[\|v_\theta(x,f,t)-u_t(x\mid x_0,x_1)\|_2^2\right].$$4 fixed changes halo images systematically with mass and concentration. Conversely, holding the guided coordinates fixed and varying the residual block produces samples in which the center of the residual latent distribution appears relaxed and single-peaked, while tail samples exhibit multi-peaked or disturbed morphologies. The authors interpret these as possible mergers, disturbed systems, or outliers not captured by the two auxiliary variables alone. In that sense, $$\mathcal{L}_{\text{Latent-CFM}} = \mathbb{E}_{t,q(x_0,x_1),q(f\mid x_0,x_1),p_t(x\mid x_0,x_1)} \left[\|v_\theta(x,f,t)-u_t(x\mid x_0,x_1)\|_2^2\right].$$5 is treated not as nuisance variation but as a repository of residual physical complexity (Ganguli et al., 26 Feb 2026).

The broader significance claimed in the paper is that auxiliary guidance preserves generative flexibility while yielding physically meaningful, disentangled embeddings, thereby turning the latent space into a diagnostic tool for cosmological structure. A plausible implication is that the method is simultaneously a generative model and a representation-analysis instrument: one part of latent space is reserved for known factors, while the remaining part can be interrogated for atypical structure.

5. The parallel DL-CFM formulation in disentangled representation learning

A distinct paper uses the same acronym, DL-CFM, for a more general disentangled representation learning framework based on flow matching rather than for the astrophysical auxiliary-variable construction (Chi et al., 5 Feb 2026). There, disentanglement is cast as learning factor-conditioned flows in a compact latent space. The method encodes an image into a latent target with a pretrained VQ-GAN encoder,

$$\mathcal{L}_{\text{Latent-CFM}} = \mathbb{E}_{t,q(x_0,x_1),q(f\mid x_0,x_1),p_t(x\mid x_0,x_1)} \left[\|v_\theta(x,f,t)-u_t(x\mid x_0,x_1)\|_2^2\right].$$6

extracts factor tokens

$$\mathcal{L}_{\text{Latent-CFM}} = \mathbb{E}_{t,q(x_0,x_1),q(f\mid x_0,x_1),p_t(x\mid x_0,x_1)} \left[\|v_\theta(x,f,t)-u_t(x\mid x_0,x_1)\|_2^2\right].$$7

samples

$$\mathcal{L}_{\text{Latent-CFM}} = \mathbb{E}_{t,q(x_0,x_1),q(f\mid x_0,x_1),p_t(x\mid x_0,x_1)} \left[\|v_\theta(x,f,t)-u_t(x\mid x_0,x_1)\|_2^2\right].$$8

and interpolates linearly,

$$\mathcal{L}_{\text{Latent-CFM}} = \mathbb{E}_{t,q(x_0,x_1),q(f\mid x_0,x_1),p_t(x\mid x_0,x_1)} \left[\|v_\theta(x,f,t)-u_t(x\mid x_0,x_1)\|_2^2\right].$$9

A conditional latent vector field

$\sim 50\%$0

is then trained with

$\sim 50\%$1

The defining mechanism is factorized velocity decomposition: $\sim 50\%$2 implemented by an output-attention routing mask with

$\sim 50\%$3

Each factor-specific component is gated by its routing weight, and a non-overlap regularizer penalizes cosine similarity between factor-specific velocities: $\sim 50\%$4 where $\sim 50\%$5 is the flattened factor-specific velocity. The full objective is

$\sim 50\%$6

with $\sim 50\%$7 in all experiments.

This formulation is evaluated on Cars3D, Shapes3D, MPI3D-toy, and CelebA. The reported results include best-in-class performance on Shapes3D, especially strong gains on MPI3D-toy with FactorVAE score $\sim 50\%$8 and DCI $\sim 50\%$9, and on CelebA, TAD $y$0 and FID $y$1. The paper also reports an ablation in which adding the orthogonality regularizer improves Cars3D, Shapes3D, and MPI3D-toy over bare flow matching. At the same time, it explicitly notes that orthogonality is a proxy and not a perfect guarantee of true causal disentanglement (Chi et al., 5 Feb 2026).

For encyclopedia purposes, this naming overlap is significant. The term “DL-CFM” therefore does not denote a single universally fixed algorithm; it denotes at least two closely related flow-matching-based disentanglement programs, one centered on auxiliary physical variables in scientific imaging and one centered on factor-conditioned latent transport with output-attention and orthogonality constraints.

6. Relation to adjacent work, misconceptions, and open limitations

DL-CFM is best viewed as part of a broader movement that combines latent conditioning with disentanglement-oriented structure in generative transport. It is not equivalent to plain Latent-CFM, whose principal contribution is efficient latent-conditioned flow matching with pretrained latent variable models and conditional generation based on latent features. Nor is it identical to diffusion-style time-axis disentanglement, although the latter provides a nearby conceptual precedent: “Disentanglement in T-space for Faster and Distributed Training of Diffusion Models with Fewer Latent-states” argues that timesteps can be disentangled into independently trained experts and then composed at inference, framing this as complete disentanglement in T-space (Gupta et al., 20 Aug 2025).

Several misconceptions are therefore best avoided. First, disentanglement in DL-CFM does not mean that every latent coordinate becomes automatically interpretable without auxiliary structure or routing constraints. In the astrophysical version, interpretability is produced by the conditional prior and correlation-based alignment penalties; in the factor-conditioned version, it is produced by token conditioning plus non-overlap regularization. Second, disentanglement does not imply causal identification. The representation-learning paper explicitly treats orthogonality as a proxy rather than a proof of true causal disentanglement, and the astrophysical paper confines its strongest claims to recovery of mass-concentration structure and organization of residual morphology. Third, the method family remains dependent on the quality of its latent machinery. Latent-CFM relies on pretrained latent models, and the astrophysical DL-CFM relies on auxiliary variables that are available and scientifically meaningful.

The current limitations are correspondingly concrete. Latent-CFM approximates sampling from the latent marginal by reusing empirical training samples, which the paper presents as practical but not ideal. The auxiliary-guided astrophysical DL-CFM is demonstrated on simulated halo maps with known mass and concentration labels; the paper argues for generalizability in scientific machine learning, but broader deployment would still depend on suitable auxiliary variables. The representation-learning DL-CFM fixes the number of factors and uses a pretrained VQ-GAN latent space, so its behavior is partly inherited from that latent representation. These points suggest that DL-CFM is less a single closed-form recipe than a design pattern: latent-conditioned flow matching equipped with structural constraints that force part of the latent transport to align with interpretable factors.