Research map / Probabilistic and scientific modeling / Molecules and proteins

Disentangled Representation Learning via Flow Matching

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Equivariant and disentangled latent representations · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a flow-matching framework for disentangled representation learning that conditions a latent flow on N factor embeddings and adds an orthogonality regularizer over factor-specific velocity components, reporting gains over VAE, GAN and diffusion baselines on Cars3D, Shapes3D, MPI3D-toy and CelebA.

Problem

Diffusion-based disentangled representation methods rely on inductive biases for statistical independence but lack strong semantic alignment between latent units and generative factors.

Approach

Casts disentanglement as factor-conditioned flow matching in a VQ-GAN latent space, decomposing the predicted velocity into factor-specific components via output attention (softmax routing over factor embeddings) and penalizing pairwise squared cosine similarity between these components.

Key results

• On MPI3D-toy, the method reaches 0.907 FactorVAE and 0.649 DCI, versus 0.410/0.300 for the best VAE baseline (DAVA) and 0.601/0.312 for DisDiff.
• On CelebA, TAD improves to 1.154 (vs. 0.954 for DyGA) and FID to 8.1 (vs. 12.0 for DyGA).
• Ablation on MPI3D-toy: adding the orthogonality loss raises FactorVAE from 0.745 to 0.907 and DCI from 0.571 to 0.649.

Surprising findings

• Gains over diffusion baselines on Cars3D and Shapes3D are small, and FactorVAE on Cars3D is slightly below the EncDiff baseline (0.948 vs. 0.964 is above, but FDAE 0.912 and DisDiff 0.924 are below).

Limitations

• Several baselines use pre-trained encoders in their original setups, so comparisons may not be controlled.
• Reported gains over diffusion baselines are modest on synthetic datasets, and the method adds a factor encoder and orthogonality hyperparameter tuned on validation.
• Factor count N must be fixed in advance (set to 10), and performance varies noticeably with N.

Experimental setup

Datasets: Cars3D Shapes3D MPI3D-toy CelebA

Benchmarks: Cars3D Shapes3D MPI3D-toy CelebA

Metrics: FactorVAE score DCI TAD FID Acc100 Acc1000

Baselines: FactorVAE beta-TCVAE DAVA ClosedForm GANSpace DisCo-GAN DisDiff FDAE EncDiff DyGA

Builds on: conditional flow matching DisDiff VQ-GAN FactorVAE beta-TCVAE

Closest papers by meaning

• Flow Factorized Representation Learning (NeurIPS 2023 · similarity 0.91)
• Learning Disentangled Representation by Exploiting Pretrained Generative Models: A Contrastive Learning View (ICLR 2022 · similarity 0.90)
• Leveraging sparse and shared feature activations for disentangled representation learning (NeurIPS 2023 · similarity 0.90)
• Graph-based Unsupervised Disentangled Representation Learning via Multimodal Large Language Models (NeurIPS 2024 · similarity 0.90)
• Variational Inference of Disentangled Latent Concepts from Unlabeled Observations (ICLR 2018 · similarity 0.90)
• Transferring disentangled representations: bridging the gap between synthetic and real images (NeurIPS 2024 · similarity 0.89)