Research map / Label-efficient and robust learning / Recognition and long-tail
Disentangled Representation Learning via Modular Compositional Bias
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 modular "compositional bias" for disentangled representation learning, where the only factor-specific component is a mixing strategy over latents from two images, combined with a diffusion-based prior loss and a compositional consistency loss, enabling attribute, object, and joint style-object disentanglement under one objective and architecture.
Problem
Existing disentangled representation learning methods embed factor-specific inductive biases in objectives or architectures (information-theoretic losses for attributes, slot attention or additive decoders for objects), so new or co-occurring factors require redesigning the model.
Approach
Decouple inductive bias from objectives and architecture: factor-specific recombination rules (mutually exclusive for attributes, freely exchangeable for objects) are encoded as a stochastic latent mixing strategy; a frozen unconditional diffusion prior ensures composites are realistic and a relative InfoNCE-style consistency loss ensures composites match their latents. Switching the mixing strate
Key results
- On Shapes3D, the method reaches 0.975 FactorVAE and 0.837 DCI, outperforming the second-best methods by about 8% in FactorVAE and 15.7-21.4% in DCI.
- On CLEVR-Style, the method achieves over 95% style prediction accuracy, while LSD and L2C stay near chance (~10%) on style.
- On CLEVRTex object segmentation, the method reports FG-ARI 87.68 vs 82.55 for L2C and 71.64 for LSD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed framework with compositional mixing | Attribute disentanglement · Shapes3D and MPI3D | FactorVAE: 8% | Helped |
| use the proposed framework with compositional mixing | Attribute disentanglement · Shapes3D and MPI3D | DCI: 15.7--21.4% | Helped |
| use the proposed framework with compositional mixing | Attribute disentanglement · Cars3D | DCI | Helped |
| use the proposed method instead of LSD | Object segmentation · CLEVRTex | FG-ARI: 71.64 -> 87.68 | Helped |
| use the proposed method instead of L2C | Object segmentation · CLEVR | mIoU: 25.61 -> 26.54 | Helped |
| use the proposed method instead of LSD | Joint attribute and object disentanglement · CLEVR-Style | Style accuracy: 23.30 -> 96.50 | Helped |
| use the proposed method instead of Additive Decoder w/o SA | Attribute disentanglement · Shapes3D | FactorVAE: 0.000 -> 0.975 | Helped |
| replace mixing 2 images with mixing 64 images (ablation) | Attribute disentanglement · Attribute disentanglement | FactorVAE: 0.975±0.040 -> 0.966±0.032 | No clear effect |
| replace mixing 2 images with mixing 64 images (ablation) | Attribute disentanglement · Attribute disentanglement | DCI: 0.837±0.105 -> 0.802±0.088 | No clear effect |
| add the prior loss (ablation) | Proposed method · Attribute disentanglement | disentanglement performance | Helped |
| add the compositional consistency loss (ablation) | Proposed method · Attribute disentanglement | disentanglement performance | Helped |
| use all three losses together (ablation) | Proposed method · Object disentanglement | disentanglement performance | Helped |
Surprising findings
- Mixing 64 images instead of 2 yields no meaningful difference in FactorVAE/DCI, supporting the equivalence argument.
- An additive decoder alone fails on attribute disentanglement (FactorVAE 0.000 on Shapes3D) and does not disentangle objects without slot attention.
Limitations
- Assumes the target factors of variation are known in advance; the mixing strategy must be chosen by the practitioner.
- No theoretical guarantees of disentanglement are provided.
- Object segmentation on CLEVR is weaker in mIoU/mBO because backgrounds are included in object latents.
Experimental setup
Datasets: Shapes3D Cars3D MPI3D CLEVR-Easy CLEVR CLEVR-Tex CLEVR-Style MSN-Style MultiShapeNet
Benchmarks: Shapes3D Cars3D MPI3D CLEVR CLEVR-Easy CLEVRTex CLEVR-Style MSN-Style
Metrics: FactorVAE score DCI disentanglement property prediction accuracy FG-ARI mIoU mBO GRAM loss style prediction accuracy
Baselines: FactorVAE beta-TCVAE InfoGAN-CR GANspace LatentDiscovery DisCo DisDiff SA SLASH LSD L2C
Builds on: Learning to Compose (L2C) DisDiff Latent Slot Diffusion (LSD) Score distillation sampling (SDS) InfoNCE
Closest papers by meaning
- Graph-based Unsupervised Disentangled Representation Learning via Multimodal Large Language Models (NeurIPS 2024 · similarity 0.89)
- Tripod: Three Complementary Inductive Biases for Disentangled Representation Learning (ICML 2024 · similarity 0.89)
- Transferring disentangled representations: bridging the gap between synthetic and real images (NeurIPS 2024 · similarity 0.88)
- Disentangled Representation Learning via Flow Matching (NeurIPS 2026 · similarity 0.88)
- Factorized Diffusion Autoencoder for Unsupervised Disentangled Representation Learning (AAAI 2024 · similarity 0.88)
- Disentanglement Beyond Static vs. Dynamic: A Benchmark and Evaluation Framework for Multi-Factor Sequential Representations (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.