Research map / Image generation / Fast diffusion sampling
Exploring Diffusion Time-steps for Unsupervised Representation Learning
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 DiTi, which learns a time-step-partitioned feature that compensates a frozen DDPM's reconstruction error at each diffusion step, yielding disentangled attribute features for unsupervised representation learning.
Problem
Unsupervised disentangled representation learning from diffusion models, which have no explicit encoder; prior DM-based feature learners entangle attributes and leak information across time-steps.
Approach
Connects diffusion time-steps to modular attributes (early steps lose fine-grained attributes, late steps coarse ones) and trains an expanding set of feature subsets, where the first t subsets (others masked) must compensate the frozen DM's reconstruction error at step t.
Key results
- DiTi reaches 62.3% AP on CelebA attribute classification, versus 60.2% for PDAE and 60.3% for Diff-AE.
- On LFW attribute regression (trained on CelebA), DiTi achieves Pearson's r 0.617 and MSE 0.392, compared with PDAE's 0.596 and 0.410.
- On FFHQ, DiTi reaches 61.4% AP versus 59.7% for PDAE, and DiTi is the only method producing artifact-free counterfactuals in qualitative comparisons.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use time-step-specific, cumulatively masked features instead of a time-step-independent feature | PDAE-style diffusion representation · CelebA (trained on CelebA) | AP: 60.2 -> 62.3 | Helped |
| Use time-step-specific, cumulatively masked features instead of a time-step-independent feature | PDAE-style diffusion representation · LFW attribute regression (trained on CelebA) | Pearson's r: 0.596 -> 0.617 | Helped |
| Use time-step-specific, cumulatively masked features instead of a time-step-independent feature | PDAE-style diffusion representation · LFW attribute regression (trained on CelebA) | MSE: 0.410 -> 0.392 | Helped |
| Use time-step-specific, cumulatively masked features instead of a time-step-independent feature | PDAE-style diffusion representation · CelebA (trained on FFHQ) | AP: 59.7 -> 61.4 | Helped |
| Use time-step-specific, cumulatively masked features instead of a time-step-independent feature | PDAE-style diffusion representation · LFW attribute regression (trained on FFHQ) | Pearson's r: 0.603 -> 0.622 | Helped |
| Use time-step-specific, cumulatively masked features instead of a time-step-independent feature | PDAE-style diffusion representation · LFW attribute regression (trained on FFHQ) | MSE: 0.416 -> 0.384 | Helped |
| Vary the number of feature subsets k (ablation) | DiTi · CelebA attribute inference | AP | No clear effect |
| Allocate more feature dimensions to time-steps 100–300 and fewer to 500–1000 (ablation) | DiTi · CelebA attribute inference | AP | Helped |
| Detach gradients for earlier feature subsets so only the current subset is trained (ablation) | DiTi-Detach · LFW attribute regression (feature trained on CelebA) | Regression performance | Hurt |
| Remove color jittering and random grayscale (ablation) | SimCLR · CelebA (trained on CelebA) | AP: 59.7 -> 34.7\% | Hurt |
| Remove color jittering and random grayscale (ablation) | SimCLR · LFW attribute regression (trained on CelebA) | Pearson's r: 0.474 -> 0.176 | Hurt |
Surprising findings
- SimCLR without color-related augmentations drops to 34.7% AP, showing that removing augmentations does not fix its attribute-inference deficiency.
- Detaching gradients of earlier feature subsets improves disentanglement for generation but slows convergence and hurts LFW regression.
Limitations
- A disentangled feature is sufficient but not necessary for the objective, so the modularity is not guaranteed and some attributes may leak across subsets.
- Fine-grained attributes such as earrings are not well disentangled; SimCLR slightly outperforms on some of them.
- Subset-to-time-step partitioning and dimension allocation are heuristic and not searched.
Experimental setup
Datasets: CelebA FFHQ LFW LSUN Bedroom
Benchmarks: CelebA LFW FFHQ
Metrics: AP Pearson's r MSE
Baselines: beta-TCVAE IB-GAN SimCLR Diff-AE PDAE
Builds on: PDAE DDPM DDIM ProbMask
Closest papers by meaning
- Understanding Representation Dynamics of Diffusion Models via Low-Dimensional Modeling (NeurIPS 2025 · similarity 0.87)
- CleanDIFT: Diffusion Features without Noise (CVPR 2025 · similarity 0.87)
- Unsupervised Region-Based Image Editing of Denoising Diffusion Models (AAAI 2025 · similarity 0.87)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.87)
- Deconstructing Denoising Diffusion Models for Self-Supervised Learning (ICLR 2025 · similarity 0.87)
- SDMG: Smoothing Your Diffusion Models for Powerful Graph Representation Learning (ICML 2025 · similarity 0.87)
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.