Research map / Image generation / Autoregressive and latent generation
Diffusion Bridge AutoEncoders 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 Diffusion Bridge AutoEncoders (DBAE), which infers the diffusion endpoint x_T from a latent z via a feed-forward decoder and a learned h-transform bridge, so z becomes an information bottleneck for reconstruction and generation.
Problem
Auxiliary-encoder diffusion representation models (e.g., DiffAE) split information between the latent z and the diffusion endpoint x_T, which hurts reconstruction and makes x_T expensive (ODE solving) and inflexible in dimension.
Approach
Make the endpoint x_T a feed-forward function of z (x_T = Dec(z)) and define the forward diffusion as a Doob h-transform bridge from x_0 to this x_T, so z is the sole bottleneck; train with a score-matching objective that is shown equivalent to a reconstruction loss plus entropy regularization.
Key results
- On CelebA linear-probe attribute prediction with 512-dim latent, DBAE reaches AP 0.655 vs DiffAE 0.603 and DiTi 0.623.
- Reconstruction on CelebA-HQ with random x_T: DBAE-d achieves SSIM 0.953 and MSE 2.49e-3, versus DiffAE (random x_T) SSIM 0.677.
- Unconditional FFHQ generation with prior on latent: DBAE+AE reaches FID 1.77 vs DiffAE+AE 2.84; on CelebA disentanglement, DBAE+TC gets TAD 0.362 vs DisDiff 0.305.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a z-dependent endpoint x_T to make z an information bottleneck | DBAE · CelebA and FFHQ linear-probe attribute prediction | AP: CelebA: 0.623 -> 0.655; FFHQ: 0.614 -> 0.664 | Helped |
| use a z-dependent endpoint x_T to make z an information bottleneck | DBAE · CelebA-HQ reconstruction | SSIM: 0.677 -> 0.920 | Helped |
| use a deterministic encoder (ablation) | DBAE · CelebA-HQ reconstruction | SSIM: 0.920 -> 0.953 | Helped |
| use a deterministic encoder (ablation) | DBAE · CelebA linear-probe attribute prediction | AP: 0.655 -> 0.650 | Hurt |
| apply total-correlation regularization (ablation) | DBAE · CelebA disentanglement | TAD: 0.124$\pm{0.078}$ -> 0.362$\pm{0.036}$ | Helped |
| apply total-correlation regularization (ablation) | DBAE · CelebA sample quality | FID: 11.8$\pm{0.2}$ -> 13.4$\pm{0.2}$ | Mixed |
| use the inferred latent distribution q_phi(z) instead of the learned prior (ablation) | DBAE · FFHQ unconditional generation | FID 50k: 11.25 -> 1.77 | Helped |
| use the inferred latent distribution q_phi(z) instead of the learned prior (ablation) | DBAE · FFHQ unconditional generation | Recall: 0.392 -> 0.687 | Helped |
| use a z-dependent endpoint x_T to make z an information bottleneck | DBAE · FFHQ unconditional generation | FID 50k: 9.14 -> 11.25 | Hurt |
Surprising findings
- Plain DBAE (unconditional generation with learned prior) has worse FID (11.25) than DDPM (9.14) and DiffAE (9.40) on FFHQ, though it wins on precision and IS.
- The stochastic-encoder DBAE beats the deterministic DBAE-d on downstream linear probing, while the deterministic variant wins on reconstruction.
Limitations
- Generation quality depends on a separately trained shallow latent diffusion prior; without it FID is weaker on FFHQ (DBAE FID 11.25 vs DDPM 9.14).
- Paper notes sensitivity to the gap between inferred and prior latent distributions.
- Results on the stated experiments rely on datasets of faces and LSUN, with limited scale evaluation.
Experimental setup
Datasets: FFHQ CelebA CelebA-HQ LFW LSUN Horse LSUN Bedroom
Benchmarks: CelebA FFHQ CelebA-HQ LFW LSUN
Metrics: AP Pearson's r MSE SSIM LPIPS TAD ATTRS FID Precision Inception Score
Baselines: SimCLR beta-TCVAE IB-GAN DiffAE PDAE DiTi DDIM DDPM VAE beta-VAE InfoVAE InfoDiffusion
Builds on: DiffAE PDAE Doob h-transform DDIM InfoDiffusion
Closest papers by meaning
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.88)
- Deconstructing Denoising Diffusion Models for Self-Supervised Learning (ICLR 2025 · similarity 0.87)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.87)
- PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher (NeurIPS 2024 · similarity 0.87)
- Steering Optimisation Trajectories in Diffusion Representation Learning (NeurIPS 2026 · similarity 0.87)
- DiffEnc: Variational Diffusion with a Learned Encoder (ICLR 2024 · 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.