Research map / Image generation / Fast diffusion sampling
Denoising Diffusion Bridge Models
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 Denoising Diffusion Bridge Models (DDBMs), which learn the score of a diffusion bridge between paired endpoint distributions to translate between images, and reduce to standard diffusion models when one endpoint is Gaussian.
Problem
Standard diffusion models map noise to data and cannot directly translate between two arbitrary paired distributions (e.g., image-to-image translation), forcing cumbersome guidance or projected sampling; existing bridge/transport methods are either empirically weak or require expensive iterative tra
Approach
Learn the score of a diffusion bridge conditioned on both endpoints by denoising score matching against the closed-form Gaussian bridge conditional score, then sample via a reverse SDE or probability flow ODE with a hybrid Euler-Maruyama/Heun sampler; a generalized EDM-style preconditioning is derived for the bridge setting.
Key results
- On Edges→Handbags 64x64, DDBM (VP) reaches FID 1.83 vs 7.43 for I2SB and 25.3 for Rectified Flow at 40 NFE.
- On DIODE 256x256, DDBM (VP) reaches FID 4.43 vs 9.34 for I2SB, with lower LPIPS (VE 0.226 vs I2SB 0.373).
- On unconditional generation the method is comparable to EDM, with negligible FID degradation on CIFAR-10 and marginal improvement on FFHQ-64x64.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DDBM (VP) instead of the strongest baseline | Pixel-space image translation · Edges→Handbags-64×64 | FID: 7.43 -> 1.83 | Helped |
| Use DDBM (VE) instead of the strongest baseline | Pixel-space image translation · Edges→Handbags-64×64 | LPIPS: 0.241 -> 0.131 | Helped |
| Use DDBM (VE) instead of the strongest baseline | Pixel-space image translation · Edges→Handbags-64×64 | MSE: 0.088 -> 0.013 | Helped |
| Use DDBM (VP) instead of the strongest baseline | Pixel-space image translation · DIODE-256×256 | FID: 9.34 -> 4.43 | Helped |
| Use DDBM (VE) instead of the strongest baseline | Pixel-space image translation · DIODE-256×256 | LPIPS: 0.373 -> 0.226 | Helped |
| Use DDBM (VE) instead of the strongest baseline | Pixel-space image translation · DIODE-256×256 | MSE: 0.145 -> 0.0107 | Helped |
| Increase Euler step ratio s from 0 (ODE-only) toward a tuned value (ablation) | DDBM hybrid sampler · Edges→Handbags and DIODE | FID | Helped |
| Vary guidance strength w (ablation) | DDBM (VE) · Edges→Handbags and DIODE | FID | No clear effect |
| Set guidance strength w to 1 (ablation) | DDBM (VP) · Edges→Handbags and DIODE | FID | Helped |
| Add the proposed preconditioning (ablation) | DDBM · Image translation | generation performance | Helped |
| Use the hybrid sampler instead of the ODE sampler (s=0) (ablation) | DDBM · Edges→Handbags | image sharpness | Helped |
Surprising findings
- VP bridges need the guidance strength w=1 to perform well, while VE bridges are insensitive to w.
- The probability flow ODE alone produces blurry outputs for translation, so injecting noise via the hybrid sampler is necessary.
Limitations
- Unconditional results only match EDM rather than surpassing it
- VP bridges depend on guidance strength w=1 and a specific symmetric noise schedule
- Latent-space results are described but limited in reported detail
Experimental setup
Datasets: Edges→Handbags DIODE-Outdoor CIFAR-10 FFHQ-64x64
Benchmarks: Edges→Handbags DIODE-Outdoor CIFAR-10 FFHQ-64x64
Metrics: FID IS LPIPS MSE NFE
Baselines: Pix2Pix SDEdit DDIB Rectified Flow I2SB DDPM DDIM EDM NCSN++ DDPM++
Builds on: EDM Doob h-transform score-based SDE Rectified Flow I2SB
Closest papers by meaning
- Diffusion Bridge Implicit Models (ICLR 2025 · similarity 0.95)
- Consistency Diffusion Bridge Models (NeurIPS 2024 · similarity 0.94)
- Assessing the quality of denoising diffusion models in Wasserstein distance: noisy score and optimal bounds (NeurIPS 2025 · similarity 0.91)
- Training Diffusion-based Generative Models with Limited Data (ICML 2025 · similarity 0.91)
- Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models (NeurIPS 2024 · similarity 0.91)
- Exploring the Design Space of Diffusion Bridge Models (NeurIPS 2025 · similarity 0.91)
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.