Research map / Image generation / Fast diffusion sampling
Exploring the Design Space of 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.
Extends stochastic interpolants with endpoint conditioning, preconditioning, a noise-controlled sampler, and a modified base distribution to improve image-to-image bridge models and conditional diversity.
Problem
Diffusion bridge models for image-to-image translation have restricted path design, coupled parameters, and complicated tuning, while stochastic interpolants lack practical refinements; one-to-many translation also suffers from low conditional output diversity.
Approach
Endpoint-Conditioned Stochastic Interpolants decouple bridge-path and sampler parameters (alpha, beta, gamma, epsilon), unify DDBM/DBIM/I2SB samplers as special cases, and perturb the base distribution with Gaussian noise to raise conditional diversity.
Key results
- On Edges2handbags (64x64), ECSI (gamma_max=0.125) reaches FID 0.56 at 20 NFE versus 1.74 for DBIM at 20 NFE.
- On DIODE-Outdoor (256x256), ECSI (gamma_max=0.25) reaches FID 3.27 at 20 NFE versus 4.99 for DBIM and 4.43 for DDBM.
- Modifying the base distribution raises AFD on Edges2handbags from 6.25 (ECSI) to 9.65 at NFE=20, at the cost of a slightly higher FID (1.74).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ECSI sampling instead of DBIM sampling | Pretrained DDBM VP · Edges→Handbags (64×64), NFE=5 | FID: 3.60 -> 2.36 | Helped |
| Use ECSI sampling instead of DBIM sampling | Pretrained DDBM VP · DIODE-Outdoor (256×256), NFE=5 | FID: 14.25 -> 10.87 | Helped |
| Use ECSI rather than DBIM | Edges→Handbags image translation · Edges→Handbags (64×64), NFE=20 | FID: 1.74 -> 0.56 | Helped |
| Use ECSI rather than DBIM | DIODE-Outdoor image translation · DIODE-Outdoor (256×256), NFE=20 | FID: 4.99 -> 3.27 | Helped |
| Use ECSI for deblurring | ImageNet deblurring · Uniform kernel | FID: DDRM 9.9; DDNM 3.0; Pallette 4.1; CDSB 15.5; I²SB 3.9; ECSI 1.11 | Helped |
| Use ECSI for deblurring | ImageNet deblurring · Gaussian kernel | FID: DDRM 6.1; DDNM 2.9; Pallette 3.1; CDSB 7.7; I²SB 3.0; ECSI 0.41 | Helped |
| Add noise to the base distribution (ablation) | ECSI checkpoint + sampler · Edges2Handbags (64×64), NFE=5 | AFD: 6.00 -> 8.53 | Helped |
| Add noise to the base distribution (ablation) | ECSI checkpoint + sampler · Edges2Handbags (64×64), NFE=5 | FID: 0.89 -> 3.31 | Hurt |
| Replace the DDBM checkpoint with an ECSI checkpoint while using ECSI sampling (ablation) | Edges2Handbags (64×64), NFE=5 · Edges2Handbags (64×64) | FID: 2.36 -> 0.89 | Helped |
| Set εt=0 for the last two sampling steps and use the endpoint update (ablation) | ECSI sampler · Image-to-image translation | FID | Helped |
| Vary γmax in the linear bridge path (ablation) | ECSI · DIODE (64×64) | FID | Mixed |
| Vary η in the sampling noise control (ablation) | ECSI · DIODE (64×64) | FID | Mixed |
Surprising findings
- Modifying the base distribution with added noise, which by data processing should only lose information, substantially improves conditional diversity.
- FID can be nearly identical across pseudo-generative models with very different conditional diversity, so FID fails to capture conditional diversity.
Limitations
- Optimal path design may vary by task.
- Guidance techniques are not incorporated and could further improve performance.
- Modifying the base distribution slows sampling and slightly increases FID.
Experimental setup
Datasets: Edges2handbags DIODE-Outdoor ImageNet
Benchmarks: Edges2handbags 64x64 DIODE-Outdoor 256x256 ImageNet deblurring 256x256
Metrics: FID IS LPIPS MSE AFD
Baselines: Pix2Pix DDIB SDEdit Rectified Flow I2SB DDBM DBIM DDRM DDNM Palette CDSB
Builds on: DDBM DBIM I2SB Stochastic Interpolants EDM
Closest papers by meaning
- UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control (ICML 2025 · similarity 0.93)
- Diffusion Bridge Implicit Models (ICLR 2025 · similarity 0.92)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.91)
- Solving Diffusion ODEs with Optimal Boundary Conditions for Better Image Super-Resolution (ICLR 2024 · similarity 0.90)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.89)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.89)
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.