Research map / Image generation / Fast diffusion sampling
Diffusion Bridge Implicit 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.
Introduces diffusion bridge implicit models (DBIMs), a training-free family of non-Markovian samplers for pretrained denoising diffusion bridge models that gives up to 25x faster sampling with deterministic, stochastic and high-order variants.
Problem
DDBMs need hundreds of network evaluations to sample from diffusion bridges between paired endpoints, making image translation and restoration slow.
Approach
Generalizes the DDBM forward bridge to non-Markovian bridges on discrete timesteps controlled by a variance parameter eta that share marginals and training objective, yielding implicit (DDIM-like) deterministic or stochastic samplers, a novel ODE for the bridge, and high-order solvers, with a booting noise in the first step.
Key results
- On ImageNet 256x256 inpainting, DBIM with 20 NFE beats all baselines including DDBM with 500 NFE (FID 4.07 vs 4.27), a ~25x speed-up.
- On DIODE-Outdoor, DBIM with 100 NFE reaches FID 2.57 vs DDBM's 3.34 at 200 NFE.
- DBIM with 100 NFE reaches FID 3.88 on ImageNet inpainting, the first reported FID below 4 in this setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace DDBM sampling with DBIM at NFE 100 | DDBM · DIODE-Outdoor (256×256) | FID: 3.34 -> 2.57 | Helped |
| replace DDBM sampling with DBIM at NFE 20 | DDBM · ImageNet inpainting (256×256) | FID: 4.27 -> 4.07 | Helped |
| increase DBIM sampling from NFE 20 to NFE 100 | DBIM · Edges→Handbags (64×64) | FID: 1.74 -> 0.89 | Helped |
| set η=0 rather than η=1.0 (ablation) | DBIM · Edges→Handbags (64×64), NFE 20 | FID: 2.72 -> 1.76 | Helped |
| set η=0 rather than η=1.0 (ablation) | DBIM · DIODE-Outdoor (256×256), NFE 20 | FID: 8.85 -> 4.97 | Helped |
| increase stochasticity to η=0.8 rather than η=0 (ablation) | DBIM · ImageNet inpainting (256×256), NFE 100 | FID: 3.91 -> 3.88 | Helped |
| use a 2nd-order solver instead of first-order (ablation) | DBIM · Edges→Handbags (64×64), NFE 20 | FID: 1.76 -> 1.48 | Helped |
| use a 3rd-order solver instead of first-order (ablation) | DBIM · DIODE-Outdoor (256×256), NFE 20 | FID: 4.97 -> 4.20 | Helped |
| use a 2nd-order solver instead of first-order (ablation) | DBIM · ImageNet inpainting (256×256), NFE 20 | FID: 4.11 -> 4.07 | Helped |
| increase NFE from 100 to 200 at η=0 (ablation) | DBIM · ImageNet inpainting (256×256) | FID: 3.91 -> 3.91 | No clear effect |
| increase η from 0 to 1.0 (ablation) | DBIM · ImageNet inpainting (256×256), NFE 20 | FID: 4.11 -> 4.51 | Hurt |
Surprising findings
- For image translation, the deterministic sampler (eta=0) outperforms stochastic variants, whereas for diverse inpainting larger stochasticity helps at high NFE.
- High-order solvers give only modest gains in diffusion bridges, smaller than the gains from tuning eta.
Limitations
- Quality is poor at very small NFE and blurry regions remain even with high-order solvers.
- Training-free method cannot exceed the capability of the pretrained diffusion bridge model; fails on faces and hands in inpainting.
- Lags GAN-based methods in one-step generation.
Experimental setup
Datasets: Edges→Handbags DIODE-Outdoor ImageNet
Benchmarks: Edges→Handbags DIODE-Outdoor ImageNet inpainting (center 128x128 mask)
Metrics: FID IS LPIPS MSE classifier accuracy diversity score
Baselines: DDBM DDIB SDEdit Pix2Pix I2SB DDRM ΠGDM DDNM Palette
Builds on: DDBM DDIM DDPM probability flow ODE I2SB
Closest papers by meaning
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.95)
- Consistency Diffusion Bridge Models (NeurIPS 2024 · similarity 0.94)
- Exploring the Design Space of Diffusion Bridge Models (NeurIPS 2025 · similarity 0.92)
- Inverse Bridge Matching Distillation (ICML 2025 · similarity 0.91)
- Directly Denoising Diffusion Models (ICML 2024 · similarity 0.91)
- Learning to Discretize Denoising Diffusion ODEs (ICLR 2025 · similarity 0.90)
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.