Research map / Image generation / Fast diffusion sampling
Consistency 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 consistency diffusion bridge models (CDBM) that learn the PF-ODE consistency function of diffusion denoising bridge models via distillation (CBD) or training (CBT), enabling 4x-50x faster sampling than DDBM.
Problem
DDBM sampling requires hundreds of network evaluations, limiting practical deployment for coupled-data generation tasks such as image-to-image translation and inpainting.
Approach
Extends consistency models to diffusion denoising bridges by learning the consistency function of the bridge PF-ODE, using a unified noise-schedule/parameterization view and a general-form first-order exponential-integrator bridge ODE solver, with both a distillation (CBD) and a training (CBT) paradigm that connects them.
Key results
- On Edges→Handbags (64x64), CBT with NFE=2 reaches FID 0.80, better than DDBM at NFE=100 (FID 0.89) with ODE-1 solver.
- On DIODE-Outdoor (256x256), CBT at NFE=2 achieves FID 2.93 vs. DDBM ODE-1 at NFE=100 with FID 2.57 (close), and CBD achieves FID 3.66.
- Fine-tuning DDBM with CBT generally outperforms CBD across the three datasets; CBT at NFE=4 matches DDBM at NFE=10 on ImageNet inpainting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CBD with two function evaluations instead of DDBM ODE-1 with 100 evaluations | DDBM on Edges→Handbags · Edges→Handbags (64 × 64) | FID: 0.89 -> 1.30 | Hurt |
| Use CBT with two function evaluations instead of DDBM ODE-1 with 100 evaluations | DDBM on Edges→Handbags · Edges→Handbags (64 × 64) | FID: 0.89 -> 0.80 | Helped |
| Use CBD with two function evaluations instead of DDBM ODE-1 with 100 evaluations | DDBM on DIODE-Outdoor · DIODE-Outdoor (256 × 256) | FID: 2.57 -> 3.66 | Hurt |
| Use CBT with two function evaluations instead of DDBM ODE-1 with 100 evaluations | DDBM on DIODE-Outdoor · DIODE-Outdoor (256 × 256) | FID: 2.57 -> 2.93 | Hurt |
| Fine-tune the pretrained DDBM with CBT rather than distill it with CBD | CDBM · Edges→Handbags (64 × 64) | FID: 1.30 -> 0.80 | Helped |
| Fine-tune the pretrained DDBM with CBT rather than distill it with CBD | CDBM · DIODE-Outdoor (256 × 256) | FID: 3.66 -> 2.93 | Helped |
Surprising findings
- Fine-tuning a pretrained bridge model with consistency bridge training outperforms distillation with a pretrained score model in all three datasets.
- The original DDBM bridge ODE has a singularity at the starting point x_T, so the ODE must start from T-gamma after a stochastic step.
Limitations
- Experiments limited to relatively small datasets; performance on larger or more complex datasets unexplored.
- Consistency model training suffers from numerical instability.
- CBD requires careful tuning of timestep interval Δt or b to ensure stable training.
Experimental setup
Datasets: Edges→Handbags DIODE-Outdoor ImageNet
Benchmarks: Edges→Handbags DIODE-Outdoor ImageNet 256x256 inpainting
Metrics: FID IS LPIPS MSE classifier accuracy
Baselines: Pix2Pix DDIB SDEdit Rectified Flow I2SB DDBM
Builds on: consistency models denoising diffusion bridge models EDM Bridge-TTS I2SB DBIM
Closest papers by meaning
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.94)
- Diffusion Bridge Implicit Models (ICLR 2025 · similarity 0.94)
- Truncated Consistency Models (ICLR 2025 · similarity 0.91)
- SCott: Accelerating Diffusion Models with Stochastic Consistency Distillation (AAAI 2025 · similarity 0.90)
- Simplifying, Stabilizing and Scaling Continuous-time Consistency Models (ICLR 2025 · similarity 0.89)
- Consistency Trajectory Models: Learning Probability Flow ODE Trajectory of Diffusion (ICLR 2024 · 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.