Research map / Image generation / Fast diffusion sampling
Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps
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 Time-Shift Sampler, a training-free sampling wrapper that shifts the next diffusion time step within a window to better couple with the predicted state, reducing exposure bias and improving FID for DDPM/DDIM/PNDM-style samplers.
Problem
Exposure bias in diffusion probabilistic models: training uses ground-truth noisy inputs while sampling uses network-predicted states, so errors accumulate over many sampling steps; prior fixes require retraining.
Approach
For each backward step, search a window of time steps around t-1 for a step t_s whose noise level best matches the variance of the current predicted sample (cutoff below which shifting is disabled), giving a training-free plug-in for existing samplers.
Key results
- TS-F-PNDM reaches FID 3.88 on CIFAR-10 with 10 steps (44.49% better than F-PNDM), beating DDIM with 100 steps (4.23).
- TS-DDIM (quadratic) improves CIFAR-10 FID at 100 steps from 4.23 to 3.81 (+9.93%) without retraining.
- With ADM backbone, TS-DDIM gets FID 26.94 at 5 steps vs 28.98 for DDIM, and outperforms retrained ADM-IP at 10 and 20 steps, with negligible extra sampling time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Time-Shift Sampler to DDIM with quadratic time selection | DDIM · CIFAR-10, 10 steps | FID: 13.70 -> 11.93 | Helped |
| Add Time-Shift Sampler to DDIM with uniform time selection | DDIM · CIFAR-10, 10 steps | FID: 18.71 -> 12.21 | Helped |
| Add Time-Shift Sampler to DDPM with uniform time selection | DDPM · CIFAR-10, 10 steps | FID: 42.04 -> 33.36 | Helped |
| Add Time-Shift Sampler to S-PNDM with uniform time selection | S-PNDM · CIFAR-10, 10 steps | FID: 9.49 -> 5.14 | Helped |
| Add Time-Shift Sampler to F-PNDM with uniform time selection | F-PNDM · CIFAR-10, 10 steps | FID: 6.99 -> 3.88 | Helped |
| Add Time-Shift Sampler to DDIM with quadratic time selection | DDIM · CelebA 64×64, 10 steps | FID: 10.93 -> 9.36 | Helped |
| Add Time-Shift Sampler to DDIM with uniform time selection | DDIM · CelebA 64×64, 10 steps | FID: 17.18 -> 10.61 | Helped |
| Add Time-Shift Sampler to DDPM with uniform time selection | DDPM · CelebA 64×64, 10 steps | FID: 34.12 -> 27.17 | Helped |
| Add Time-Shift Sampler to S-PNDM with uniform time selection | S-PNDM · CelebA 64×64, 10 steps | FID: 11.36 -> 10.50 | Helped |
| Add Time-Shift Sampler to F-PNDM with uniform time selection | F-PNDM · CelebA 64×64, 10 steps | FID: 9.23 -> 6.96 | Helped |
| Add Time-Shift Sampler to DDIM | ADM with DDIM · CIFAR-10, 10 steps | FID: 12.11 -> 10.73 | Helped |
| Add Time-Shift Sampler to DDIM | DDIM · LSUN-bedroom 256×256, 10 steps | FID: 16.90 -> 16.66 | Helped |
Surprising findings
- ADM-IP, which retrains with input perturbation, performs much worse than the original ADM at 5 and 10 sampling steps.
- Improvements on high-order solvers (DPM-solver, DEIS) are smaller than for DDIM/DDPM, which the authors attribute to their semi-linear ODE structure.
Limitations
- Introduces two hyperparameters (window size and cutoff value) that are tuned empirically from statistics rather than derived analytically.
- Hyperparameters for LSUN-bedroom were not tuned due to computational limits.
- Theoretical derivation relies on approximations (large t, Gaussian prediction error, high dimensionality).
Experimental setup
Datasets: CIFAR-10 CelebA LSUN-bedroom ImageNet MSCOCO val2017
Benchmarks: CIFAR-10 CelebA LSUN-bedroom ImageNet MSCOCO val2017
Metrics: FID precision recall sampling time
Baselines: DDPM DDIM S-PNDM F-PNDM ADM-IP Analytic-DPM DPM-solver-2 DEIS
Builds on: DDPM DDIM PNDM ADM-IP DPM-solver
Closest papers by meaning
- Elucidating the Exposure Bias in Diffusion Models (ICLR 2024 · similarity 0.92)
- Multi-Step Denoising Scheduled Sampling: Towards Alleviating Exposure Bias for Diffusion Models (AAAI 2024 · similarity 0.92)
- Accelerating Diffusion Sampling with Optimized Time Steps (CVPR 2024 · similarity 0.90)
- Tuning Timestep-Distilled Diffusion Model Using Pairwise Sample Optimization (ICLR 2025 · similarity 0.90)
- T-Stitch: Accelerating Sampling in Pre-Trained Diffusion Models with Trajectory Stitching (ICLR 2025 · similarity 0.89)
- Rethinking Diffusion Posterior Sampling: From Conditional Score Estimator to Maximizing a Posterior (ICLR 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.