Research map / Image generation / Fast diffusion sampling
Multi-Step Denoising Scheduled Sampling: Towards Alleviating Exposure Bias for Diffusion 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 multi-step denoising scheduled sampling (MDSS), a fine-tuning method that reduces exposure bias in DDPM/DDIM/DPM-Solver by modeling multi-step prediction error accumulation and denoising the input signal.
Problem
Exposure bias in diffusion models: training feeds ground-truth-noised inputs while inference feeds model-generated inputs, and prior fixes (input perturbation, scheduled sampling) model only one-step errors partially.
Approach
Model the accumulated prediction error with a closed-form multi-step posterior formulation initialized from a one-step model prediction error, and train with a noise-free objective plus an extra denoising channel that removes the input-signal error; applied as a short fine-tune of a pretrained DDPM.
Key results
- MDSS achieves FID 3.86 with 100 DDIM steps on CIFAR-10, versus 4.78 for the second best (SS).
- With DDPM at 30 steps on CIFAR-10, MDSS reaches FID 5.54 versus 7.81 for the DDPM baseline (multi-step training k=4 on DDPM).
- Fine-tuning with MDSS converges in roughly 20,000 iterations, about a 10x reduction versus training from scratch (~200,000 iterations).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MDSS instead of the strongest listed baseline | iDDPM · CIFAR-10 DDPM, 30–250 steps | FID: 5.54 vs 6.40; 3.49 vs 3.25; 3.62 vs 3.02 | Mixed |
| Use MDSS instead of the strongest listed baseline | iDDPM · ImageNet 64×64 DDPM, 30–250 steps | FID: 31.76 vs 32.32; 24.28 vs 23.22; 23.69 vs 22.48 | Mixed |
| Use MDSS instead of the strongest listed baseline | ADM · LSUN tower 64×64 DDPM, 30–250 steps | FID: 5.57 vs 5.52; 2.26 vs 2.24; 1.92 vs 1.90 | Mixed |
| Use MDSS instead of the strongest listed baseline | iDDPM · CIFAR-10 DDIM, 30–250 steps | FID: 5.25 vs 5.35; 3.86 vs 4.78; 3.92 vs 3.94 | Helped |
| Use MDSS instead of the strongest listed baseline | iDDPM · ImageNet 64×64 DDIM, 30–250 steps | FID: 27.18 vs 27.47; 24.96 vs 24.75; 23.86 vs 23.44 | Mixed |
| Use MDSS instead of the strongest listed baseline | ADM · LSUN tower 64×64 DDIM, 30–250 steps | FID: 3.84 vs 3.94; 2.67 vs 2.61; 4.13 vs 4.38 | Mixed |
| Add input-signal denoising to SDSS (ablation) | iDDPM · CIFAR-10 DDPM, 30–250 steps | FID: 7.93 -> 7.92; 3.55 -> 3.47; 3.09 -> 3.08 | Helped |
| Add input-signal denoising to SDSS (ablation) | iDDPM · CIFAR-10 DDIM, 30–250 steps | FID: 6.93 -> 6.92; 4.60 -> 4.53; 4.05 -> 3.94 | Helped |
| Use 4-step multi-step training instead of single-step training (ablation) | iDDPM with MDSS · CIFAR-10 DDPM, 30–250 steps | FID: 7.92 -> 5.54; 3.47 -> 3.49; 3.08 -> 3.62 | Mixed |
| Use 10-step multi-step training instead of single-step training (ablation) | iDDPM with MDSS · CIFAR-10 DDPM, 30–250 steps | FID: 7.92 -> 5.92; 3.47 -> 4.20; 3.08 -> 3.75 | Mixed |
| Use 20-step multi-step training instead of single-step training (ablation) | iDDPM with MDSS · CIFAR-10 DDPM, 30–250 steps | FID: 7.92 -> 6.70; 3.47 -> 3.64; 3.08 -> 3.58 | Mixed |
| Use 50-step multi-step training instead of single-step training (ablation) | iDDPM with MDSS · CIFAR-10 DDPM, 30–250 steps | FID: 7.92 -> 8.31; 3.47 -> 4.12; 3.08 -> 3.32 | Mixed |
Surprising findings
- Input Perturbation performs poorly with DDIM (e.g., FID 8.60 at 250 steps on CIFAR-10 vs 4.48 for DDPM baseline), suggesting predefined Gaussian perturbation mismatches real inference errors.
- MDSS gives little or no improvement over the DDPM baseline for DPM-Solver-2 and -3 on CIFAR-10.
Limitations
- Does not compatibly address multi-evaluation solvers such as DPM-Solver-2/3; gains vanish or reverse there.
- Assumes model prediction noise is additive; other noise types are left for future work.
- Choice of number of multi-step accumulation steps k needs careful tuning, as too many steps can worsen results.
Experimental setup
Datasets: CIFAR-10 ImageNet 64x64 LSUN tower 64x64
Benchmarks: CIFAR-10 ImageNet 64x64 LSUN tower 64x64
Metrics: FID CLIP-FID MSE between model prediction and ground truth
Baselines: DDPM DDIM DPM-Solver Input Perturbation (IP) Scheduled Sampling (SS)
Builds on: DDPM DDIM DPM-Solver scheduled sampling input perturbation
Closest papers by meaning
- Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps (ICLR 2024 · similarity 0.92)
- Denoising Diffusion Step-aware Models (ICLR 2024 · similarity 0.92)
- Score-Optimal Diffusion Schedules (NeurIPS 2024 · similarity 0.91)
- UDPM: Upsampling Diffusion Probabilistic Models (NeurIPS 2024 · similarity 0.90)
- Learning to Discretize Denoising Diffusion ODEs (ICLR 2025 · similarity 0.90)
- D^2-DPM: Dual Denoising for Quantized Diffusion Probabilistic Models (AAAI 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.