Research map / Image generation / Fast diffusion sampling
On Inference Stability 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 a sequence-aware (SA) loss that jointly penalizes noise-prediction errors over K consecutive diffusion timesteps to reduce cumulative estimation gap, improving FID/IS of DDPM/DDIM and covariance-estimation variants on CIFAR10, CelebA, and CelebA-HQ.
Problem
Standard DPM training minimizes per-timestep noise-prediction MSE and ignores timestep correlations, so errors accumulate along the sampling trajectory, degrading sample quality especially at few sampling steps.
Approach
Add a sequence-aware loss that penalizes the norm of the averaged weighted noise-prediction error over K consecutive timesteps, combined with the simple loss; the authors prove it is a tighter upper bound of the total trajectory estimation gap than the weighted conventional loss.
Key results
- SA-2-DPM at T=10 on CIFAR10 (LS) reduces FID from 41.41 (DDPM) to 30.51.
- On CelebA-HQ 256x256 with DDPM at T=10, FID improves from 54.19 to 53.23; SA+DDIM at T=100 gives 18.98 vs 22.19 baseline.
- Combined with covariance-estimation methods, SA variants further lower FID (e.g., CIFAR10 LS T=10: SA+A 21.66 vs B+A 34.19); total gap at final step is reduced about 2.5x for SA-2-DPM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add sequence-aware loss | DDPM · CIFAR10 32×32, linear schedule, T=10 | FID: 41.41 -> 30.51 | Helped |
| Add sequence-aware loss | DDPM · CIFAR10 32×32, linear schedule, T=1000 | FID: 5.92 -> 4.33 | Helped |
| Add sequence-aware loss | DDPM · CIFAR10 32×32, cosine schedule, T=10 | FID: 34.98 -> 24.59 | Helped |
| Add sequence-aware loss | DDPM · CelebA 64×64, DDPM, T=10 | FID: 36.69 -> 32.15 | Helped |
| Add sequence-aware loss | DDIM · CelebA 64×64, T=10 | FID: 20.54 -> 12.88 | Helped |
| Add sequence-aware loss | DDPM · CIFAR10 32×32, linear schedule, T=10 | Inception Score: 6.93 -> 7.55 | Helped |
| Use λ=2.0 for SA-2-DPM (ablation) | SA-2-DPM · CIFAR10 32×32, T=10 | FID: 41.41 -> 19.14 | Helped |
| Use λ=2.0 for SA-2-DPM (ablation) | SA-2-DPM · CIFAR10 32×32, T=1000 | FID: 5.92 -> 14.20 | Hurt |
| Use K=3 and λ=1.5 (ablation) | SA-3-DPM · CIFAR10 32×32, T=10 | FID: 41.41 -> 15.59 | Helped |
| Use K=3 and λ=1.5 (ablation) | SA-3-DPM · CIFAR10 32×32, T=1000 | FID: 5.92 -> 19.49 | Hurt |
| Use K=4 and λ=0.4 (ablation) | SA-4-DPM · CIFAR10 32×32, T=10 | FID: 41.41 -> 26.68 | Helped |
| Use K=4 and λ=0.4 (ablation) | SA-4-DPM · CIFAR10 32×32, T=1000 | FID: 5.92 -> 6.95 | Hurt |
Surprising findings
- Setting lambda too high (e.g., 2.0 for SA-2-DPM) degrades FID at large timestep counts (T=1000 FID 14.20 vs 5.92 baseline), even as it helps at small T.
- Increasing the number of consecutive steps K does not much change the total gap or model performance beyond K=2 in the appendix analysis.
Limitations
- The new loss requires network evaluations at multiple timesteps, making training slower than the vanilla loss.
- Performance is sensitive to the weight lambda, with large values hurting high-step sampling quality.
Experimental setup
Datasets: CIFAR10 CelebA CelebA-HQ
Benchmarks: CIFAR10 CelebA 64x64 CelebA-HQ 256x256
Metrics: FID Inception Score estimation gap
Baselines: DDPM DDIM Analytic-DPM NPR-DPM SN-DPM
Builds on: DDPM DDIM Analytic-DPM NPR-DPM SN-DPM
Closest papers by meaning
- Assessing the quality of denoising diffusion models in Wasserstein distance: noisy score and optimal bounds (NeurIPS 2025 · similarity 0.91)
- Learning to Discretize Denoising Diffusion ODEs (ICLR 2025 · similarity 0.91)
- D^2-DPM: Dual Denoising for Quantized Diffusion Probabilistic Models (AAAI 2025 · similarity 0.90)
- O(d/T) Convergence Theory for Diffusion Probabilistic Models under Minimal Assumptions (ICLR 2025 · similarity 0.90)
- EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.90)
- Improved Diffusion-based Generative Model with Better Adversarial Robustness (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.