Research map / Image generation / Fast diffusion sampling
Improved Noise Schedule for Diffusion Training
Tiankai Hang, Shuyang Gu, Jianmin Bao, Fangyun Wei, Dong Chen, Xin Geng, Baining Guo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace cosine schedule with Laplace noise schedule | DiT · ImageNet-256 (CFG=3.0) | FID-10K: 11.06 -> 7.96 | Improved |
| replace cosine with Laplace schedule for each prediction target | DiT · ImageNet-256 (500K iters) | FID-10K: 11.16 -> 8.48 (x0), 11.06 -> 7.96 (v), 10.46 -> 9.53 (epsilon) | Improved |
| replace cosine with Laplace schedule at 512 resolution | DiT · ImageNet-512 | FID-10K: 11.91 -> 9.09 | Improved |
| replace cosine with Laplace schedule on unconditional faces | DiT · CelebA 64x64 | FID: 7.93795 -> 6.58359 | Improved |
| replace cosine with Laplace schedule for text-to-image | DiT-base · CC12M | FID: 58.3619 -> 54.3492 | Improved |
| change noise schedule instead of loss weighting (Min-SNR, Soft-Min-SNR) (ablation) | DiT · ImageNet-256 (CFG=2.0) | FID-10K: 9.07 -> 7.98 | Improved |
| use cosine-ply importance sampling as a loss weight instead of schedule (ablation) | DiT · ImageNet-256 (CFG=2.0) | FID-10K: 7.98 -> 8.88 | Hurt |
| shift the cosine schedule by mu=1 (ablation) | Cosine schedule · ImageNet-256 | FID-10K: 10.85 -> 11.67 (CFG=2.0) | Hurt |
| shift the cosine schedule by mu=-1 (ablation) | Cosine schedule · ImageNet-256 | FID-10K: 10.85 -> 11.48 (CFG=2.0) | Hurt |
| scale cosine schedule with 1/s=0.5 to concentrate near log SNR 0 (ablation) | Cosine schedule · ImageNet-256 (CFG=2.0) | FID-10K: 10.85 -> 8.04 | Improved |
| reduce 1/s from 0.5 to 0.25 in cosine scaled (ablation) | Cosine Scaled · ImageNet-256 | FID-10K: 12.74 -> 15.83 (CFG=1.5), 11.02 -> 8.26 (CFG=3.0) | Mixed |
| move Cauchy location mu from 0 to -1 or 1 (gamma=1) (ablation) | Cauchy schedule · ImageNet-256 (CFG=2.0) | FID-10K: 8.93 -> 10.38 (mu=-1), 10.19 (mu=1) | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Cosine — COSINE: Community-Preserving Social Network Embedding From Information Diffusion Cascades (AAAI 2018) · better on 9 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Align Your Steps: Optimizing Sampling Schedules in Diffusion Models (ICML 2024 · similarity 0.91)
- Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.91)
- Schedule On the Fly: Diffusion Time Prediction for Faster and Better Image Generation (CVPR 2025 · similarity 0.90)
- One More Step: A Versatile Plug-and-Play Module for Rectifying Diffusion Schedule Flaws and Enhancing Low-Frequency Controls (CVPR 2024 · similarity 0.90)
- Improving Training Efficiency of Diffusion Models via Multi-Stage Framework and Tailored Multi-Decoder Architecture (CVPR 2024 · similarity 0.90)
- Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models (NeurIPS 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.
