Research map / Image generation / Fast diffusion sampling
Towards More Accurate Diffusion Model Acceleration with A Timestep Tuner
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 TimeTuner, a plug-in method that learns a single replacement timestep per sampling interval so the denoiser is conditioned on a better timestep, reducing truncation error in accelerated diffusion sampling.
Problem
Accelerated diffusion samplers that skip most denoising steps suffer large quality degradation because the integral direction from the initial state is inaccurate over a large timestep interval, causing accumulating truncation error.
Approach
Replace the timestep input of the pre-trained noise prediction network with a learned timestep tau_i per interval, optimized sequentially (from T down to 0) to match the sampling distribution to the real one, motivated by a mean-value-theorem view of integration and an error bound; the pre-trained model is not updated.
Key results
- On LSUN Bedroom with DDIM at NFE=10, FID improves from 9.65 to 6.07 (sequential training) with TimeTuner.
- On CIFAR10 with DDIM at NFE=10 (quadratic trajectory), FID drops from 13.65 to 11.77, outperforming OMS-DPM (12.34) and RS-DDIM.
- On LDM ImageNet 256x256 at NFE=10 with DDIM, FID improves from 10.13 to 9.63 and sFID from 12.52 to 7.29.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| tune DDIM input timesteps with TimeTuner | DDIM with LDM · LSUN Bedroom 256x256 | FID: 9.46 -> 5.85 | Helped |
| tune DDIM input timesteps with TimeTuner | DDIM with LDM · FFHQ 256x256 | FID: 23.58 -> 14.80 | Helped |
| tune DDIM input timesteps with TimeTuner | DDIM with LDM · CelebA-HQ 256x256 | FID: 18.72 -> 16.59 | Helped |
| tune DDIM input timesteps with TimeTuner | DDIM with LDM · ImageNet 256x256 | FID: 10.13 -> 9.63 | Helped |
| add TimeTuner to Consistency Distillation | CD · LSUN Bedroom, NFE = 1 | FID: 8.14 -> 7.14 | Helped |
| add TimeTuner to Consistency Distillation | CD · LSUN Bedroom, NFE = 2 | FID: 5.91 -> 5.18 | Helped |
| add TimeTuner to DDIM | DDIM · CIFAR10, NFE = 10, linear trajectory | FID: 13.65 -> 11.77 | Helped |
| train timestep values in parallel instead of sequentially (ablation) | DDIM + TimeTuner · CIFAR10 | FID: 11.77 -> 11.75 | No clear effect |
| train timestep values in parallel instead of sequentially (ablation) | DDIM + TimeTuner · CelebA | FID: 11.84 -> 12.14 | Hurt |
| train timestep values in parallel instead of sequentially (ablation) | DDIM + TimeTuner · LSUN Bedroom | FID: 6.07 -> 6.08 | No clear effect |
| train timestep values in parallel instead of sequentially (ablation) | DDIM + TimeTuner · FFHQ | FID: 14.80 -> 15.03 | Hurt |
| train timestep values in parallel instead of sequentially (ablation) | DDIM + TimeTuner · CelebA-HQ | FID: 17.75 -> 18.03 | Hurt |
Surprising findings
- Parallel training of timesteps, which is much cheaper, yields nearly the same FID as sequential training on most datasets, though it is slightly worse on some.
Limitations
- Parallel training strategy introduces extra error and is somewhat worse than sequential training.
- Does not find a globally optimal schedule; acts as a plug-in improvement over existing trajectories.
- Improvements shrink as the number of function evaluations grows.
Experimental setup
Datasets: CIFAR10 CelebA LSUN Bedroom FFHQ CelebA-HQ ImageNet MS-COCO
Benchmarks: CIFAR10 CelebA LSUN Bedroom FFHQ CelebA-HQ ImageNet MS-COCO
Metrics: FID IS sFID Precision Recall
Baselines: DDIM DDPM Analytic-DDIM Analytic-DDPM DPM-Solver-2 EDM Consistency Distillation GGDM RS-DDIM OMS-DPM
Builds on: DDIM DDPM DPM-Solver Analytic-DPM EDM
Closest papers by meaning
- Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training (CVPR 2025 · similarity 0.90)
- TimeStep Master: Asymmetrical Mixture of Timestep LoRA Experts for Versatile and Efficient Diffusion Models in Vision (ICML 2025 · similarity 0.88)
- Boosting Diffusion Models with Moving Average Sampling in Frequency Domain (CVPR 2024 · similarity 0.87)
- Improving Training Efficiency of Diffusion Models via Multi-Stage Framework and Tailored Multi-Decoder Architecture (CVPR 2024 · similarity 0.87)
- PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores (ICLR 2025 · similarity 0.87)
- Accelerating Convergence of Score-Based Diffusion Models, Provably (ICML 2024 · similarity 0.87)
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.