Research map / Image generation / Fast diffusion sampling
RaSS: Improving Denoising Diffusion Samplers with Reinforced Active Sampling Scheduler
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 RaSS, an RL-trained lightweight step scheduler that adaptively chooses per-instance sampling strides for pre-trained diffusion samplers (DDIM, DPM-Solver, DPM-Solver++), improving quality and speed.
Problem
Existing sampling schedulers for diffusion models use fixed or heuristic step schedules that ignore the instance-specific sampling state, yielding suboptimal quality-efficiency trade-offs.
Approach
An RL agent monitors each generated instance (prompt, current image, timestep) and selects the next step stride from a discrete action set; training uses a long-term perception mechanism with a historical network and a loss-aware evaluator reward balancing PSNR and step stride, with the timeline split into five stages.
Key results
- On ImageNet 64x64 with ADM-G, RaSS (α=2.5, 4.8 steps, 1.25x speed) reaches FID 11.02 vs. 23.71 for DDIM at 6 steps.
- On Stable Diffusion MS-COCO text-to-image, RaSS (α=1.5, ~13 steps, 1.89x speed) matches DDIM-25 quality (FID 24.27 vs 24.26, CLIP 0.2631 vs 0.2634).
- On CIFAR-10 with DPM-Solver, α=1.5 gives FID 4.51 at 21 steps (1.35x speed) vs. the baseline's higher step count.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply RaSS to DDIM on ImageNet 64 × 64 at 6 steps | ADM-G · ImageNet 64 × 64 | FID: 23.71 -> 11.02 | Helped |
| Apply RaSS to DDIM on ImageNet 64 × 64 at 10 steps | ADM-G · ImageNet 64 × 64 | FID: 8.86 -> 6.22 | Helped |
| Apply RaSS to DDIM on ImageNet 64 × 64 at 15 steps | ADM-G · ImageNet 64 × 64 | FID: 5.38 -> 4.85 | Helped |
| Apply RaSS to DDIM at 10 steps | ADM on LSUN-Bedroom · LSUN-Bedroom | FID: 10.01 -> 8.62 | Helped |
| Apply RaSS to DDIM at 10 steps | ADM on LSUN-Cat · LSUN-Cat | FID: 20.05 -> 17.23 | Helped |
| Apply RaSS to DDIM with 25-step baseline | Stable Diffusion on MS-COCO 2017 validation · MS-COCO 2017 validation | FID: 24.26 -> 24.27 | Hurt |
| Apply RaSS to DPM-solver++ with 25-step baseline | Stable Diffusion on MS-COCO 2017 validation · MS-COCO 2017 validation | FID: 17.71 -> 17.45 | Helped |
| Increase α from 0.5 to 1.5 (ablation) | DDIM+RaSS with DDPM · CIFAR-10 | FID: 6.96 -> 10.43 | Hurt |
| Increase α from 0.5 to 1.5 (ablation) | DPM-Solver+RaSS with DDPM · CIFAR-10 | FID: 3.45 -> 4.51 | Hurt |
| Increase α from 0.5 to 1.5 (ablation) | DDIM+RaSS with DDPM · LSUN-Bedroom | FID: 8.75 -> 9.86 | Hurt |
| Increase α from 0.5 to 1.5 (ablation) | DPM-Solver+RaSS with DDPM · LSUN-Bedroom | FID: 2.41 -> 3.23 | Hurt |
| Add LTP while leaving out DSP (ablation) | RaSS with DDPM · CIFAR-10 | FID: 8.33 -> 8.25 | Helped |
Surprising findings
- Text-to-image RaSS with half the steps achieves image quality matching or slightly exceeding the 25-step DDIM and DPM-Solver++ baselines.
- Instance-specific random step sequences showed the highest achievable FID potential, even though average non-uniform schedules were less stable.
Limitations
- Gains are reported as modest or mixed in some settings (e.g., some FID values comparable to baselines); quality-efficiency trade-off controlled by hand-tuned α.
- Evaluation focuses on FID/IS/CLIP without human study; hardware and compute details are not stated.
Experimental setup
Datasets: CIFAR-10 LSUN-Bedroom LSUN-Cat ImageNet 64x64 MS-COCO 2017
Benchmarks: CIFAR-10 LSUN-Bedroom LSUN-Cat ImageNet 64x64 MS-COCO 2017
Metrics: FID IS CLIP score speed (relative) MACs parameters
Baselines: DDIM AutoDiff DPM-Solver++ Adadiff
Builds on: DDIM DPM-Solver DPM-Solver++ AutoDiffusion Adadiff
Closest papers by meaning
- Learning to Sample From Diffusion Models via Inverse Reinforcement Learning (NeurIPS 2026 · similarity 0.89)
- Align Your Steps: Optimizing Sampling Schedules in Diffusion Models (ICML 2024 · similarity 0.88)
- Score-Optimal Diffusion Schedules (NeurIPS 2024 · similarity 0.88)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.88)
- Multi-Step Denoising Scheduled Sampling: Towards Alleviating Exposure Bias for Diffusion Models (AAAI 2024 · similarity 0.87)
- Interpreting and Improving Diffusion Models from an Optimization Perspective (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.