Research map / Image generation / Fast diffusion sampling
Diffusion Rejection Sampling
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 Diffusion Rejection Sampling (DiffRS), which uses a time-dependent discriminator to estimate per-timestep density ratios and applies sequential rejection sampling with noise re-initialization to refine samples from a fixed pre-trained diffusion model.
Problem
Pre-trained diffusion models produce sampling error because their learned reverse transition kernels differ from the true ones; improving sample quality usually requires more steps or retraining the diffusion model.
Approach
Applies rejection sampling to each reverse transition kernel of a fixed pre-trained diffusion model, estimating the density-ratio acceptance probability with a time-dependent discriminator, and re-initializes rejected samples by injecting forward noise adaptively rather than restarting from the prior.
Key results
- Unconditional CIFAR-10 FID 1.59 (64 NFE) and class-conditional 1.52 (88 NFE) with EDM, new SOTA among compared methods.
- Class-conditional ImageNet 64x64 FID 1.26 with 273.93 NFE, close to the best reported 1.23 (RIN, 1000 NFE).
- ImageNet 256x256 with DiT-XL/2: DG+DiffRS FID 1.76 vs DG 1.83 at 375 NFE and DDPM 2.30 at 250 NFE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DiffRS to EDM (Heun) | EDM on CIFAR-10 unconditional · CIFAR-10 unconditional | FID: 2.01 -> 1.59 | Helped |
| Apply DiffRS to EDM (Heun) | EDM on CIFAR-10 conditional · CIFAR-10 conditional | FID: 1.83 -> 1.52 | Helped |
| Apply DiffRS to EDM | EDM on ImageNet 64×64 class-conditional · ImageNet 64×64 class-conditional | FID: 2.18 -> 1.26 | Helped |
| Apply DiffRS to DG | DG on ImageNet 256×256 class-conditional · ImageNet 256×256 class-conditional | FID: 1.83 -> 1.76 | Helped |
| Add sequential rejection sampling over transition kernels (ablation) | DiffRS on CIFAR-10 unconditional · CIFAR-10 unconditional | FID: 3.73 -> 1.59 | Helped |
| Use transition-kernel ratios instead of marginal probabilities for sequential rejection sampling (ablation) | DiffRS on CIFAR-10 unconditional · CIFAR-10 unconditional | FID: 1.66 -> 1.59 | Helped |
| Use the proposed re-initialization with additional rejection after the forward step (ablation) | DiffRS on CIFAR-10 unconditional · CIFAR-10 unconditional | FID: 1.84 -> 1.59 | Helped |
| Use the proposed re-initialization instead of restarting from the prior (ablation) | DiffRS on CIFAR-10 unconditional · CIFAR-10 unconditional | NFE: 138.07 -> 64.06 | Helped |
| Increase rejection percentile γ (ablation) | DiffRS on CIFAR-10 unconditional · CIFAR-10 unconditional | NFE: 41.37 -> 91.17 | Hurt |
| Increase rejection percentile γ beyond the usual range (ablation) | DiffRS on CIFAR-10 unconditional · CIFAR-10 unconditional | FID: 1.60 -> 1.79 | Hurt |
| Apply DiffRS to DPM-Solver++ | DPM-Solver++ on CIFAR-10 unconditional · CIFAR-10 unconditional | FID: 3.08 -> 2.41 | Helped |
| Apply DiffRS to consistency distillation | Consistency Distillation on ImageNet 64×64 · ImageNet 64×64 class-conditional | FID: 3.30 -> 2.95 | Helped |
Surprising findings
- Naively applying marginal (not transition-level) rejection sampling degrades FID (1.66 vs 1.59 for full DiffRS), and applying rejection only at t=0 yields FID 3.73 with ~295 NFE.
Limitations
- NFE increases with rejections; FID can worsen when rejection percentile gamma is too large (NFE beyond 2-3x base).
- Rejection constant M_t is estimated empirically from percentiles rather than exactly; discriminator overfitting with few training samples harms density-ratio estimation.
- Gains on large-scale text-to-image are reported with a modest number of images (5,000 COCO) and only on a single base sampler (DDIM 100 NFE).
Experimental setup
Datasets: CIFAR-10 ImageNet 64x64 ImageNet 256x256 FFHQ AFHQv2 COCO
Benchmarks: CIFAR-10 ImageNet 64x64 ImageNet 256x256 FFHQ AFHQv2 COCO
Metrics: FID NFE sFID Inception Score Precision Recall F1 CLIP score
Baselines: EDM (Heun) EDM+DG EDM+Restart DDPM DDIM DG Restart DPM-Solver++ Consistency Distillation RIN CDM ADM
Builds on: DG (Discriminator Guidance) Restart sampling EDM Consistency Models DPM-Solver++
Closest papers by meaning
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.90)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.89)
- Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps (ICLR 2024 · similarity 0.89)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.89)
- Multistep Distillation of Diffusion Models via Moment Matching (NeurIPS 2024 · similarity 0.88)
- Align Your Steps: Optimizing Sampling Schedules in Diffusion Models (ICML 2024 · similarity 0.88)
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.