Research map / Image generation / Fast diffusion sampling
Boost-and-Skip: A Simple Guidance-Free Diffusion for Minority Generation
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 Boost-and-Skip, a guidance-free minority-sample sampler for diffusion models that boosts the variance of the initial noise and skips the earliest timesteps, matching guided minority samplers at much lower compute.
Problem
Generating minority samples (low-density instances on the data manifold) with diffusion models, where existing high-performing methods depend on computationally expensive guidance such as classifier guidance or backpropagation-based low-density guidance.
Approach
Replace standard N(0,I) initialization with variance-boosted N(0,γ²I) (γ>1) and start the reverse SDE at a skipped timestep T_skip = T − Δ so that the boosted initial noise retains non-negligible signal weight; stochastic contraction of the reverse SDE then corrects early errors, which PF-ODE cannot do.
Key results
- On ImageNet 64x64, Boost-and-Skip reaches cFID 12.68 and sFID 3.18 versus 11.24 to 12.47 for the guidance-based minority samplers, with ~0.98 s/sample inference versus 2.27 to 2.80 s.
- On CelebA 64x64 it achieves the best sFID (6.87) and recall (0.51) among compared methods, with cFID 19.79 close to the best 19.34.
- Compared with the guidance-based sampler on ImageNet 64x64, the paper reports 65% less wall-clock time and 4.5x lower peak memory, and it avoids the more than 16 days of extra classifier training that guided methods need.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Boost-and-Skip instead of ADM | ADM · CelebA 64×64 | cFID: 75.41 -> 19.79 | Helped |
| Use Boost-and-Skip instead of ADM | ADM · ImageNet 64×64 | cFID: 18.37 -> 12.68 | Helped |
| Use Boost-and-Skip instead of ADM | ADM · LSUN Bedrooms 256×256 | cFID: 63.30 -> 44.77 | Helped |
| Use Boost-and-Skip instead of ADM | ADM · ImageNet 256×256 | cFID: 13.22 -> 10.04 | Helped |
| Add boosted initialization without timestep skipping (ablation) | Boost-and-Skip · CelebA 64×64 minority | cFID: 19.79 -> 81.75 | Hurt |
| Add timestep skipping without variance boosting (ablation) | Boost-and-Skip · CelebA 64×64 minority | cFID: 19.79 -> 55.15 | Hurt |
| Set boosted initialization to γ²=9.0 with Δt=3 (ablation) | Boost-and-Skip · CelebA 64×64 minority | cFID: 23.56 -> 219.21 | Hurt |
| High-pass filter boosted initialization at cutoff 8 (ablation) | Boost-and-Skip · CelebA | cFID: 23.56 -> 60.16 | Hurt |
| Low-pass filter boosted initialization at cutoff 8 (ablation) | Boost-and-Skip · CelebA | cFID: 23.56 -> 33.08 | Hurt |
| Reduce initialization variance to γ²=0.6 (ablation) | Boost-and-Skip · CelebA test set | Precision: 0.851 -> 0.881 | Helped |
| Reduce initialization variance to γ²=0.6 (ablation) | Boost-and-Skip · CelebA test set | Recall: 0.627 -> 0.586 | Hurt |
| Use stochastic reverse-SDE rather than PF-ODE (ablation) | Boost-and-Skip · Two-circle toy data | On-manifold minority generation | Helped |
Surprising findings
- Boosting alone or skipping alone yields only marginal gains; only the combination produces significant minority generation, and PF-ODE samplers fail with the method because they lack contraction.
- Using gamma < 1 with skipping can instead improve sample quality, so the same mechanism acts as a quality-diversity control.
Limitations
- Naive application to deterministic (ODE) samplers fails due to the absence of a contracting effect.
- Sensitive to the choice of gamma and skip amount; too large gamma (e.g., 9.0 with skip 3) collapses precision.
Experimental setup
Datasets: CelebA LSUN-Bedrooms ImageNet
Benchmarks: CelebA 64x64 LSUN Bedrooms 256x256 ImageNet 64x64 ImageNet 256x256
Metrics: cFID sFID Precision Recall AvgkNN LOF Rarity Score Density Coverage
Baselines: ADM BigGAN StyleGAN LDM EDM DiT CADS ADM-ML temperature sampling
Builds on: DDPM score-based SDE truncation trick temperature sampling
Closest papers by meaning
- Don't Play Favorites: Minority Guidance for Diffusion Models (ICLR 2024 · similarity 0.91)
- Minority-Focused Text-to-Image Generation via Prompt Optimization (CVPR 2025 · similarity 0.90)
- Rethinking Training for De-biasing Text-to-Image Generation: Unlocking the Potential of Stable Diffusion (CVPR 2025 · similarity 0.87)
- AdaDiff: Adaptive Step Selection for Fast Diffusion Models (AAAI 2025 · similarity 0.87)
- Free Lunch for Pass@$k$? Low Cost Diverse Sampling for Diffusion Language Models (NeurIPS 2026 · similarity 0.87)
- SPARKE: Scalable Prompt-Aware Diversity and Novelty Guidance in Diffusion Models via RKE Score (NeurIPS 2025 · 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.