Research map / Probabilistic and scientific modeling / Variational and generative sampling

Multilevel and Sequential Monte Carlo for Training-Free Diffusion Guidance

NeurIPS 2026 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Hamiltonian Monte Carlo · 0 citations (OpenAlex)

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 an SMC sampler for training-free diffusion guidance that uses an unbiased Monte Carlo (multilevel MC) estimator of the marginal likelihood p(y|x_t) as particle weights, instead of point-estimate posterior approximations.

Problem

Training-free conditional sampling from diffusion models relies on point-estimate approximations of the marginal likelihood p(y|x_t), which are biased and fail to capture multimodal denoising distributions, causing weight degeneracy and poor sample validity.

Approach

Use an unbiased MLMC estimator of p(y|x_t) by integrating over the full denoising distribution via coupled discretized reverse trajectories, employed as SMC importance weights with a fixed signal-adaptive resampling schedule; TFG can serve as the proposal on high-dimensional tasks.

Key results

• On CIFAR-10 label guidance, SMC-MLMC reaches 95.6% accuracy (FID 46.3) versus 77.1% for TFG-4 and 62.0% for FreeDoM, with cost-per-success 23.4s vs 68.5s for TFG-1 (about 3x lower).
• On ImageNet 256x256 (six classes), success rate rises from 20.2% (TFG) to 58.3%, with FID improving from 231.9 to 182.3 and a 1.5x cost-per-success advantage.

Surprising findings

• The authors could not reproduce the published TFG ImageNet accuracy using the released code, reporting lower accuracy than originally stated.

Limitations

• Per-run cost is higher than heuristic baselines; advantage is measured as cost-per-success.
• Resampling schedule and proposal choice are problem-dependent and require pilot tuning.
• ImageNet results rely on TFG as proposal and small particle count (N=4) over six classes.

Experimental setup

Datasets: CIFAR-10 ImageNet

Benchmarks: CIFAR-10 label guidance ImageNet 256x256 class guidance (6 classes)

Metrics: top-1 accuracy FID cost-per-success

Baselines: DPS LGD FreeDoM TFG Twisted Diffusion Sampler Universal Guidance

Builds on: Diffusion Posterior Sampling Twisted Diffusion Sampler multilevel Monte Carlo Universal Guidance

Closest papers by meaning

• No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.91)
• Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.90)
• Characteristic Guidance: Non-linear Correction for Diffusion Model at Large Guidance Scale (ICML 2024 · similarity 0.90)
• Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.90)
• Tempered Guided Diffusion (NeurIPS 2026 · similarity 0.90)
• TFG: Unified Training-Free Guidance for Diffusion Models (NeurIPS 2024 · similarity 0.90)