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

BridgeTwist: Twisting Schrödinger Bridges for Training-Free Conditional Sampling

NeurIPS 2026 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Posterior sampling with diffusion guidance

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 BridgeTwist, a training-free conditional sampler for pretrained Schrödinger bridge generative models using a tempered twisted SMC with a velocity-score endpoint predictor and a bridge-shaped diffusion schedule.

Problem

Training-free conditional sampling from pretrained unconditional Schrödinger bridge (SB) generative models, which has been developed for diffusion and flow models but not for SBs.

Approach

Combine the SB velocity and score into a closed-form endpoint predictor E[X1|Xt] used as a plug-in twist in SMC, temper the twist with a time-varying exponent t^a, and use a bridge-shaped inference diffusion schedule g(t) ∝ sqrt(t(1-t)); telescoping importance weights make the sampler asymptotically exact.

Key results

• On CIFAR-10 class-conditional sampling at K=64, BridgeTwist reaches ACCE 0.979 with KID 0.0403, versus 0.865 ACCE for the best baseline IS.
• At matched NFE budgets on CIFAR-10, BridgeTwist gives the best ACCE at all budgets above the smallest (0.984 at the largest budget).
• On CelebA-HQ deblurring BridgeTwist reaches PSNR 31.13 and SSIM 0.847, the best among compared methods; on super-resolution it ranks second to PnP-Flow.

Surprising findings

• Velocity-only and score-only endpoint predictors each substantially degrade class alignment, showing both channels are needed.
• FMPS leads at the smallest NFE budget on CIFAR-10, so BridgeTwist is not best in every regime.

Limitations

• Restricted to Schrödinger bridges with a Brownian reference; other references need reference-specific endpoint predictors.
• Twist uses only the conditional endpoint mean, which can be inaccurate for strongly curved or multimodal conditional endpoint laws.
• Requires an additional score head when preparing the SB prior, a one-time overhead relative to pretrained flow or diffusion models.

Experimental setup

Datasets: MNIST CIFAR-10 CelebA-HQ

Benchmarks: MNIST CIFAR-10 CelebA-HQ 2D Gaussian mixture toy benchmark

Metrics: ACCL ACCE KID Density Coverage CLIP similarity LPIPS diversity PSNR SSIM MMD

Baselines: FMPS TFTF TDS DPS-SB IS PnP-Flow Flower FlowDPS

Builds on: twisted diffusion sampler SF2M (simulation-free score and flow matching) diffusion posterior sampling Feynman-Kac/twisted SMC annealed importance sampling

Closest papers by meaning

• Adjoint Schrödinger Bridge Sampler (NeurIPS 2025 · similarity 0.88)
• DEFT: Efficient Fine-tuning of Diffusion Models by Learning the Generalised $h$-transform (NeurIPS 2024 · similarity 0.87)
• Outsourced Diffusion Sampling: Efficient Posterior Inference in Latent Spaces of Generative Models (ICML 2025 · similarity 0.87)
• Transport meets Variational Inference: Controlled Monte Carlo Diffusions (ICLR 2024 · similarity 0.87)
• Tempered Guided Diffusion (NeurIPS 2026 · similarity 0.87)
• Stochastic Interpolants via Conditional Dependent Coupling (NeurIPS 2026 · similarity 0.86)