Skip to main content
Back to timeline
arXivSource publication:

Schrödinger bridges get training-free conditional sampling: BridgeTwist fuses velocity and score into an endpoint predictor, no retraining per task

Synopsis

Pretrained Schrödinger bridge generative models previously needed a task-specific network for each likelihood family, so they could not be reused for training-free conditioning the way diffusion and flow models are. BridgeTwist combines the velocity and score fields exposed by bridge pretraining into a closed-form endpoint predictor, tempers that unreliable early twist with a time-varying exponent, and pairs it with a bridge-shaped diffusion schedule and importance-weight correction to form an asymptotically exact sampler. It outperforms training-free baselines on class-conditional and text-prompted image generation and is competitive with specialized solvers on deblurring, super-resolution and inpainting.

Source-provided article image: BridgeTwist: Twisting Schrödinger Bridges for Training-Free Conditional Sampling
Figure 1 · arXiv · Page 2

Interpretation

Training-free conditional sampling can now be built on a pretrained unconditional Schrödinger bridge, so one bridge prior serves many downstream conditional tasks through a differentiable likelihood without retraining a network per likelihood. Diffusion and flow-matching priors already had training-free routes such as gradient guidance, projection and sequential Monte Carlo, whereas existing Schrödinger-bridge conditional methods trained a task-specific network for each likelihood family. BridgeTwist outperforms training-free baselines on class-conditional and text-prompted image generation, and is competitive with specialized restoration solvers PnP-Flow, Flower and FlowDPS on CelebA-HQ super-resolution and Gaussian deblurring and on CIFAR-10 inpainting with half-image and random 50% masks.

The sampler advances each particle along the bridge and resamples by weight along the way, so the particle population converges toward the endpoint posterior under the given likelihood. The ideal lookahead twist is the predictive likelihood, which integrates over all future trajectories and is too costly to estimate by direct Monte Carlo, so a plug-in surrogate replaces it. The plug-in twist evaluates the likelihood at an endpoint conditional mean built from the bridge's velocity and score fields; a temperature exponent downweights it early, a bridge-shaped diffusion schedule concentrates proposal noise in the path interior, and telescoping importance weights cancel the intermediate twist factors so the endpoint target is unchanged.

Perspective

The demonstrated scope is Schrödinger bridge priors under a Brownian reference, on MNIST and CIFAR-10 class-conditional sampling, CelebA-HQ text-prompted generation, CelebA-HQ super-resolution and deblurring, and CIFAR-10 inpainting; it addresses generative-modeling researchers who want to reuse a pretrained bridge model for conditional sampling.

The method currently applies only to Brownian-reference Schrödinger bridges, and other references require reference-specific endpoint predictors; the twist uses only the conditional endpoint mean and can be inaccurate for strongly curved likelihoods or multimodal endpoint laws, with the second-order extension improving only low-dimensional cases; the sampler targets the posterior induced by the learned and discretized prior, leaving prior-learning and discretization errors; preparing the bridge prior also needs an extra score head, a one-time overhead.