BridgeTwist
A training-free conditional sampler for pretrained Schrödinger bridge models. Its core method is annealed twisted sequential Monte Carlo, which uses the velocity field and score field to construct a closed-form twisting proxy, enabling conditional generation without additional training.
Timeline (last 90 days)
Before BridgeTwist appeared, SB models lacked a training-free conditional sampling scheme.
BridgeTwist provides a training-free conditional sampler for pretrained SB models.
The core of BridgeTwist is a tempered twisted sequential Monte Carlo scheme.
BridgeTwist introduces a time-varying exponential tempering strategy that reduces the twisting influence early in sampling and gradually strengthens it later.
BridgeTwist designs a bridge-shaped inference-time diffusion schedule that concentrates the proposal noise in the interior region of the path.
With BridgeTwist combined with telescoping importance weights, all intermediate twisting factors are canceled layer by layer, ultimately yielding a sampler that is asymptotically exact for the posterior distribution.
BridgeTwist surpasses existing training-free baseline methods in class-conditional image generation and text-prompted image generation tasks.
BridgeTwist is competitive with solvers specifically designed for inverse problems in inverse problem solving (image inpainting, deblurring, super-resolution).
BridgeTwist is still at the research stage, and issues such as its performance in actual large-scale deployment and its robustness to different SB pretraining qualities remain to be verified.
All Facts (9)
Before BridgeTwist appeared, SB models lacked a training-free conditional sampling scheme.
50%UnverifiedBridgeTwist provides a training-free conditional sampler for pretrained SB models.
50%UnverifiedThe core of BridgeTwist is a tempered twisted sequential Monte Carlo scheme.
50%UnverifiedBridgeTwist introduces a time-varying exponential tempering strategy that reduces the twisting influence early in sampling and gradually strengthens it later.
50%UnverifiedBridgeTwist designs a bridge-shaped inference-time diffusion schedule that concentrates the proposal noise in the interior region of the path.
50%UnverifiedWith BridgeTwist combined with telescoping importance weights, all intermediate twisting factors are canceled layer by layer, ultimately yielding a sampler that is asymptotically exact for the posterior distribution.
50%UnverifiedBridgeTwist surpasses existing training-free baseline methods in class-conditional image generation and text-prompted image generation tasks.
50%UnverifiedBridgeTwist is competitive with solvers specifically designed for inverse problems in inverse problem solving (image inpainting, deblurring, super-resolution).
50%UnverifiedBridgeTwist is still at the research stage, and issues such as its performance in actual large-scale deployment and its robustness to different SB pretraining qualities remain to be verified.
50%