ENCORE
A method for inference-time control of diffusion models that employs replica exchange and trajectory caching to achieve the first exact parallel control scheme. It steers pre-trained generative models toward a target distribution without requiring retraining.
Timeline (last 90 days)
ENCORE is a method proposed for inference-time control of diffusion models, with the paper titled "ENCORE: Exact Non-equilibrium COntrol with Replica Exchange for Diffusion Generation," published on arXiv.
ENCORE focuses on tilted target distributions of the form π₀∝G₀·p₀, where p₀ is the sampler output distribution and G₀ is a directly computable reweighting function.
ENCORE claims to be the first exact parallel control method.
The key design of ENCORE is that each replica stores its own generation trajectory, making upward moves a simple trajectory truncation, thereby eliminating the need to actually simulate the time-reversal process.
The paper theoretically proves the target invariance of ENCORE, meaning the method converges to the correct target distribution without bias.
The paper's diffusion analysis points out that under certain regularity conditions, both sequential control and parallel control become unstable as temporal discretization is refined in the absence of guidance.
ENCORE has been validated on multiple types of tasks including synthetic target distributions, Boltzmann sampling of biomolecules, and image generation.
Experiments show that ENCORE achieves competitive levels of accuracy and diversity, and remains robust to sampler perturbations.
ENCORE can be applied to distilled samplers, a scenario in which existing RE correction methods cannot be used.
All Facts (9)
ENCORE can be applied to distilled samplers, a scenario in which existing RE correction methods cannot be used.
50%UnverifiedENCORE focuses on tilted target distributions of the form π₀∝G₀·p₀, where p₀ is the sampler output distribution and G₀ is a directly computable reweighting function.
50%UnverifiedENCORE claims to be the first exact parallel control method.
50%UnverifiedThe key design of ENCORE is that each replica stores its own generation trajectory, making upward moves a simple trajectory truncation, thereby eliminating the need to actually simulate the time-reversal process.
50%UnverifiedThe paper theoretically proves the target invariance of ENCORE, meaning the method converges to the correct target distribution without bias.
50%UnverifiedThe paper's diffusion analysis points out that under certain regularity conditions, both sequential control and parallel control become unstable as temporal discretization is refined in the absence of guidance.
50%UnverifiedENCORE has been validated on multiple types of tasks including synthetic target distributions, Boltzmann sampling of biomolecules, and image generation.
50%UnverifiedExperiments show that ENCORE achieves competitive levels of accuracy and diversity, and remains robust to sampler perturbations.
50%UnverifiedENCORE is a method proposed for inference-time control of diffusion models, with the paper titled "ENCORE: Exact Non-equilibrium COntrol with Replica Exchange for Diffusion Generation," published on arXiv.
50%