Minwoo Kim, Seunghyeok Shin, Dabin Kim, Hongki Lim
原文摘要
Training-free diffusion priors are powerful for inverse problems, but measurement guidance during reverse sampling is local: aggressive updates can improve immediate data fit while disrupting later denoising. We introduce an anchored inertial correction rule that separates proposal generation from acceptance. At each corrected step, the denoiser prediction anchors a frozen clean-space objective combining measurement fit with a noise-level-dependent proximal term. Heavy-ball dynamics with step size and momentum computed from the measured curvature generate a clean-state candidate; a separate verifier evaluates it with the same objective, with the denoiser prediction retained as fallback; both stages are derived from one objective for the whole trajectory, considered one step at a time. The accepted clean state is re-noised to continue sampling. The rule applies to pixel and latent diffusion with differentiable measurement operators, requires no retraining, and adds no denoiser calls inside the correction loop. We analyze how objective verification yields local nonincrease and displacement control. Across natural-image benchmarks, the method achieves competitive reconstruction quality. Code is available at https://github.com/Seunghyeok0715/PIVOT2026.
Autonomous vehicles in enclosed spaces such as tunnels and mines rely on range maps that smoke, dust and fog degrade. Conformal prediction calibrates a keepout around the mapped walls that covers the true unsafe set at a chosen probability. In narrow passages this margin and the planner's buffer close the route, and existing remedies recover reach by relaxing the coverage level. This paper conditions the conformal margin on a label-free severity estimate per viewing direction and sweep, holding per-class coverage near its target as degradation varies. A planning and control layer enforces one braking-distance condition on the keepout at plan time and at every command, takes the measured tracking residual as its only margin and never alters the keepout. The keepout is validated on ground-robot smoke recordings and the closed loop on simulated quadrotor flights in a foggy tunnel. The proposed system substantially reduces entries into the calibrated keepout and reaches numerically more goals than the deployment pattern in both calibration regimes. The per-class keepout holds its coverage near the target on held-out recordings.