Neural Radiance Fields (NeRF) have attracted extensive attention because they can reconstruct high‑quality 3D geometry and synthesize novel views from multi‑view images. Existing pipelines usually assume sharp inputs, yet real‑world capture often suffers from motion or focus blur, which degrades NeRF reconstruction.
To tackle this, we introduce WS‑NeRF, a Mamba‑driven world‑state‑aware adaptive deblurring neural radiance field. The alternating optimization of the radiance field is modeled as a dynamic evolution process equipped with temporal memory. By leveraging comprehensive multi‑dimensional world states and a mixture‑of‑experts mechanism, the confidence of deblurring priors is adjusted on‑the‑fly during each iteration.
Experiments on synthetic and real blurry datasets show that WS‑NeRF markedly improves PSNR, SSIM, and LPIPS compared with baselines, while exhibiting more stable iterative convergence.
Review: The integration of temporal memory and expert mixing effectively mitigates blur‑induced inconsistencies, offering a robust NeRF solution for practical imaging scenarios.