Is-NeRF: In-scattering Neural Radiance Field for Blurred Images

Nan Luo, Chenglin Ye, Jiaxu Li, Gang Liu, Bo Wan, Di Wang, Lupeng Liu, Jun XiaoarXiv 2025 · DOI · arXiv

Abstract

Neural Radiance Fields (NeRFs) typically assume a straight-line volume rendering paradigm, which fails to accurately represent complex optical phenomena like reflection and refraction, especially in the presence of motion blur. We present Is-NeRF, an in-scattering neural radiance field framework that reconstructs high-quality 3D scenes from blurred images. By replacing traditional ray tracing with a unified In-scattering Lightpath Model (ISLM), Is-NeRF captures diverse light transport effects. Our approach includes scattering-aware volume rendering and self-guided adaptive sampling to resolve geometric ambiguities. Experimental results on public and self-captured datasets demonstrate that Is-NeRF significantly outperforms existing state-of-the-art deblurring methods in both visual quality and quantitative metrics.

Introduction

Most Neural Radiance Fields (NeRFs) rely on a "straight-line" volume rendering paradigm, assuming light travels in a perfectly linear path from a surface to the camera. While effective for simple, matte (Lambertian) surfaces, this assumption fails in real-world environments containing glass, water, or metallic objects. These materials cause light to reflect or refract, creating high-frequency color changes that a standard NeRF cannot easily map to a single 3D point.

The In-scattering Lightpath Model (ISLM) unifies various optical phenomena like reflection, refraction, and diffusion into a single representation.

This challenge is intensified when input images are motion-blurred. Standard deblurring NeRFs attempt to model camera motion but maintain straight-line rendering, leading to "geometric ambiguities." The model becomes unable to distinguish whether blur is caused by camera motion or complex optical reflections, resulting in hazy reconstructions and lost structural detail.

Methodology

Is-NeRF introduces three primary innovations to handle complex light transport and motion blur:

  1. In-scattering Lightpath Model (ISLM): Instead of treating reflection, refraction, and diffusion as separate problems, ISLM unifies them. It assumes that light from various directions can "scatter" into the primary path toward the camera at any point in the volume.
  2. Scattering-aware Volume Rendering: The classic NeRF rendering equation is extended to integrate contributions from "scattering paths" that branch off the primary ray, allowing the model to correctly place radiance for non-Lambertian effects.
  3. Self-Guided Adaptive Sampling: To maintain efficiency, the model uses a learning strategy to autonomously determine optimal scattering directions and sampling intervals. This "local density correction" helps preserve thin structures and small-scale objects that might otherwise be missed.
The Is-NeRF pipeline jointly optimizes scene parameters, scattering paths, and camera trajectories to produce sharp images from blurry inputs.

Key Insights

A critical finding in the development of Is-NeRF was the value of sampling diversity over multiplicity. The authors discovered that emitting a single scattering ray from five different points along the primary path yielded significantly better results than emitting five rays from a single intersection point. This suggests that capturing the spatial origin of scattering across a volume is more vital for accuracy than high-density sampling at a single location.

Furthermore, by modeling the physical scattering process alongside camera motion, the system can effectively distinguish between motion blur and optical highlights, preventing the "hallucination" of incorrect geometry behind reflective surfaces.

Experimental Results

Is-NeRF was evaluated on the public Deblur-NeRF Dataset and a new Self-Captured Dataset featuring complex light propagation (e.g., refraction through glass bottles). It consistently outperformed state-of-the-art (SOTA) methods like BAD-NeRF and DP-NeRF.

  • Quantitative Performance: On the reflection-heavy "Pool" scene, Is-NeRF achieved a PSNR of 33.53, compared to the baseline's 32.15.
  • Qualitative Fidelity: Visual comparisons show that Is-NeRF successfully renders sharp details, such as a straw inside a glass bottle or crisp reflections on car paint, where previous methods produced indistinct, muddy results.

Conclusion

Is-NeRF represents a significant advancement in 3D reconstruction from imperfect data. By explicitly modeling in-scattering, it bridges the gap between simplified volume rendering and complex real-world optics. While the additional scattering paths introduce computational overhead, future integration with faster architectures like Instant-NGP or 3D Gaussian Splatting offers a promising path toward real-time, high-fidelity scene reconstruction.

Summary updated

BibTeX
@misc{luo2025is,
  author = {Nan Luo and Chenglin Ye and Jiaxu Li and Gang Liu and Bo Wan and Di Wang and Lupeng Liu and Jun Xiao},
  title = {{Is-NeRF: In-scattering Neural Radiance Field for Blurred Images}},
  year = {2025},
  eprint = {2508.13808},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.13808},
}
Figure