PaperlayerView exact revision on arXiv

Source-linked AI summary

NeRF: Neural Radiance Field in 3D Vision: A Comprehensive Review (Updated Post-Gaussian Splatting)

Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, Jonathan Li

arXiv:2210.00379v8cs.CV

TL;DR

NeRF research has shifted toward Gaussian Splatting in many novel-view-synthesis areas, while NeRF and adjacent neural rendering persist in applications favoring implicit or volumetric representations. This survey organizes the field, introduces NeRF theory, and compares models, applications, and datasets.

  • Problem

    NeRF and adjacent neural rendering methods span rapidly evolving architectures and applications, but the field's shift toward Gaussian Splatting makes a consolidated account of their continuing role valuable.

  • Method

    The survey categorizes NeRF, neural-field, and implicit or hybrid representation methods by architecture, application, and active research area, while covering theory, differentiable volume rendering, benchmarks, and datasets.

  • Results

    NeRF and implicit or hybrid neural-field methods have progressed in speed, quality, and training-view requirements and remain useful in applications where implicit, hybrid, or volumetric representations are advantageous.

  • Takeaways & Limitations

    The survey provides a reference for NeRF models and datasets and identifies continuing applications including SLAM, human avatars, urban mapping, image processing, and semantic 3D interaction.

  • Takeaways & Limitations

    NeRF models are typically trained per scene and require relatively dense images with varied poses, although some methods address sparse or unposed inputs.

Abstract

from arXiv · show

In March 2020, Neural Radiance Field (NeRF) revolutionized Computer Vision, allowing for implicit, neural network-based scene representation and novel view synthesis. NeRF models have found diverse applications in robotics, urban mapping, autonomous navigation, virtual reality/augmented reality, and more. In August 2023, Gaussian Splatting, a direct competitor to the NeRF-based framework, was proposed, gaining tremendous momentum and overtaking NeRF-based research in terms of interest as the dominant framework for novel view synthesis. We present a comprehensive survey of NeRF papers from the past five years (2020-2025). These include papers from the pre-Gaussian Splatting era, where NeRF dominated the field for novel view synthesis and 3D implicit and hybrid representation neural field learning. We also include works from the post-Gaussian Splatting era where NeRF and implicit/hybrid neural fields found more niche applications. Our survey is organized into architecture and application-based taxonomies in the pre-Gaussian Splatting era, as well as a categorization of active research areas for NeRF, neural field, and implicit/hybrid neural representation methods. We provide an introduction to the theory of NeRF and its training via differentiable volume rendering. We also present a benchmark comparison of the performance and speed of classical NeRF, implicit and hybrid neural representation, and neural field models, and an overview of key datasets.

1 Introduction

NeRF uses implicit neural representations and differentiable volume rendering for photorealistic novel-view synthesis, but Gaussian Splatting shifted research interest after 2023. This survey extends earlier work to cover post-Gaussian-Splatting neural-field methods and organizes the literature by methods and applications.

  • NeRF and neural representations: NeRF represents scene geometry and lighting with neural fields stored in MLPs, then synthesizes images from user-specified viewpoints.The representation is typically implicit and learned through differentiable volume rendering.
  • Shift after Gaussian Splatting: Gaussian Splatting overtook NeRF and adjacent methods on many novel-view-synthesis benchmarks and applications in 2023.Research interest consequently shifted toward Gaussian Splatting, although NeRF-adjacent research persisted.
  • NeRF advantages: NeRF methods require only multiview images and poses for self-supervised scene learning, without 3D or depth supervision.Camera poses can be estimated with Structure from Motion tools such as COLMAP.
  • Representation trade-offs: Gaussian Splatting generally produces higher-quality images, trains faster, and renders faster, while implicit or hybrid NeRF methods require less storage and memory.The comparison describes trade-offs between explicit point-based and implicit neural representations.
  • Survey scope: The survey covers pre- and post-Gaussian-Splatting work, including neural fields that use NeRF-like differentiable volume rendering under authors’ newer terminology.Post-Gaussian-Splatting methods are included when they combine neural-field representations with NeRF-like rendering.

2 Background

The background reviews prior NeRF surveys, defines neural radiance fields and differentiable volume rendering, and introduces datasets, metrics, and representation terminology. A NeRF maps spatial coordinates and viewing directions to density and color, which are integrated along camera rays to render images.

  • Existing surveys: NeRF surveys differ in scope, with this review emphasizing detailed NeRF coverage, taxonomies, datasets, metrics, and more recent work.The paper positions itself as focused on NeRF rather than broad neural rendering or multimodal synthesis.
  • NeRF theory: A NeRF approximates a five-dimensional radiance field with MLPs that predict view-independent density and view-dependent color.The two-stage architecture first processes position, then combines learned features with viewing direction to predict color.
  • Differentiable volume rendering: Novel-view synthesis traces camera rays, evaluates color and density at sampled points, and applies volume rendering to produce each pixel.The ray bounds extend from the camera center to the scene’s bounding volume.
  • Differentiable volume rendering: Accumulated transmittance represents the probability that a ray travels without interception before reaching a point.This quantity weights contributions along the ray during rendering.
  • Depth and geometry: Expected depth and depth regularization can constrain densities toward surfaces and help geometry converge earlier with fewer input views.Depth cues may come from point clouds or other supervision sources.
  • Datasets and representations: The review distinguishes implicit, hybrid, and explicit 3D representations and discusses datasets spanning synthetic objects, indoor scenes, and city-scale driving.KITTI challenges NeRF training because its camera coverage is relatively sparse.

3 Neural Radiance Field (NeRF) pre Gaussian Splatting

The pre-Gaussian-Splatting survey section organizes selected NeRF and adjacent papers through method and application taxonomies. Its method taxonomy tracks improvements in quality, speed, sparse-view learning, generative editing, composition, dynamics, and pose estimation.

  • Scope and organization: The survey benchmarks influential pure novel-view-synthesis NeRF, neural-rendering, and adjacent methods on the Synthetic NeRF dataset.Publications are sequenced using first arXiv preprint dates.
  • Application taxonomy: Application-based classification is presented separately from the method-based taxonomy.This separation distinguishes what methods improve from where NeRF techniques are applied.
  • Method taxonomy: The method taxonomy covers photometric fidelity, learned geometry, training and inference speed, sparse or few-shot radiance fields, generative fields, compositional scenes, dynamics, and pose estimation.These themes span Sections 3.1 through 3.6.

3.1 Improvements in the Quality of Synthesized Views and Learned Geometry

Pre-Gaussian-Splatting research improved NeRF view quality and geometry through multiscale rendering, reflection-aware radiance, depth cues, point clouds, symmetry, and single-image geometric inference.

  • Better view synthesis: Mip-NeRF approximated cone tracing with Integrated Positional Encoding, producing multiscale anti-aliased rendering that outperformed baseline NeRF, especially at lower resolutions.Mip-NeRF 360 extended the approach to unbounded scenes using proposal-based density sampling and regularization.
  • Better view synthesis: Ref-NeRF modeled radiance using reflected viewing directions and predicted surface normals, roughness, diffuse color, and specular color.It performed particularly well on reflective surfaces and specular highlights.
  • Depth supervision and point clouds: Depth supervision from LiDAR, SfM, or learned priors helped models converge faster, reach higher final quality, and require fewer training views than baseline NeRF.Depth regularization can restrict densities toward scene surfaces or enforce depth smoothness.
  • Depth supervision and point clouds: PointNeRF used feature point clouds as an intermediate representation before volume rendering, combining geometric probabilities with image features.A 3D CNN generated depth and surface probabilities, while a 2D CNN supplied image features.
  • Geometry refinement: SNeS imposed soft symmetry constraints, while S3-NeRF used shadow and shading cues with an occupancy field to recover geometry from single images.S3-NeRF reported excellent depth-map and surface-normal reconstruction on synthetic and real-world datasets.

3.2 Improvements to Training and Inference Speed

NeRF speed improvements either bake MLP evaluations into accessible structures or use hybrid and explicit representations to reduce computation. These methods accelerate training, inference, or both, often trading memory or representation flexibility against speed.

  • Baked models precompute and store NeRF MLP evaluations in accessible data structures, improving inference speed without affecting training time.
  • Hybrid representations learn scene features separately from MLP parameters, enabling smaller MLPs and faster training and inference at increased memory cost.
  • 3000× faster inference was achieved by SNeRG and PlenOctree relative to the original NeRF implementation through sparse voxel-grid or octree caching.
  • 2000× acceleration was achieved by KiloNeRF through thousands of distilled cell-specific MLPs combined with empty-space skipping and early ray termination.
  • 8× faster training was achieved by pivot-based sampling and zero-density point rejection, followed by caching the trained scene into a NeRF tree.
  • 26–35× fewer FLOPs and 28–31× faster wall-clock rendering were achieved by R2L while surpassing NeRF and other efficient methods in visual quality.

3.3 Few Shot/Sparse Training View NeRF

Few-shot and sparse-view NeRF methods incorporate image features, geometric costs, regularization, or augmentation to compensate for insufficient viewpoint coverage. The survey describes approaches spanning generalization across scenes, single-view synthesis, and sparse-view reconstruction.

  • Sparse or insufficiently varied training views can cause NeRF overfitting and nonsensical geometry, motivating priors and regularization.
  • PixelNeRF, MVSNeRF, and related methods condition NeRF rendering on CNN-derived image features or 3D cost-volume encodings.
  • NeuRay generalized well to new scenes and could be fine-tuned to exceed baseline NeRF performance.
  • RegNeRF combines depth smoothness, color-patch likelihood regularization, and sampled-space annealing to train without pretraining under sparse views.
  • RegNeRF outperformed Mip-NeRF and DietNeRF under sparse-view conditions and performed comparably to pretrained PixelNeRF, SRF, and MVSNeRF models.

3.4 Generative and Conditional Models

Generative and conditional NeRFs use latent codes, text, images, or diffusion and adversarial models to synthesize and edit 3D scenes. The surveyed methods progress from latent-conditioned generation toward more controllable text- and image-guided pipelines.

  • Generative NeRFs condition 3D geometry on text, images, or latent codes, commonly using 2D generative models to create training images.
  • Early GAN-based NeRF models relied on latent codes, which provided less direct text and image control than later diffusion- and flow-matching-based approaches.
  • NeRF-VAE generalized to out-of-distribution scenes without training from scratch and outperformed baseline NeRF with 5–20 scene views.
  • GAN-based models such as GRAF, π-GAN, EG3D, and StyleNeRF combine conditional scene representations with adversarial image synthesis.
  • CLIP-NeRF maps text or image inputs into shape and appearance latent displacements that modify a scene’s NeRF representation.
  • SSDNeRF jointly optimizes NeRF and diffusion components end-to-end, supporting robust sparse-view learning and flexible test-time sampling.

3.5 Unbounded Scene and Scene Composition

NeRF extensions address outdoor and unbounded scenes, reflective surfaces, foreground-background composition, and dynamic or articulated content. They do so by separating scene components, introducing specialized radiance fields, or adding temporal representations.

  • Outdoor and unbounded scenes require separating foreground and background while handling image-to-image lighting and appearance variation.
  • NeRF-W models per-image appearance variation and transient objects using appearance and transient embeddings while keeping scene density fixed.
  • NeRF++ handles unbounded scenes by separating foreground and background with a sphere, radial inversion, and two independently trained NeRFs.
  • Fig-NeRF uses separate foreground and background NeRFs for object interpolation and amodal segmentation, achieving good results across ShapeNet, Gelato, and Objectron.
  • NeRFRen separates transmitted and reflected radiance fields and uses depth-based losses to reconstruct reflective-surface geometry.
  • Dynamic-scene methods add time through deformation fields, canonical representations, explicit voxel grids, or hybrid spacetime feature planes.
  • MAV3D synthesizes time-varying 3D scenes from text using a 4D radiance field guided by a pretrained text-to-video diffusion model without 4D datasets.

3.6 Pose Estimation

NeRF pose-estimation research relaxed the requirement for known camera poses by jointly optimizing poses and radiance fields, using inverse, generative, bundle-adjusted, and SLAM-based approaches. Results improved across several settings, although initialization and real-time constraints remained important boundaries.

  • NeRF– jointly estimated camera and radiance-field parameters end-to-end, achieving results comparable to COLMAP-based NeRF but struggling with rotational motion and object tracking.Its pose initialization limitations made it most suited to front-facing scenes.
  • BARF exceeded NeRF– by 1.49 PSNR on average across eight LLFF forward-facing scenes with unknown camera poses.It jointly optimized poses and the radiance field using coarse-to-fine registration through adaptive positional-encoding masking.
  • GNeRF used pose as a generative latent code and was slightly outperformed by COLMAP-based NeRF on Synthetic-NeRF but outperformed it on DTU.Adversarial training produced coarse pose–image pairs before joint photometric refinement.
  • NICE-SLAM improved pose estimation and reconstruction over iMAP while using approximately one-quarter of its FLOPs, one-third of tracking time, and half of mapping time.Its hierarchical grid-based geometry representation filled gaps in large-scale unobserved scene features, though it remained far from real-time.
  • NICER-SLAM improved on NICE-SLAM with end-to-end dense SLAM using only RGB inputs and a hierarchical SDF-based neural implicit representation.Monocular geometric cues, optical flow, and warping loss guide optimization without depth supervision.

3.7 Adjacent Methods for Neural Rendering

Adjacent neural-rendering methods replaced or compressed the standard NeRF MLP while retaining neural or volumetric scene-rendering principles. Explicit, factorized, compressed, and transformer-based designs targeted faster training, lower memory, or alternative ray-processing strategies.

  • Plenoxels voxelized scenes and directly optimized density and spherical-harmonic color features without MLP training, achieving NeRF++- and JaxNeRF-comparable results.Training was faster by a factor of a few hundred, suggesting volumetric rendering rather than density and color MLPs was the primary contribution.
  • TensoRF represented density and directional features with factorized tensors, reducing memory requirements from Plenoxels by a factor of 200 using CP factorization.Its vector–matrix factorization achieved better performance than CP, while training remained much faster than implicit NeRF models.
  • Streaming Radiance Fields achieved approximately 90-times-faster training than Plenoxels with memory requirements 100 to 300 times smaller.Model-difference compression, narrow-band tuning, and training acceleration targeted video-based NeRF learning.
  • Dictionary Fields and Factor Fields factorized neural fields into a spatial coefficient field and shared basis field to improve quality per parameter and accelerate training.The shared basis supports multi-signal learning and few-shot or sparse-view reconstruction while coefficients remain scene-specific.
  • IBRNet and SRT used image features and transformers to predict or render target rays, with SRT querying an entire ray at once without producing a density function.These methods differ from conventional NeRF ray sampling and explicit density modeling.
  • NeRFormer processed ray-depth-ordered feature sequences from multiple source views using pooling and ray-wise attention, and introduced the Common Objects in 3D dataset.

4 Applications of NeRF and adjacent methods pre-Gaussian Splatting

Pre-Gaussian-Splatting NeRF research applied neural rendering across urban reconstruction, robotics, image processing, and semantic understanding. These methods adapted representations, supervision, and optimization to sparse views, large outdoor scenes, difficult lighting, noise, and limited labels.

  • Robotics and navigation: NeRF supported robotics navigation and grasping by representing environments for collision checking and using learned density to estimate depths for transparent-object grasp planning.Dex-NeRF introduced one synthetic and two real-world transparent-object datasets; Evo-NeRF improved sequential grasping and planning with unreliable geometry.
  • The survey classifies applications into urban reconstruction, human faces and articulated bodies, surface reconstruction, and low-level image processing, with navigation and generative methods covered separately.
  • Urban reconstruction: Urban NeRF methods addressed unbounded environments, limited camera diversity, and large-scale reconstruction using LiDAR, scene partitioning, aerial imagery, or transient-object masking.Urban Radiance Fields added depth, sight, and skybox segmentation losses, while Mega-NeRF and Block-NeRF scaled reconstruction to large outdoor scenes.
  • Image processing: RawNeRF adapted Mip-NeRF to HDR view synthesis, denoising, and low-light scenes by training on raw linear images and applying exposure and tone mapping after rendering.HDR-NeRF instead used variable-exposure low-dynamic-range images for HDR view synthesis.
  • Image processing: NaN achieved state-of-the-art burst denoising under large motion and high noise while avoiding per-scene training through inter-view and spatial awareness.It built on IBRNet’s unseen-scene generalization with minimal input.
  • Semantic understanding: Semantic NeRF methods trained semantic labels with categorical cross-entropy, operating with sparse labels and recovering labels from pixel-wise and region or instance-wise noise.They also supported label super-resolution, sparse point-wise propagation, and multi-view semantic fusion; a KITTI-360 method outperformed prior semantic-label-transfer approaches.

5 Post-Gaussian Splatting Neural Rendering and NeRF

Post-Gaussian Splatting research continues to develop NeRF and neural-field methods for efficient rendering, scene representation, semantics, language grounding, and specialized physical effects. These methods increasingly target niche capabilities while combining implicit or hybrid representations with complementary models.

  • Post-Gaussian Splatting landscape: Gaussian Splatting uses anisotropic 3D Gaussians for fast differentiable rendering, typically producing slightly better images while requiring more memory and storage than NeRF.Each Gaussian encodes position, scale, orientation, opacity, and color.
  • Post-Gaussian Splatting landscape: Post-Gaussian Splatting work emphasizes efficient volume rendering, spectral generalization, view-dependent effects, semantic integration, vision-language grounding, and generative priors.The survey categorizes these developments across rendering, scene representation, diffusion, and SLAM applications.
  • Differentiable volume rendering: NeuRBF combines adaptive radial basis functions with grid-based interpolation and multi-frequency composition, then decodes features with an MLP for volume rendering and SDF reconstruction.It reports state-of-the-art performance across 2D image fitting, 3D SDF reconstruction, and neural radiance field synthesis.
  • Differentiable volume rendering: PIE-NeRF couples elastodynamic simulation with a meshless discretization over a NeRF density field, supporting large deformations and nonlinear hyperelastic material simulation.The method avoids converting the implicit surface into a mesh or voxel grid.
  • Differentiable volume rendering: Hybrid approaches combine Gaussian geometry with NeRF-based color and opacity, preserving real-time inference while reducing view-dependent artifacts.VDGS uses 3D Gaussian Splatting for geometry and NeRF-based encoding for appearance.
  • Differentiable volume rendering: MulFAGrid integrates multiplicative filters with Fourier features and Grid Tangent Kernel theory to improve high-frequency spectral efficiency and generalization.It supports regular and irregular grids through joint optimization of grid and kernel features.
  • Differentiable volume rendering: NU-NeRF reconstructs nested transparent objects through staged reflection-refraction separation, outer-surface fitting, explicit ray tracing, and inner-surface reconstruction.The method uses neural signed distance fields and refraction or transparency models across its stages.
  • Differentiable volume rendering: VD-NeRF disentangles view-independent appearance from view-dependent effects while modeling visibility and shadowing for consistent editing and high-frequency relighting.Its two-branch architecture separates base appearance from view-dependent residuals.

5.3 Diffusion and Neural Fields

Diffusion models are combined with NeRFs and neural fields for language-driven editing, sparse-view reconstruction, 4D generation, inpainting, super-resolution, and relighting. These methods use multi-view constraints, learned priors, or iterative synchronization to improve consistency and quality.

  • Diffusion for 3D generation and editing: Text-to-image diffusion models enable language-driven object insertion and removal in NeRF scenes through synthesized multi-view images and iterative refinement.The system demonstrates high-quality, view-consistent edits with minimal manual input.
  • Diffusion for 3D generation and editing: ReconFusion uses a multiview-conditioned diffusion prior as a regularizer within NeRF training to improve fidelity under sparse views and mitigate floaters and fog.The diffusion model is fine-tuned from a pretrained latent diffusion backbone using real and synthetic data.
  • Diffusion for 3D generation and editing: Comps4D generates animated 3D scenes by separating static NeRF-based asset creation from LLM-guided global motion planning and learned local deformations.The framework extends beyond earlier object-centric approaches.
  • Diffusion for 3D generation and editing: LN3Diff maps images into a compact 3D-aware latent space, decodes triplanes with a transformer, and supports reconstruction and generation from as few as two views.It reports state-of-the-art performance on ShapeNet, FFHQ, and Objaverse, with up to 3× faster inference than existing GAN and diffusion baselines.
  • Diffusion for 3D generation and editing: Inpaint4DNeRF uses Stable Diffusion on seed views, infers coarse geometry, and refines remaining views using diffusion guided by seed images and geometry.The pipeline extends to 4D dynamic scenes and enforces multiview consistency during refinement.
  • Diffusion aiding image processing: DiSR-NeRF alternates 2D diffusion super-resolution with NeRF training and optimizes intermediate denoised latents to produce sharper, more consistent high-resolution NeRFs from low-resolution views.It does not require high-resolution training data and outperforms existing methods in high-fidelity super-resolved NeRF generation.
  • Diffusion aiding image processing: MVIP-NeRF jointly optimizes masked regions across views with text-conditioned score distillation, RGB reconstruction, normal-map distillation, and a multi-view SDS formulation.The method reports state-of-the-art results for NeRF inpainting, including under large view changes.
  • Diffusion aiding image processing: Neural Gaffer provides category-agnostic single-view relighting using a 2D diffusion prior trained with physically based materials and HDR environment maps.It extends to 3D relighting through NeRFs and outperforms existing approaches on synthetic and real data.

5.4 SLAM with Implicit and Hybrid Neural Fields

Implicit and hybrid neural fields support diverse SLAM systems for semantic mapping, dynamic scenes, loop closure, multi-agent collaboration, and real-time operation. Recent methods combine neural representations with geometric, semantic, distributed, and point-based components to improve robustness and scalability.

  • Semantic SLAM: SNI-SLAM targets accurate real-time 3D semantic mapping by jointly modeling appearance, geometry, and semantics.It uses hierarchical semantic encoding, cross-attention, and a one-way decoder to improve inter-modal information flow.
  • Semantic SLAM: DNS-SLAM uses class-wise scene decomposition and a hybrid point-based neural field for semantic-aware SLAM.The framework links object classes to scene representation and introduces the INS dataset for realistic benchmarking.
  • Graph-based and global mapping: Neural Graph Mapping anchors lightweight neural fields to pose-graph keyframes, allowing deformation during loop closure without costly reintegration.The RGB-D framework combines sparse visual tracking with dense neural mapping across diverse scenes.
  • Point-based and SDF representations: PIN-SLAM replaces grid structures with neural feature points and combines hybrid neural SDF mapping with correspondence-free odometry.Its representation supports spatial flexibility and elastic correction during loop closure.
  • Multi-agent SLAM: MNE-SLAM enables fully distributed multi-agent mapping and camera tracking without centralized training or raw-data exchange.It uses a triplane+MLP hybrid neural field and an intra-to-inter loop-closure strategy.
  • Hybrid rendering: A hybrid SDF-plus-Gaussian renderer achieves real-time to high-frame-rate operation while maintaining or improving reconstruction quality relative to Gaussian-only RGB-D SLAM.It raycasts SDF depth and base color, then splats surface-near Gaussians with SDF-guided depth culling, avoiding expensive volumetric sampling and Gaussian depth sorting.

5.5 Human avatars with Implicit and Hybrid Neural Fields

Human-avatar research uses implicit and hybrid neural fields for photorealistic, controllable, relightable, and animatable representations. Recent methods improve efficiency through hash grids, learned manifolds, mesh anchoring, distillation, and reduced rendering spaces.

  • Facial avatars: HQ3D generates photorealistic facial avatars from multi-view video while requiring only monocular RGB input at test time.Its hybrid field uses a voxelized feature grid, multiresolution hash encoding, and a decoding MLP without mesh templates or space pruning.
  • Facial avatars: Spatially-Varying Expression replaces uniform global expression codes with position-dependent features for finer facial control.The method combines 3DMM expression parameters with spatial features to control geometry and rendering locally.
  • Efficient rendering: BakedAvatar uses a learned head-surface manifold and layered mesh proxies to approximate volumetric rendering with fast rasterization.The representation targets photorealistic real-time 4D head-avatar rendering on commodity devices.
  • Efficient rendering: LightAvatar removes dependence on explicit meshes and volume rendering through neural light-field distillation.Training combines pseudo and real data, while a warping field corrects imperfect 3DMM fitting on real data.
  • Animatable and relightable avatars: NeRFFaceShop learns animatable and relightable human heads from large-scale in-the-wild videos by deforming feature maps rather than 3D points.A tri-plane generator and unified animation-lighting representation support detailed, controllable motion.
  • Hybrid mesh-field avatars: MeshAvatar combines explicit triangular meshes with neural signed-distance and material fields for hardware-compatible avatar rendering.Differentiable marching tetrahedra bridges mesh and implicit components, while physics-based rendering and neural encoders add relighting and high-frequency detail.

6 Discussion

NeRF and Gaussian Splatting differ in representation and rendering: neural fields use implicit or hybrid functions with differentiable volume rendering, while Gaussian Splatting uses explicit point-like primitives. Gaussian Splatting generally offers greater speed and quality, whereas neural fields retain advantages in memory efficiency, continuous modeling, and applications requiring implicit representations.

  • NeRF vs. Gaussian Splatting: NeRF-like methods use implicit or hybrid neural fields and differentiable volume rendering, whereas Gaussian Splatting uses explicit 3D point-cloud-like representations.This distinction reflects different scene representations and rendering paradigms.
  • NeRF vs. Gaussian Splatting: NeRF-like methods are typically more memory- and storage-efficient but much slower than Gaussian Splatting and often produce slightly lower view-synthesis quality.Gaussian methods use explicit Gaussian primitives rather than sampling color values along camera rays.
  • Research landscape: Gaussian Splatting has largely overtaken NeRF-adjacent methods in novel view synthesis and several adjacent research areas because of faster training and inference and higher view-synthesis quality.The shift includes 3D model generation, semantic view synthesis, and scene representation, reconstruction, and editing.
  • Hybrid methods: Hybrid Gaussian-neural-field methods seek to combine fast rasterization with continuous, view-dependent, and expressive neural modeling.They use Gaussians or local feature points for primitives and neural networks for rendering parameters or local deformation and motion functions.
  • Post-Gaussian applications: Neural fields remain useful for SLAM and human avatars because they require less memory and storage and are directly queried at spatial points.Their implicit structure also fits articulated-avatar pipelines and can benefit onboard SLAM and autonomous navigation.
  • Post-Gaussian applications: Neural implicit signed distance functions remain active for high-fidelity meshing, continuous surface modeling, inverse rendering, relighting, and material editing.NeRF’s formulation supports lighting and deformation fields as well as physically based recovery of scene properties.

7 Conclusion

This survey reviews NeRF and implicit or hybrid neural-field research from 2020 to 2025, covering theory, taxonomies, applications, benchmarks, and datasets. It concludes that Gaussian Splatting shifted research interest, while neural fields remain relevant where implicit or volumetric representations are advantageous.

  • Conclusion: The survey covers NeRF and implicit or hybrid neural-field progress in speed, quality, training-view requirements, and applications.Applications include urban mapping, photogrammetry, image processing, 3D reconstruction, and human avatars.
  • Conclusion: Although research interest shifted toward Gaussian Splatting, NeRF-based methods remain active where implicit or hybrid representations or volumetric rendering are advantageous.The survey also notes that many Gaussian Splatting methods drew inspiration from earlier NeRF work.
  • Conclusion: The review aims to introduce practitioners to NeRF, provide references for existing models and datasets, and motivate future research.It presents NeRF as a paradigm for novel view synthesis, 3D reconstruction, and 3D scene representation.
  • Scope: The review draws on published literature and public knowledge, with cited sources accessible through standard scholarly channels.
Loading 2210.00379v8…