nerf
Table of Contents
1. LIGHT
- 1
- Relighting Neural Radiance Fields with Shadow and Highlight Hints
- moving point light source
- second multi layer perceptron which takes shadow and highlight hints
1.1. REFLECTIONS
- nerf reflections: https://youtu.be/qrdRH9irAlk
- Mirror-NeRF: Learning Neural Radiance Fields for Mirrors with Whitted-Style Ray Tracing
- introducing the reflection probability
- SpecNeRF: Gaussian Directional Encoding for Specular Reflections
- learnable Gaussian directional encoding to model view-dependent effects under near-field light
- UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections
- faster reconstructions; explicitly blending these representations in 3D space
1.1.1. SNAP-IT
- Snap-it, Tap-it, Splat-it: Tactile-Informed 3D Gaussian Splatting for Reconstructing Challenging Surfaces
- incorporates touch data (local depth maps) with multi-view vision data to achieve surface reconstruction and novel view synthesis
- fewer images required
1.2. MESH INTEGRATION
- Dynamic Mesh-Aware Radiance Fields
- two-way rendering-simulation coupling between mesh and NeRF
- realistic light from NeRF media onto surfaces, cast shadows on the NeRF = enhanced realism
2. QUALITY
- Neuralangelo: High-Fidelity Neural Surface Reconstruction (3d mesh augmented)
- PyNeRF: Pyramidal Neural Radiance Fields
- modification to grid-based models by training model heads at different spatial grid resolutions
- reduce error rates by 20% while training over 60x faster against Mip-NeRF
- ReconFusion: 3D Reconstruction with Diffusion Priors
- leverages diffusion for novel view synthesis, trained on multiview; few photos
- DyBluRF: Dynamic Deblurring Neural Radiance Fields for Blurry Monocular Video
- deals with challenge of view synthesis from motion blur
3. CREATING NERFS
- FrameNeRF: A Simple and Efficient Framework for Few-shot Novel View Synthesis
- regularization model (sd) as a data generator to produce dense views from sparse inputs
3.1. POSE PREDICTION
- MELON: NeRF with Unposed Images Using Equivalence Class Estimation (no poses)
- DINER: Depth-aware Image-based NEural Radiance fields
- COLMAP-Free 3D Gaussian Splatting
- continuity of the input video = no need for camera poses
3.2. FROM VIDEO
- Efficient Neural Radiance Fields for Interactive Free-viewpoint Video (people)
- Progressively Optimized Local Radiance Fields for Robust View Synthesis
- turn video into nerf
- F2-NeRF: Fast Neural Radiance Field Training with Free Camera Trajectories
- turn video into nerf
- HOSNeRF: Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video
- free 360% viewpoint
- Consistent4D: Consistent 360° Dynamic Object Generation from Monocular Video
- no need for multi-view data collection and camera calibration
- to train DyNeRF
3.2.1. VIDEO NERF
- SceNeRFlow: Time-Consistent Reconstruction of General Dynamic Scenes (dynamic-NeRF method)
- correspondences even for long-term, long-range motions
- ResFields: Residual Neural Fields for Spatiotemporal Signals
- effectively represent complex temporal signals
- matrix factorization technique to reduce the number of trainable parameters
3.3. FROM TEXT
- HyperFields: Towards Zero-Shot Generation of NeRFs from Text
- distills scenes encoded in individual NeRFs into one dynamic hypernetwork
4. OPERATING UPON, EDITING
- ARF: Artistic Radiance Fields https://www.cs.cornell.edu/projects/arf/
- nerf style transfer
- Instruct-NeRF2NeRF: Editing 3D Scenes with Instructions
- training to NeRF in an iterative fashion
- integrated to nerfstudio
- FaceCLIPNeRF: Text-driven 3D Face Manipulation using Deformable Neural Radiance Fields
- train a latent code-conditional deformable NeRF, over a dynamic scene
- learns to represent a manipulated scene with spatially varying latent codes using clip
- Seal-3D: Interactive Pixel-Level Editing for Neural Radiance Fields
- preview, instantly; local pretraining and global finetuning
- proxy function mapping the editing instructions to the original space
- SIGNeRF: Scene Integrated Generation for Neural Radiance Fields
- Generatively edits NeRF scenes
- TIP-Editor: An Accurate 3D Editor Following Both Text-Prompts And Image-Prompts
- accepts text and image prompts and a 3D bounding box to specify the editing region
- Freditor: High-Fidelity and Transferable NeRF Editing by Frequency Decomposition
- lift 2D stylization results to 3D scenes
- enabling stable intensity control and novel scene transfer
4.1. INPAINTING
- Reference-guided Controllable Inpainting of Neural Radiance Fields
- use a mask and a single view image to force it on
- NeRFiller: Completing Scenes via Generative 3D Inpainting
- leveraging a 2D inpainting diffusion model
- InseRF: Text-Driven Generative Object Insertion in Neural 3D Scenes
- GO-NeRF: Generating Virtual Objects in Neural Radiance Fields
- utilizing scene context for high-quality and harmonious 3D object generation within an existing NeRF
- ReplaceAnything3D: Text-Guided 3D Scene Editing with Compositional Neural Radiance Fields
- replace while maintaining 3D consistency across multiple viewpoints
5. OUTPUT
5.1. TEXTURE
- TUVF : Learning Generalizable Texture UV Radiance Fields
- nerf baked to texture
- Blended-NeRF: Zero-Shot Object Generation and Blending in Existing Neural Radiance Fields
- change nerf objects, multiple nerf views
- Mesh Neural Cellular Automata, instead of uv map, 3d texture feel
5.2. NERF FACE
- NeRSemble: Multi-view Radiance Field Reconstruction of Human Heads
- nerf to texture to faces, very realistic (73 fps)
- fix faces generated no texture-sticking issue https://www.youtube.com/watch?v=j1ZY7LInN9g&t=272s
5.3. GRID
- Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields (house interiors)
- Progressively Optimized Local Radiance Fields for Robust View Synthesis
- dynamically allocate new local radiance
5.3.1. URBAN - CITY
- city nerf: https://city-super.github.io/gridnerf/
- Neural Fields meet Explicit Geometric Representation for Inverse Rendering of Urban Scenes
- nerf inserting 3d things
- (not nerf) UrbanIR: Large-Scale Urban Scene Inverse Rendering from a Single Video (google maps)
- City-on-Web: Real-time Neural Rendering of Large-scale Scenes on the Web
- real life stream
- VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction
- large scenes 3D Gaussian Splatting
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation
- guide the densification of the 3D Gaussians across large scenes
- Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians
- LOD-structured 3D Gaussian approach
6. FASTNESS
- parent: computerillumination
- 7.3.2
- NERF for real-time view synthesis https://arxiv.org/abs/2103.14645
- AdaNeRF https://arxiv.org/pdf/2207.10312.pdf (40 ms, two nerfs)
- VR-NeRF: High-Fidelity Virtualized Walkable Spaces (36 Hz)
6.1. GEOMETRY
- INGeo: Accelerating Instant Neural Scene Reconstruction with Noisy Geometry Priors, faster than Instant-NGP
- instant-ngp: fractions of screen-space repeatedly, all neural primitives in seconds
- PlenOctree: https://github.com/sxyu/volrend 150 fps
- Tetra-NeRF: Representing Neural Radiance Fields Using Tetrahedra
- Reconstructive Latent-Space Neural Radiance Fields for Efficient 3D Scene Representations
- nerf with autoencoder latent field, 13 times faster rendering
- VideoRF: Rendering Dynamic Radiance Fields as 2D Feature Video Streams
=best=
- streaming-rendering nerf online or mobile devices
- feature image stream can be efficiently compressed by 2D video codecs
- HybridNeRF: Efficient Neural Rendering via Adaptive Volumetric Surfaces
=surfaces instead of volumes=
, real time better speed
6.1.1. HARMONICS
- Plenoxels (plenoptic voxels), 3D grid with spherical harmonics https://arxiv.org/pdf/2112.05131
- Patch-based 3D Natural Scene Generation from a Single Example (3d patches as codebook)
- content + structure separation
- Patch-based 3D Natural Scene Generation from a Single Example (3d patches as codebook)
6.1.2. TRIANGLES
- MobileNERF = small neural network (Small MLP) for view dependant per pixel, deferred neural shader
- https://youtu.be/ofVgAEb1FiE
- https://youtu.be/nIqmuylmpFY
- 10 minutes
- mobileNERF (polygons, triangles) 124.3 fps
- Re-ReND: Real-time Rendering of NeRFs across Devices (facebook) 329.6 fps
- using rendering pipeline gpu geometry (like that one which used triangles)
- Re-ReND: Real-time Rendering of NeRFs across Devices (facebook) 329.6 fps
7. NOT NERF
7.1. 4D GENERATION
- 7.3.2.4
- 4K4D: Real-Time 4D View Synthesis at 4K Resolution (30x faster than previous)
- 4D feature grid with points naturally regularized and optimized
- learn the proposed model from RGB videos
- Motion2VecSets: 4D Latent Vector Set Diffusion for Non-rigid Shape Reconstruction and Tracking
- dynamic reconstruction from point cloud sequences
7.2. NERF ALIKES
7.2.1. VIDEO
- NeRV: Neural Representations for Videos (nerf video)
- https://github.com/haochen-rye/NeRV
- https://github.com/haochen-rye/HNeRV
- FFNeRV: Flow-Guided Frame-Wise Neural Representations for Videos
- incorporates flow information
- FFNeRV: Flow-Guided Frame-Wise Neural Representations for Videos
- Streaming Video Model
- Fast View Synthesis of Casual Videos
- synthesize high-quality novel views from a monocular video efficiently, real time
7.2.2. IMAGES
- NeuRBF: A Neural Fields Representation with Adaptive Radial Basis Functions (hd images)
- general radial bases with flexible kernel position and shape, to fit target signals
- LightSpeed: Light and Fast Neural Light Fields on Mobile Devices
- a direct mapping from a ray representation to the pixel color, neural light field using a light slab representation
7.3. NERF ALTERNATIVE
- parent: nerf
=nerf alternative=
ROOMDREAMER NOT NERF SUPERPRIMITIVE- ViewFormer: no NeRF, instead Transformers
- Geometry-Free View Synthesis: Transformers and no 3D Priors; no 3d prior
- AutoMerge: A Framework for Map Assembling and Smoothing in City-scale Environments (google maps) without gps
- M-SDF: Mosaic-SDF for 3D Generative Models
=best=
- approximates the Signed Distance Function (SDF) of shape by using set of local grids spread near the boundary
- VOLUME DIFFUSION
=best=
7.3.1. GIBR =best=
- GIBR: Denoising Diffusion via Image-Based Rendering
- IB-planes,
=new neural scene representation=
accurately represent large 3D scenes dynamically allocating more capacity as needed for details - denoising-diffusion framework to learn prior over IB-planes
- only 2D images no need for masks or depths
- single image as input, synthesises plausible details in hidden regions
- IB-planes,
7.3.2. GAUSSIAN
- 7.3.2.4
- a point cloud > gaussian cloud
- ray tracing(nerf) > ray marching
- vs mobilenerf? which seems faster with lower system requirements
- 3D Gaussian Splatting for Real-Time Radiance Field Rendering
- represent the scene with 3D Gaussians
- it has NO neural networks at all
=best nerf=
far better than instant-ngp
- Mip-Splatting: Alias-free 3D Gaussian Splatting
- smoothing filter eliminating multiple artifacts and achieving alias-free renderings
- Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering
=best=
- reduces redundant Gaussians while delivering higher-quality rendering
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (vs 3DGS)
- guide the densification of the 3D Gaussians
- RadSplat: Radiance Field-Informed Gaussian Splatting for Robust Real-Time Rendering with 900+ FPS
- GRM: Large Gaussian Reconstruction Model for Efficient 3D Reconstruction and Generation
- recovering a 3D asset from sparse-view images in around 0.1s
- reconstructs 3d gaussians-meshes from various sources: zero123++, instant3d, v3d, and sv3d
7.3.2.1. IMPROVED GAUSSIAN
- TRIPS: Trilinear Point Splatting for Real-Time Radiance Field Rendering
- approach that combines ideas from both Gaussian Splatting and ADOP(crisper images)
- real-time frame rate of 60fps
- Robust Gaussian Splatting
- fixing blur, imperfect camera poses, color inconsistencies(caused by ambient light, shadows)
- GAUSSIAN QUANTIZATION
- FisherRF: Active View Selection and Uncertainty Quantification for Radiance Fields using Fisher Informations
- view selections at 70~fps, better quality
- by leveraging fisher information, longer needing density distribution assumptions
- FisherRF: Active View Selection and Uncertainty Quantification for Radiance Fields using Fisher Informations
- MEMORY
- GES: Generalized Exponential Splatting for Efficient Radiance Field Rendering
- requiring far fewer particles to represent a scene, half the memory seed up to 40%
- GES: Generalized Exponential Splatting for Efficient Radiance Field Rendering
- TRAIN
- KEP-SVGP: Self-Attention through Kernel-Eigen Pair Sparse Variational Gaussian Processes
- attention kernels are in essence asymmetric, thus KEP-SVGP as attention kernel to fully characterizes the asymmetry
- KEP-SVGP: Self-Attention through Kernel-Eigen Pair Sparse Variational Gaussian Processes
7.3.2.2. 4D GAUSSIAN
- 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering
- video and in real time, 20 min
- PhysGaussian: Physics-Integrated 3D Gaussians for Generative Dynamics
- integrates Newtonian dynamics within 3D Gaussians for motion synthesis
- negates the necessity for triangle/tetrahedron meshing
- 4DGen: Grounded 4D Content Generation with Spatial-temporal Consistency
- 4D representation using dynamic 3D Gaussians, generation from images or videos
- specify geometry and motion offering superior control over content creation
- 4D Gaussian Splatting: Towards Efficient Novel View Synthesis for Dynamic Scenes
- 4DGS, anisotropic 4D XYZT Gaussian
- modeling complicated dynamics and fine details, especially for scenes with abrupt motions
- Spacetime Gaussian Feature Splatting for Real-Time Dynamic View Synthesis
- 3d gaussians enhanced with temporal opacity and parametric motion/rotation
- replaces spherical harmonics with neural features, so small size and fast at 60 FPS
- GaussianFlow: Splatting Gaussian Dynamics for 4D Content Creation
- smooth and natural, even in highly dynamic regions, no artifacts
- MESH CONTROL
- SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering
- hybrid gaussian-mesh for easy animation by manipulating the mesh
- GaMeS: Mesh-Based Adapting and Modification of Gaussian Splatting
- hybrid of mesh and gaussian, that pin all gaussians splats on the object surface (mesh)
- allowing for adjustments in position, scale, and rotation during animation
- hybrid of mesh and gaussian, that pin all gaussians splats on the object surface (mesh)
- 7.3.2.3.2.2
- Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling
- 2D CNNs(StyleGAN-based) and 3D Gaussian splatting
- Gaussian Frosting: Editable Complex Radiance Fields with Real-Time Rendering
- extracting a base mesh from gaussians; the fuzzier the material, the thicker the frosting
- editing and animation by modifying the mesh
- SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering
7.3.2.3. HUMAN BODY
- GPS-Gaussian: Generalizable Pixel-wise 3D Gaussian Splatting for Real-time Human Novel View Synthesis
- lift 2D parameter maps(depth estimation) to 3D space
- HiFi4G: High-Fidelity Human Performance Rendering via Compact Gaussian Splatting
- gaussian from non-rigid tracking
- compression rate of approximately 25 times, less than 2MB of storage per frame
- GPS-Gaussian: Generalizable Pixel-wise 3D Gaussian Splatting for Real-time Human Novel View Synthesis
- synthesizing novel views of any unseen characters instantly without any fine-tuning or optimization
- Human101: Training 100+FPS Human Gaussians in 100s from 1 View (Gaussian Animation method)
- BODY AVATAR
- GSM: Gaussian Shell Maps for Efficient 3D Human Generation
- 3D Gaussian rendering primitives for controllable poses and diverse appearances
- 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting
- HAHA: Highly Articulated Gaussian Human Avatars with Textured Mesh Prior
- model learns to apply Gaussian splatting only in areas of mesh where it is necessary
- like hair and out-of-mesh clothing
- so it can handle the animation of small body parts such as fingers
- model learns to apply Gaussian splatting only in areas of mesh where it is necessary
- GSM: Gaussian Shell Maps for Efficient 3D Human Generation
- GAUSSIAN FACE
- Relightable Gaussian Codec Avatars
- high-fidelity relightable(real time) head avatars, eye reflections, animated to novel expressions
- Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians
- controllable 3D Gaussians, webcam controlled expressions
- PSAvatar: A Point-based Morphable Shape Model for Real-Time Head Avatar Creation with 3D Gaussian Splatting
- parametric morphable for poses and expressions
- MAGICMIRROR
- MagicMirror: Fast and High-Quality Avatar Generation with a Constrained Search Space
- conditional NeRF and stable diffusion geometric prior
- creation of custom avatars with unparalleled quality and better adherence to input text prompts
- MagicMirror: Fast and High-Quality Avatar Generation with a Constrained Search Space
- SPLATTING AVATAR
- SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting
- disentangle the motion and appearance of a virtual human
- control the rotation and translation of the Gaussians directly by mesh
- SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting
- Relightable Gaussian Codec Avatars
7.3.2.4. GAUSSIAN GENERATION
- GaussianDreamer: Fast Generation from Text to 3D Gaussian Splatting with Point Cloud Priors
- 3D diffusion makes point cloud priors and then 2D model enriches the geometry and appearance, 25 min
- D3GA: Drivable 3D Gaussian Avatars
=best=
- multi-view videos as input
- LGM: Large Multi-View Gaussian Model for High-Resolution 3D Content Creation
=best=
- 3D models from text prompts or single-view images, 5 seconds
- GaussianDreamer: Fast Generation from Text to 3D Gaussians by Bridging 2D and 3D Diffusion Models
- 3D diffusion model provides priors for initialization, 2D model enriches the geometry and appearance; 15 minutes
- FDGAUSSIAN
- FASTER
- DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation (1 min)
- gaussian splatting with mesh extraction and texture refinement in uv space
- high-quality textured meshes, just 2 minutes from a single-view image
- FSGS: Real-Time Few-shot View Synthesis using Gaussian Splatting
- real-time and photo-realistic synthesis with three training views
- DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation (1 min)
- LUCIDDREAMER
- LucidDreamer: Domain-free Generation of 3D Gaussian Splatting Scenes
- LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching
- interval-based score matching to counteract over-smoothing
- generated from any text or image prompt (pseudo-depth alignment algorithm)
- incorporated 3D Gaussian Splatting
- project a portion of point cloud to the desired view and provide the projection
- painted images are lifted to 3D space with estimated depth maps, composing a new points
- DreamPolisher: Towards High-Quality Text-to-3D Generation via Geometric Diffusion
- coarse 3D generation refined via geometric optimization
- then ControlNet driven refiner coupled with the geometric consistency to improve texture and consistency
- coarse 3D generation refined via geometric optimization
- LucidDreamer: Domain-free Generation of 3D Gaussian Splatting Scenes