Authors: Grzegorz Wilczyński, Mikołaj Zieliński, Krzysztof Byrski, Joanna Waczyńska, Dominik Belter, Przemysław Spurek
Organizations: Jagiellonian University, Faculty of Mathematics and Computer Science · IDEAS Research Institute · Poznań University of Technology, Institute of Robotics and Machine Intelligence
Neural Radiance Fields achieve high-fidelity scene representation but suffer from costly training and rendering, while 3D Gaussian splatting offers real-time performance with strong empirical results. Recently, solutions that harness the best of both worlds by using Gaussians as proxies to guide neural field evaluations, still suffer from significant computational inefficiencies. They typically rely on stochastic volumetric sampling to aggregate features, which severely limits rendering performance. To address this issue, a novel framework named IRIS (Intersection-aware Ray-based Implicit Editable Scenes) is introduced as a method designed for efficient and interactive scene editing. To overcome the limitations of standard ray marching, an analytical sampling strategy is employed that precisely identifies interaction points between rays and scene primitives, effectively eliminating empty space processing. Furthermore, to address the computational bottleneck of spatial neighbor lookups, a continuous feature aggregation mechanism is introduced that operates directly along the ray. By interpolating latent attributes from sorted intersections, costly 3D searches are bypassed, ensuring geometric consistency, enabling high-fidelity, real-time rendering, and flexible shape editing. Code can be found at https://github.com/gwilczynski95/iris.
Figures & tables
Figure 1: The framework introduces the Ray Intersection Selector (RIS), which eliminates redundant sampling to achieve high-throughput inference compared to volumetric baselines. This explicit anchor-based representation facilitates seamless integration with physics engines and robust part-level scene editing. Consequently, high-fidelity reconstruction is maintained.
Figure 2: Overview of the IRIS pipeline. The rendering process begins with the Ray Intersection Selector (RIS) , which analytically determines the precise intersection points between camera rays and 3D Gaussians, thereby significantly reducing the number of required samples. In the subsequent Ray-Coherent Aggregation (RCA) stage, local neural features are aggregated from these intersected primitives utilizing Mahalanobis distance-based weighting to synthesize a continuous latent vector. Shallow MLPs decode this representation into view-dependent color and material density, σ . Finally, the neural density is spatially modulated by the explicit Gaussian opacity and integrated via point-based volumetric rendering to produce the final image.
Figure 3: Qualitative comparison on indoor scenes from the Mip-NeRF 360 dataset. Reconstruction fidelity is evaluated against editable baselines: Radiance Meshes and EKS. It is observed that IRIS consistently achieves superior visual quality, effectively recovering high-frequency details that are blurred or lost in competing methods.
Figure 4: Sampling strategy comparison. The proposed Ray Intersection Selector (RIS) (left) analytically determines precise interaction points with scene primitives, generating meaningful samples in a single pass without processing empty space. In contrast, standard volumetric sampling (right) relies on a heuristic, multi-stage process (coarse and fine) that inherently allocates the computational budget to unoccupied regions.
Figure 5: Qualitative comparison against static state-of-the-art baselines on complex real-world scenes. TS denotes Triangle Splatting [ 9 ] . It is observed that IRIS achieves reconstruction fidelity highly competitive with leading static representations such as 3DGS or Mip-NeRF 360. Despite the additional architectural constraints required for editability, fine geometric details-such as the thin structures in the Bicycle scene and the surface textures of the Truck -are preserved with precision.
Figure 6: Qualitative comparison on the NeuralEditor deformation benchmark. The rendered views demonstrate the capability of IRIS to maintain high visual fidelity under significant geometric deformation. It is observed that while baselines such as GaMeS and NeuralEditor exhibit artifacts of texture degradation, IRIS produces crisp, photorealistic renderings with preserved view-dependent effects (e.g., specular highlights on the cymbals).
Deep Blending
Tanks&Temples
drjohnson
playroom
train
truck
Avg.
Static
INGP
28.26
21.67
20.46
23.38
23.44
MNeRF360
29.14
29.66
19.52
24.91
25.81
3DGS
28.77
30.04
21.10
25.19
26.27
TS
28.91
29.35
21.33
24.94
26.13
Table 1: Quantitative analysis. TS: Triangle Splatting, RM: Radiance Meshes ( ⋆ re-evaluated). IRIS achieves state-of-the-art fidelity on Deep Blending and competitive performance on Tanks&Temples, confirming its overall robustness against all baselines.
Figure 8
Model
PSNR ↑
SSIM ↑
LPIPS ↓
Training time ↓
FPS ↑
Static
InstantNGP
29.31
0.939
0.053
7min 27s
1.894
Nerfacto
29.19
0.941
0.095
9m 38s
0.860
Editable
NeuralEditor
31.41
0.955
0.036
103h 45min
0.083
EKS
33.12
0.962
0.031
26h 2min
0.183
Table 2: Averaged quantitative comparisons on the NeRF Synthetic dataset. While reconstruction fidelity remains comparable to the state-of-the-art editable baseline (EKS), a decisive advantage is demonstrated in efficiency, with significantly reduced training times and rendering speeds approximately 30 × higher compared to the EKS method.
Figure 9: Qualitative comparison on the challenging HuSc3D [ 22 ] dataset. Compared to editable baselines, IRIS effectively suppresses floating artifacts prevalent in Radiance Meshes (RM) and preserves fine textures and sharp geometry that EKS over-smooths.
Figure 10: Qualitative comparison on manual (top) and physics-based (bottom) editing. It is observed that EKS exhibits visible ghosting artifacts, whereas IRIS maintains superior fidelity and clean geometry, effectively eliminating trailing noise during motion.
Outdoor scenes
Indoor scenes
bicycle
flowers
garden
stump
treehill
Avg.
room
counter
kitchen
bonsai
Avg.
Static
INGP
22.17
20.65
25.07
23.47
22.37
22.75
29.69
26.69
29.48
30.69
29.14
Mip-NeRF 360
24.37
21.73
26.98
26.40
22.87
24.47
31.63
29.55
32.23
33.46
31.72
3DGS
25.25
21.52
27.41
26.55
22.49
24.64
30.63
28.70
30.32
31.98
30.41
Triangle Splatting
24.90
20.85
27.20
26.29
21.94
24.27
31.05
28.90
31.32
31.95
30.80
Table 3: Quantitative comparisons of reconstruction capability (PSNR) on the Mip-NeRF 360 dataset. It is observed that in indoor environments, IRIS outperforms all editable baselines and explicit static methods, being surpassed only by the implicit Mip-NeRF 360. In outdoor scenarios, while a performance gap is observed relative to Radiance Meshes, competitive fidelity is maintained against other baselines, with results significantly exceeding those of EKS. Radiance Meshes results ( ⋆ ) were re-evaluated using the official implementation.
Figure 11: Editing on large-scale scenes. Non-rigid warping (top) and object scaling (bottom) are demonstrated on Bicycle and Garden scenes. It is observed that high fidelity and geometric consistency are preserved in these challenging unbounded environments.
Figure 12: While standard NeRFs clip objects exiting the training volume, IRIS maintains full rendering integrity by dynamically sampling explicit geometry independent of static global bounds.
Figure 13: Qualitative ablation on Hotdog and Lego . It is observed that restricted sampling (16 IPR) results in significant geometric noise. While increasing the budget (64 IPR) or removing hash initialization with full budget recovers structure, the full model is required to effectively resolve high-frequency details and specular highlights.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Chair
Drums
Lego
Mic
Mater.
Ship
Hotdog
Ficus
Avg.
PSNR
16 IPR
32.39
24.81
30.69
32.87
27.18
28.35
34.60
28.05
29.87
64 IPR
34.94
25.61
35.43
35.16
29.40
30.80
36.95
31.45
32.47
w/o hash
34.70
25.91
34.87
35.79
29.35
30.58
36.68
32.47
32.54
full
35.00
25.51
35.54
35.87
29.46
30.69
37.02
32.12
32.65
Appendix
Table 4: Ablation study conducted on the static NeRF Synthetic dataset. The impact of the sampling budget is evaluated by varying the maximum number of intersections processed per ray (denoted as IPR), comparing restricted budgets of 16 and 64 against the full model’s 128. Additionally, the role of feature initialization is assessed by substituting the Hash Grid strategy with random initialization ”w/o hash”). It is observed that increasing the intersection quota significantly improves reconstruction fidelity, validating the need of sufficient depth coverage, while the Hash Grid initialization further contributes to the optimal performance of the full model.
Chair
Drums
Lego
Mic
Mater.
Ship
Hotdog
Ficus
Avg.
PSNR
Naive Plotting
24.58
21.54
25.38
27.56
21.59
22.21
26.72
24.62
24.27
Neuraleditor
25.29
21.93
27.14
27.49
23.04
24.12
27.14
24.83
25.12
GaMeS
24.51
22.02
26.65
27.07
21.73
22.19
27.26
26.65
24.76
EKS
26.03
22.08
28.04
27.85
23.14
24.43
28.23
27.58
25.92
IRIS (our)
25.97
21.72
28.16
27.46
22.89
24.16
27.93
27.34
25.70
Appendix
Table 5: Quantitative comparisons on the NeuralEditor benchmark. The evaluation is conducted on the NeRF Synthetic dataset, where ground-truth deformations are applied to the objects for rigorous comparison. It is observed that IRIS achieves reconstruction quality comparable to the state-of-the-art editable method EKS, even surpassing it on specific scenes.
Figure 14: Extended qualitative comparison (close-ups). This figure extends Fig. 5 from the main paper by providing additional crop views from scenes on the Mip-NeRF 360 and Deep Blending datasets. TS denotes Triangle Splatting. It is observed that IRIS preserves high-frequency structural details, achieving quality competitive with static state-of-the-art baselines like Mip-NeRF 360.
Figure 15: Extended qualitative comparison (Full Views). Zoomed-out visualizations from Mip-NeRF 360, Deep Blending and Tanks&Temples datasets are presented. These views demonstrate the global consistency of the reconstruction, confirming that IRIS maintains photometric accuracy across large spatial extends.
Figure 16: Additional editing results on the Mip-NeRF 360 dataset. The capability of IRIS to perform complex, topology-agnostic manipulations in large-scale unbounded environments is demonstrated. Non-rigid deformations, such as the sinusoidal warping of the bench in the Bicycle scene (top) and the rippling effect applied to the table in the Garden scene (bottom), are executed while preserving high-frequency textural details. In all cases, geometric consistency and high fidelity are successfully maintained throughout the animation sequence.
Figure 17: Physics-based simulations on IRIS representations. The explicit point clouds optimized by our method were exported to Blender to drive dynamic deformations. The Ficus sequence (top) demonstrates a gravity simulation where leaves detach and fall realistically. The Lego sequence (bottom) showcases soft-body dynamics; the truck is dropped onto a platform, exhibiting elastic deformation and bouncing behavior upon collision. These examples highlight the capability of IRIS to support complex physical interactions while maintaining rendering fidelity.
Figure 18: Qualitative results on the NeuralEditor benchmark. Comprehensive comparison across all scenes from the deformation dataset. NE denotes NeuralEditor. It is observed that IRIS maintains high visual fidelity and geometric consistency, avoiding the blurring artifacts seen in Naive and NeuralEditor baselines, while achieving quality comparable to the state-of-the-art EKS method.
Recent advances in neural scene representations enable photorealistic novel-view synthesis, yet most methods remain tightly coupled to a single rendering paradigm, limiting their versatility and integration with conventional graphics workflows. We introduce Floating Radiance Networks (FlaRe), a neural scene representation combining explicit ray-traceable geometry with continuous neural radiance functions. A scene is represented by floating planar generalized Gaussian primitives, each carrying a compact latent descriptor of a local radiance field. A lightweight decoder shared across the scene maps this descriptor, local surface coordinates, and viewing direction to color and opacity. This formulation preserves the expressiveness of neural fields while providing an explicitly addressable structure that can be efficiently queried and manipulated. Hardware-accelerated primitive intersections enable interactive rendering and recursive ray-tracing, including reflections, refractions, transparency, and shadows. The same representation further supports primitive-level deformation, mesh extraction, and appearance stylization directly in its learned descriptor space. Experiments across standard reconstruction benchmarks demonstrate competitive rendering quality while using a compact set of primitives. Together, these results establish FlaRe as a versatile representation that brings high-fidelity neural rendering, ray-tracing, geometric manipulation, and appearance editing into a unified scene model. Source code is available online. Source code can be found at: https://github.com/KByrski/FlaRe
Krzysztof Byrski, Rafał Tobiasz, Grzegorz Wilczyński +5
Jagiellonian University, Faculty of Mathematics and Computer Science · IDEAS Research Institute · Poznań University of Technology, Institute of Robotics and Machine Intelligence
3D representations are fundamental for scene rendering, understanding, and interaction. Approaches like Neural Radiation Fields and 3D Gaussian Splatting achieve impressive photorealistic novel-view synthesis, but lack the ability to easily decompose scene elements into object level, requiring additional segmentation or grouping. We present MLP-Splatting, a method that enables scene decomposition via a few expressive light-field primitives while providing photorealistic novel-view synthesis. MLP-Splatting models each primitive as an independent compact MLP with localized spatial support that predicts radiance and opacity. In contrast to low-level Gaussian primitives or a single global radiance field, our neural primitives provide greater expressive capacity while remaining spatially localized. Rendering is performed through efficient sparse volumetric compositing over ray-primitive interactions. Our primitives are supervised using RGB supervision alone, which yields primitives that represent local scene regions often corresponding to objects or object parts, enabling interactive object-level editing without segmentation masks by selecting a handful of primitives. Our method, augmented with optional semantic feature distillation, enables open-vocabulary scene interaction and open-set instant segmentation. Compared to state-of-the-art semantic 3DGS methods, we achieve substantially lower memory usage (1/7×) and faster rendering (5×). Project Page: https://shinjeongkim.com/mlp-splatting
We introduce TRON, a rendering framework that combines 3D Gaussian ray tracing with neural rendering to enable realistic and controllable rendering of real-world 3D scenes under novel lighting, dynamic object motion, object insertion, and material editing. Prior approaches that rely solely on physically based rendering (PBR) of Gaussian representations struggle to achieve realistic relighting due to imperfections in reconstructed geometry, material estimates, and light transport estimation. At the same time, neural rendering methods often lack an explicit scene representation, limiting their ability to support interactive editing with fine-grained manipulation. TRON bridges these two paradigms. We use intrinsic decomposition priors from a learned inverse rendering model to regularize the material properties of a Gaussian field, and repurpose a ray tracer to provide radiometric guidance rather than final pixels. By treating this output as a structured 3D scaffold, we empower a lightweight neural renderer to bridge the domain gap between shading-model constrained estimates and photorealistic output. Our key insight is that the combination of explicit 3D knowledge with robust material priors provides speed and controllability, while neural rendering enables the synthesis of photorealistic images. To support real-world scenarios, we train our neural renderer with a multi-stage strategy consisting of large-scale pretraining and targeted fine-tuning on a newly constructed dataset of 2.1M rendered synthetic and real-world frames from 3D reconstructions. TRON outperforms Gaussian-based relighting methods in realism, and prior neural renderers in editability and speed. To the best of our knowledge, TRON is the first method to enable practical interactive applications in captured 3D environments, offering realistic appearance under dynamic geometric, lighting and material conditions.
Or Perel, Hassan Abu Alhaija, Zian Wang +4
1NVIDIA · University of Toronto · 3Vector Institute