3De Lens: Neural Compositing & Visualization

• 3De Lens is a family of techniques that combines parameter-free neural lens blur compositing for mixed reality with focus+context visualization using 3D and decal lenses.
• The neural pipeline estimates the circle-of-confusion directly from RGB images, enabling realistic integration of virtual objects without relying on explicit camera metadata.
• The visualization approach fuses volumetric 3D and surface-following decal lenses to render multi-geometry data seamlessly in immersive virtual environments.

Searching arXiv for the named topic and closely related papers to ground the article. I’ll check for arXiv entries matching “3De Lens” and the cited titles. Searching arXiv for “3De Lens” and “Parameter-Free Neural Lens Blur Rendering for High-Fidelity Composites”. 3De Lens denotes a small but heterogeneous family of lens-centered methods in contemporary arXiv literature. In the most explicit usage, it names a parameter-free neural lens blur compositing pipeline for mixed reality and 3D object insertion that estimates a circle-of-confusion (CoC) map directly from RGB, infers virtual-object blur through a linear signed CoC–disparity relation, and renders the result with neural reblurring (Ruan et al., 21 Nov 2025). In a distinct visualization usage, “3De lens” denotes a focus+context technique that fuses volumetric 3D lenses with surface-following decal lenses for multi-geometry data in virtual reality (Mota et al., 14 Sep 2025). The available literature therefore suggests that the term is best understood as a cross-domain label rather than a single standardized optical device.

1. Terminological scope

Two usages are primary in the available corpus. One belongs to computational photography and mixed-reality compositing; the other belongs to immersive scientific visualization.

Usage	Domain	Core idea
3De Lens	Mixed reality / 3D object insertion	Parameter-free neural lens blur compositing from RGB-derived CoC
3De lens	Virtual-environment visualization	Fusion of 3D and Decal lenses for multi-geometry focus+context

In the mixed-reality formulation, the central problem is realistic defocus transfer: a virtual object inserted into a photographed scene must inherit blur that is consistent with the photograph even when camera metadata, calibration, and explicit scene depth are unavailable. The method therefore bypasses classical metadata-driven CoC computation and estimates CoC directly from RGB before fitting blur for the inserted object (Ruan et al., 21 Nov 2025).

In the visualization formulation, the central problem is different. Multi-geometry scenes frequently contain both surfaces and non-surface data such as streamlines, volumes, or point clouds. A single lens is constructed as a pair of containment lenses so that one component can operate volumetrically and the other can conform to surfaces, thereby avoiding the need for multiple separate tools (Mota et al., 14 Sep 2025).

A broader visualization survey is consistent with this multiplicity. It defines a visualization lens as “an interactively parametrizable spatial selection according to which a base visualization is modified to provide an alternative visual representation of the data in focus,” and organizes spatially embedded lenses along 7 dimensions: Position, Orientation, and Scale; Shape; Dimensionality; Effect Scope; Effect Imagery; Effect Encoding; and Viewpoint Dependency (Mota et al., 30 Mar 2025). This suggests that “3De Lens” is best treated as a family resemblance term spanning different technical problems rather than a single canonical architecture.

2. Parameter-free neural lens blur compositing

In "Parameter-Free Neural Lens Blur Rendering for High-Fidelity Composites" (Ruan et al., 21 Nov 2025), 3De Lens is a parameter-free neural lens blur compositing pipeline for mixed reality. Its premise is that lens blur varies with scene depth, focus distance, and aperture, and that a sharp inserted object is visually inconsistent unless its blur matches the photographed background.

The paper starts from the thin-lens CoC relation. If $z$ is scene depth, $z_1$ is focus distance, $A$ the aperture diameter, and $f$ the focal length, then

$$c = A\frac{\left|z_2-z_1\right|}{z_2}\frac{f}{z_1-f} \approx Af\frac{\left|z_2-z_1\right|}{z_2z_1} = Af\left|\frac{1}{z_1}-\frac{1}{z_2}\right|.$$

The practical difficulty is that ordinary photographs often lack the required depth/disparity and camera parameters.

The defining change is to estimate the CoC map directly from the RGB image: $F_I = \mathbb{F}(I).$ The method then reformulates blur using disparity $d$, yielding

$c = Af|d-d_1|,$

and introduces a signed form,

$c_s = B(d-d_1),$

where $c_s$ is the signed CoC, $z_1$0 is a scale factor, and $z_1$1 is the focus disparity. The signed formulation encodes whether a point lies in front of or behind the focal plane.

A second difficulty is that a background CoC estimate cannot simply be copied onto a new virtual object. The method addresses this by fitting a linear relation between the signed CoC of the real image and the disparity of the virtual object inside a valid overlap region, the projection mask $z_1$2: $z_1$3 Here $z_1$4 is the observed CoC value from the real photograph, $z_1$5 is the virtual-object disparity at the corresponding pixel, and $z_1$6 are fitted linear parameters. The paper describes this fit as the bridge between real and virtual content (Ruan et al., 21 Nov 2025).

The final composite is produced not by Gaussian blur but by a neural reblurring network. The full pipeline has four parts: estimate CoC from the real RGB image, fit a linear CoC–disparity model, create a CoC map for the virtual object, and perform neural lens reblurring. The rendered composite is written as

$z_1$7

where $z_1$8 is the object mask, $z_1$9 is the object depth/disparity, $A$0 is the object RGB image, $A$1 is the background photo, and $A$2 is the linear CoC regression module (Ruan et al., 21 Nov 2025).

3. Empirical behavior of the compositing pipeline

The 3De Lens compositing paper separates blur estimation from reblurring and evaluates both. For CoC estimation, direct RGB-to-CoC estimation outperforms depth-based conversion. For reblurring, the neural network-based method outperforms Gaussian blur. The paper reports that the direct CoC estimator achieves better RMSE than the depth-based baseline, and that the neural reblurring method gives better PSNR/SSIM than Gaussian reblurring (Ruan et al., 21 Nov 2025).

On synthetic scenes with ground truth, the reported composite quality improvement over Blind Augmentation is substantial: PSNR is 34.66 vs 27.88, and SSIM is 0.97 vs 0.82. The qualitative discussion emphasizes object boundaries and objects with spatially varying blur, including the example of a horse whose head is in focus while its tail is blurred (Ruan et al., 21 Nov 2025).

The perceptual evaluation follows the same direction. A user study with 25 participants found that the proposed method was significantly harder to distinguish from real content than the baselines, and the failure rate for detecting the synthetic object was highest for the proposed method (Ruan et al., 21 Nov 2025). The paper also states that the learned linear mapping can be reused for other objects or later video frames, which positions the method not only as a single-image compositor but also as a reusable scene-consistency mechanism.

This formulation is “parameter-free” in a precise sense: the real photograph does not require camera metadata, calibration, or explicit scene depth. A plausible implication is that the method is primarily aimed at the common case of uncontrolled images, such as internet photographs or ordinary user-captured content, where classical optics-based blur rendering pipelines are difficult to apply.

4. 3De lens in immersive visualization

In "3De Interactive Lenses for Visualization in Virtual Environments" (Mota et al., 14 Sep 2025), 3De lens is a focus+context visualization technique for multi-geometry data. It combines two lens categories: 3D lenses, which operate on subvolumes and are suited to non-surface data, and Decal lenses, whose active region is the intersection of a sphere with a surface and which are suited to surface data.

The formal model places the scene in $A$3 and defines a lens $A$4 centered at $A$5 with radius $A$6. The lens region is a ball $A$7. The lens itself is defined as

$A$8

The 3D component $A$9 uses $f$0 as the selection region and displays a non-surface focus attribute $f$1. It also defines a disk $f$2 with center $f$3, radius $f$4, and normal vector $f$5 for orientation and optional angular selection. The decal component $f$6 is defined by the surface patch

$f$7

which forms the active area for a surface focus attribute $f$8 (Mota et al., 14 Sep 2025).

The interaction model is tightly coupled to VR. The implementation uses Unity and HTC Vive, with direct spatial manipulation of the lens as a single entity. Grabbing moves the lens center $f$9; while grabbing and intersecting a surface, the decal component is active; when the button is released, the 3D component becomes active for streamlines passing through the lens region. A disk-like widget supports orientation control, and only streamlines approximately aligned with the disk normal are selected, with an angular tolerance of 15 degrees. Lens radius can be changed through a separate scaling widget (Mota et al., 14 Sep 2025).

The rendering pipeline is correspondingly specialized. Surface depiction uses a two-pass silhouette algorithm and Fresnel-based opacity,

$$c = A\frac{\left|z_2-z_1\right|}{z_2}\frac{f}{z_1-f} \approx Af\frac{\left|z_2-z_1\right|}{z_2z_1} = Af\left|\frac{1}{z_1}-\frac{1}{z_2}\right|.$$0

with reported good results at $$c = A\frac{\left|z_2-z_1\right|}{z_2}\frac{f}{z_1-f} \approx Af\frac{\left|z_2-z_1\right|}{z_2z_1} = Af\left|\frac{1}{z_1}-\frac{1}{z_2}\right|.$$1 for the surface and $$c = A\frac{\left|z_2-z_1\right|}{z_2}\frac{f}{z_1-f} \approx Af\frac{\left|z_2-z_1\right|}{z_2z_1} = Af\left|\frac{1}{z_1}-\frac{1}{z_2}\right|.$$2 for the lens surface. Streamlines are rendered with piecewise view-aligned billboarding: $$c = A\frac{\left|z_2-z_1\right|}{z_2}\frac{f}{z_1-f} \approx Af\frac{\left|z_2-z_1\right|}{z_2z_1} = Af\left|\frac{1}{z_1}-\frac{1}{z_2}\right|.$$3 The implementation achieves about 45 FPS on a desktop with Intel Core i3 and GeForce GTX Titan 6G GPU, and was demonstrated on examples including 15K surface triangles with 2K streamlines, and a wind-turbine dataset with 362K surface triangles and 5K streamlines (Mota et al., 14 Sep 2025).

This usage sits naturally within the broader design-space literature on spatially embedded lenses. The survey paper cited above treats 3D lenses, 2.5D lenses, and decal imagery as distinct design choices, and identifies dynamic shapes, semi-automated 3D placement, and viewpoint-sensitive behavior as underexplored regions (Mota et al., 30 Mar 2025). An earlier geometric focus+context line of work likewise reframed lens design as 3D mesh deformation followed by controlled flattening, using energies such as

$$c = A\frac{\left|z_2-z_1\right|}{z_2}\frac{f}{z_1-f} \approx Af\frac{\left|z_2-z_1\right|}{z_2z_1} = Af\left|\frac{1}{z_1}-\frac{1}{z_2}\right|.$$4

to minimize distortion around focus and context regions (Li et al., 2013). The VR 3De lens can therefore be read as one concrete realization within a broader research program on spatially embedded lenses.

5. Optical 3D lens systems in adjacent literature

Although not all of the following systems are named “3De Lens,” they define the optical context in which 3D lensing problems are posed. One important line concerns switchable optical power for AR/VR and 3D display. "Concave/convex switchable lens using active phase-change material Ge$$c = A\frac{\left|z_2-z_1\right|}{z_2}\frac{f}{z_1-f} \approx Af\frac{\left|z_2-z_1\right|}{z_2z_1} = Af\left|\frac{1}{z_1}-\frac{1}{z_2}\right|.$$5Sb$$c = A\frac{\left|z_2-z_1\right|}{z_2}\frac{f}{z_1-f} \approx Af\frac{\left|z_2-z_1\right|}{z_2z_1} = Af\left|\frac{1}{z_1}-\frac{1}{z_2}\right|.$$6Te$$c = A\frac{\left|z_2-z_1\right|}{z_2}\frac{f}{z_1-f} \approx Af\frac{\left|z_2-z_1\right|}{z_2z_1} = Af\left|\frac{1}{z_1}-\frac{1}{z_2}\right|.$$7" proposes a wavezone lens made from phase-change material features rather than fixed-curvature glass. With optical constants $$c = A\frac{\left|z_2-z_1\right|}{z_2}\frac{f}{z_1-f} \approx Af\frac{\left|z_2-z_1\right|}{z_2z_1} = Af\left|\frac{1}{z_1}-\frac{1}{z_2}\right|.$$8 in the amorphous state and $$c = A\frac{\left|z_2-z_1\right|}{z_2}\frac{f}{z_1-f} \approx Af\frac{\left|z_2-z_1\right|}{z_2z_1} = Af\left|\frac{1}{z_1}-\frac{1}{z_2}\right|.$$9 in the crystalline state, the same structure behaves as a convex lens with $F_I = \mathbb{F}(I).$0 in one state and as a concave lens with $F_I = \mathbb{F}(I).$1 in the other, under FDTD simulation at $F_I = \mathbb{F}(I).$2 (Shi et al., 2018). The paper also states that one switchable lens can perform a negative or positive focal length, or perform negative and positive focal lengths simultaneously.

Another line addresses vergence–accommodation conflict in stereoscopic display. "Dynamic lens and monovision 3D displays to improve viewer comfort" inserts Optotune variable-focus lenses in front of each eye so that accommodation demand tracks the vergence distance implied by stereo disparity. In the reported setup, the display viewing distance is 1.77 m, corresponding to 0.57 D, and the dynamic lens produces an accommodative range of 0.48–3.2 m, corresponding to 2.06–0.31 D. In disparity detection, average threshold was 1.8 arcmin with the dynamic lens versus 2.5 arcmin in the fixed-lens control; at +1.5 D disparity, time-to-fuse was 0.96 s for the dynamic-lens condition versus 3.2 s for the fixed-lens condition (Johnson et al., 2015). The same study found that monovision did not provide the same benefit.

A third line uses electrically tunable liquid-crystal optics for autostereoscopic 2D/3D switching. "Enhancing Optical Performance of Liquid Crystal Lens Arrays via Electrode Design Optimization" reports a large-aperture LC lens array with lens pitch $F_I = \mathbb{F}(I).$3, cell gap $F_I = \mathbb{F}(I).$4, and oxide layer thickness $F_I = \mathbb{F}(I).$5. The proposed structure achieves a shortest focal length of 3.3 mm at $F_I = \mathbb{F}(I).$6 and $F_I = \mathbb{F}(I).$7, with a reported theoretical focal length of 3.27 mm and $F_I = \mathbb{F}(I).$8 near the optimal simulated operating point (Ding et al., 4 Sep 2025). Because the aperture is 800 $F_I = \mathbb{F}(I).$9m or more, the paper explicitly links the design to more pixels per lens, more viewing zones, and improved multiview 3D performance.

Taken together, these papers define a concrete optical background for 3D lens research: switchable sign of focal length, dynamic synchronization of accommodation and vergence, and electrically tuned refractive-index distributions are all mechanisms for controlling depth-dependent image formation in three-dimensional display systems.

6. Associated research directions and extensions

The literature surrounding 3De Lens also includes several adjacent directions in which “lens” no longer refers only to a conventional refractive element. One example is structural wave control. "Structural Lens Based on Variable Thickness Structures" designs a thin plate with continuously varying thickness so that it can focus a plane wave into three spots or split waves from a point source into three collimated beams. It uses the effective refractive-index relation

$d$0

examines designs with $d$1 and $d$2 for $d$3 mm and $d$4 mm, and reports broadband behavior from 50 kHz to 200 kHz (Zhao et al., 2022).

A second extension removes the lens altogether. "Coded Illumination for 3D Lensless Imaging" replaces the lens with a coded mask and improves 3D reconstruction by using multiple coded illumination patterns and a projector–camera baseline. The separable forward model is

$d$5

and with coded illumination becomes

$d$6

The prototype uses a Sony MP-CL1 laser projector, a Sony IMX183 sensor, 60 $d$7m mask features, an effective measurement size of 512 × 512, and a 55 mm baseline, with best reported results under 48 shifting lines illumination (Zheng et al., 2022).

A third extension treats the lens as a differentiable component in an end-to-end optimization pipeline. "Neural Lens Modeling" introduces NeuroLens, an invertible residual-network lens model for distortion and vignetting that supports both point projection and ray casting. It parameterizes distortion with an i-ResNet, uses a network with width 1024 and four residual blocks, and models vignetting through

$d$8

The associated SynLens benchmark is assembled from 400 lenses selected from the Lensfun database (Xian et al., 2023). In a more explicitly task-driven optical design setting, "The Differentiable Lens: Compound Lens Search over Glass Surfaces and Materials for Object Detection" optimizes surface curvatures, spacings, and glass choices jointly with RetinaNet using differentiable spherical lens simulation. On BDD100K at 1× resolution, joint optimization yields 32.0 AP for a Doublet, 33.3 AP for a Cooke, and 33.6 AP for a Tessar, despite the possibility of degraded conventional image quality (Côté et al., 2022).

Additional neighboring work expands the physical and fabrication envelope of 3D lens research. "Diffractive flat lens enables Extreme Depth-of-focus Imaging" reports an MDL with 1.8 mm aperture at $d$9, maintaining focus from 5 mm to 1200 mm and describing a depth-of-focus enhancement of approximately $c = Af|d-d_1|,$0 (Banerji et al., 2019). "Superlensing Microscope Objective Lens" combines a conventional objective with a BaTiO$c = Af|d-d_1|,$1-PDMS microsphere superlens through a 3D-printed lens adaptor, reporting a resolution of about 100 nm under white light illumination (Yan et al., 2016). "The lens was fabricated by fluidic shaping" uses a 3D-printed border, PDMS injection, and curing under free-energy minimization to fabricate spherical and free-form lenses, reporting maximum profile error 6.1 μm for the spherical lens, AFM roughness $c = Af|d-d_1|,$2 nm, and maximum contour error 13.8 μm for the free-form case (Cheng et al., 2024).

These neighboring directions do not collapse into a single definition of 3De Lens. Rather, they indicate a shared research trajectory in which depth-dependent image formation, focus+context control, and optical behavior are increasingly treated as programmable, differentiable, or geometry-coupled phenomena across computational photography, immersive visualization, display optics, wave physics, and fabrication.