Lens Flare Removal and Reconstruction

Tarun Yenamandra; Jonathon Luiten; Daniel Cremers; Nathan Matsuda

Figure 1. Lens-flare transfer to images. Given multi-view images (left column) of scenes with lens flares, our model decomposes each scene into a flare-free scene and our 3DGS-based lens-flare representation. We use the reconstructed flares to transfer them onto a diverse set of target images (top row). Each following row composites the source flare reconstructed on the left onto every target image shown above it.

Figure 2. Categories of lens flares for our discussion. Three photographs of lens flares. The left two, grouped under the label Scattering'', show a hallway ceiling light with radial streaks (labeled Streaks'') and a bright purple-white light with a hazy vertical beam (labeled Glare''). The right photo, labeled Reflective'', shows a bright light on a wooden deck railing at night with a distinct circular reflective ghost artifact nearby.

	Stock flares	Procedural	Physically-based	Ours
No lens/optical model needed	✓	✓	✗	✓
No precise 3D light geometry needed d	✓	✓	✗	✓
Viewpoint-correct optical appearance a	✗	✗	✓	✓
Reproduces the target scene’s own captured flare signature	✗	✗	✗	✓
3D-consistent transfer of a captured flare to new scenes c	✗	✗	✗	✓
Real-time rendering b	✓	✓	✓	✓

Table 1. Comparison against a stock (real, but unrelated) captured flare, procedural (2D/game-engine) simulators, and physically-based (ray-traced) simulators, the latter two used as forward renderers configured by hand or from a lens prescription rather than fitted to the capture. a Ghost/streak shapes correctly reflowing with viewpoint change; physically-based methods achieve this only when the lens prescription matches the real lens, which generic models miss (unit-specific imperfections, dust, coatings). For our method this holds when re-rendering the reconstructed scene; in transfer, appearance is queried at the source viewpoint by design (Sec. 5.4 ). b Real-time procedural and GPU physically-based demos exist, but high-fidelity physically-based rendering (diffraction, multi-bounce) is typically offline. c A stock flare is a real capture, but of an unrelated lens/light, and is pasted in 2D without adapting to the target camera’s geometry. d Our method needs only an approximate 3D light position, estimated automatically, and is robust to large errors in it (Sec. 5.5 ).

Figure 3. Varying lens flares due to lenses and light sources from our dataset. Three photographs illustrating flare variation. The first two, under the label

Same light, different lenses'', show a tree-trunk scene and a chair scene each with a different lens flare pattern despite similar lighting. The third photo, under the label

Different lights, same lens'', shows a distinct starburst-shaped light source flare.

Figure 4. Reconstructions of a 3D scene with lens flares using 3DGS ( Kerbl et al., 2023 ) , and Deformable 3DGS ( Yang et al., 2023 ) . As lens flares are not 3D-consistent, the methods struggle to reconstruct them. Three near-identical close-up photographs of a bright headlamp strapped to a tree trunk, labeled Input'', 3DGS'', and ``Def-3DGS''. The Input photo shows a small blue circular flare artifact beside the light, marked with a green circle. In the 3DGS and Def-3DGS reconstructions, this same location is marked with a red circle, and the blue flare artifact is missing or much fainter, showing that both 3D reconstruction methods fail to reproduce it.

Figure 5. Our lens flare removal model (middle). We fine-tune the diffusion model from Difix3D+ ( Wu et al., 2025 ) with LoRA ( Hu et al., 2022 ) , enabling one-step flare removal. We train two models, Ours and Ours-wGlare, without lens glare and with lens glare, respectively, as shown above (right, target supervision). We note that while the model with light glare might be subjectively preferred by an end user, the one without light glare is more realistic. To train our models, we add a flare image from our flare datasets to an image from the training dataset, yielding a corrupted input, and train them using the loss from Eq. ( 2 ). A diagram of the lens flare removal pipeline. On the left, two example inputs are formed by adding a flare image (a starburst light) to a clean image (a lake at dusk, or a tree against the sky) in linear space, producing a flare-corrupted image. These inputs pass through a VAE encoder, a UNet, and a VAE decoder, each augmented with orange LoRA weight blocks and skip connections, to produce flare-removed outputs. On the right, the outputs are compared via a loss function against two target images: one with the light source's glare removed, and one with it retained.

Figure 6. Our lens flare representation model (left, top) along with a 3DGS model(left, bottom). We represent flares as Gaussians with 1D locations. Given input camera and light positions, we deform the Gaussians and place them along a line connecting the image’s principal point to the light. We render the scene and flare Gaussians together (right, top) and supervise them using the provided scene images. We render the scene using Gaussians (right, middle) and supervise using images from our flare-removal model. A diagram illustrating the flare representation pipeline. On the left, colored 2D Gaussian ellipses represent flare components and scene Gaussians, arranged along and around a diagonal line on a camera-frustum plane labeled ``Line-based 3DGS Deformations''. On the right, three photographs of a streetlamp scene at night show the rendered composite of scene and flare Gaussians (top), the scene rendered alone (middle), and a faint flare-only render (bottom).

Figure 7. Geometric construction of a single flare. Each flare component is represented by a set of canonical 1D Gaussians with scalar mean

\mu\in\mathbb{R}

, encoding the radial offset along the image-plane line through the principal point

p

and the (corrected) light projection

l+\delta_{l}

. At render time, the scalar

\mu

is mapped to a 2D image point via

\mu^{\prime}\big((l+\delta_{l})-p\big)+p

, then unprojected one near-plane offset in front of the scene to obtain a standard 3DGS primitive in the camera frame. A diagram showing three stages of the flare geometry construction, left to right. First,

Canonical Flare Gaussians in 1D'': a row of pink ellipses of varying size and orientation arranged along a single axis, with scalar mean mu, fed into a

Deformations'' box conditioned on principal point p and light position l. Second,

Flare Gaussians in 2D'': the deformed ellipses arranged along a diagonal dashed line inside a square, connecting a black dot labeled

Principal point p'' to a yellow dot labeled Corrected light l plus delta-l''. Third, Flare Gaussians in 3D'': a camera icon projects the same line through an image plane onto a near-plane offset in front of a scene panel, placing the flare Gaussians as standard 3DGS primitives along the optical axis between the camera and the light source.

Figure 8. Workshop test view: the generating light source lies outside the image, yet our reconstructed flare combined with the scene reproduces the ground truth. See Sec. 5.3 for further experimental details and results. A strip of four images of a workshop tool cabinet labeled Ground Truth'', Render'', Scene'', and Flare''. Ground Truth and Render show near-identical photographs of the cabinet with a faint blue-green flare glow entering from the left edge. The Scene image shows the cabinet without any flare, appearing cleaner and slightly darker on the left. The Flare image is a mostly dark frame showing only the isolated bluish-green flare glow, with the light source itself outside the visible frame.

Figure 9. Examples of real large lens flares from our dataset compiled using videos from ( ActionVFX, 2026 ) (left, middle), and procedural ring-shaped reflective flares (right). Three dark photographs of large reflective lens flares. The first two, under

Full image real flares'', show a soft glowing oval shape filling most of a dark frame and a vertical greenish-blue streak with a bright spot at its base. The third, under

Ring flares'', shows a procedurally generated green ring-shaped flare with opacity fading from the ring toward the center of the image.

									Ablations on Ours-wGlare
Benchmark	Metric	LightsOut	FR	Flare7K++	ACL-FR	FlareX	Ours-wGlare	Ours	wo	wo	wo
		( Tsai et al., 2025 )	( Wu et al., 2021 )	( Dai et al., 2023a )	( Zhou et al., 2025 )	( Lishen et al., 2025 )			light	enc FT	VFX
Flare7k++	PSNR $\uparrow$	16.04	24.49	26.42	24.48	25.26	26.68	25.76	25.84	25.64	26.96
	SSIM $\uparrow$	0.6954	0.8797	0.8926	0.8689	0.8870	0.8993	0.8925	0.8831	0.8696	0.8850
	LPIPS $\downarrow$	0.2686	0.0980	0.0908	0.0925	0.0934	0.0777	0.0848	0.0850	0.0856	0.0767
vfx_dataset	PSNR $\uparrow$	13.04	21.04	21.37	20.20	24.17	25.93	27.06	25.03	23.75	24.77

Table 2. Quantitative comparisons (left) and ablations (right) on flare removal. On each benchmark, one of our two models (Ours-wGlare and Ours, with and without glare) outperforms all baselines on every metric. LPIPS uses the VGG backbone. Since Flare7k++ contains residual light glare, Ours-wGlare performs best. Our models train in

20

hours, whereas baselines take

>4

days. Ablations: Performance on the Flare7k++ benchmark degrades without light sources in target images or without VAE encoder fine-tuning (since flare-corrupted images are OOD). Performance on the VFX benchmark degrades without VFX training data, which, in exchange, slightly improves Flare7k++ PSNR and LPIPS. Best results among the compared methods are shown in red bold , second-best in blue underline ; ablations are not ranked.

Figure 10. Qualitative comparison of different flare removal models, from left to right, given image, LightsOut ( Tsai et al., 2025 ) , FR ( Wu et al., 2021 ) , ACL-FR ( Zhou et al., 2025 ) , Flare7K++ ( Dai et al., 2023a ) , FlareX ( Lishen et al., 2025 ) , Ours-wGlare, Ours, and ground truth. In the top five rows, we show the results on the Flare7K++ benchmark. As shown, our models consistently remove flares without residual flares and more effectively recover the light source in the image, leading to improved quantitative performance. In the last five rows, we show the results from our VFX benchmark. As shown, existing models mistakenly bake large flare corruptions into the sky, whereas our model performs much better at removing them. A grid of ten image strips, each row showing the same scene processed by different flare-removal methods side by side: the flare-corrupted input, then LightsOut, FR, ACL-FR, Flare7K++, FlareX, our model with lens glare, our model without lens glare, and the ground truth, with a PSNR value overlaid on each result. The top five rows are scenes from the Flare7K++ benchmark with small scattering flares; baseline outputs retain visible residual glare and streaks, while our two models' outputs closely resemble the flare-free ground truth. The bottom five rows are scenes from the VFX benchmark with large reflective flares covering much of the sky; baseline methods leave large corrupted regions, while our models recover plausible sky and scene detail.

Scene	Vanilla 3DGS			w BilateralGrid			w Flare7k++			w FR			w ACL-FR			w FlareX			w Ours-wGlare			w Ours
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
amp	12.25	0.518	0.599	14.40	0.637	0.413	17.24	0.714	0.435	18.50	0.689	0.450	16.75	0.669	0.455	17.03	0.681	0.440	19.84	0.743	0.379	21.06	0.751	0.377
billboard	18.93	0.645	0.327	19.59	0.707	0.273	19.16	0.656	0.345	19.84	0.686	0.319	18.61	0.668	0.336	19.43	0.680	0.301	20.67	0.740	0.256	20.47	0.705	0.286
dog	21.77	0.869	0.185	21.74	0.903	0.146	27.22	0.931	0.118	25.88	0.917	0.149	25.01	0.888	0.173	25.44	0.916	0.135	28.42	0.945	0.090	28.48	0.943	0.089
guitar	20.07	0.835	0.255	20.03	0.866	0.242	24.53	0.903	0.171	25.65	0.905	0.173	22.64	0.870	0.214	24.97	0.900	0.176	26.21	0.911	0.151	25.72	0.910	0.162
plant	20.36	0.836	0.226	19.05	0.820	0.266	23.83	0.872	0.192	23.40	0.848	0.218	22.53	0.848	0.215	21.32	0.815	0.258	21.06	0.831	0.252	23.91	0.874	0.186

Table 3. Quantitative comparison of performance of different flare removal models on 3D reconstruction. We evaluate the flare removal performance across multiple views of FR ( Wu et al., 2021 ) , Flare7k++ ( Dai et al., 2023a ) , ACL-FR ( Zhou et al., 2025 ) , FlareX ( Lishen et al., 2025 ) , Ours and Ours-wGlare by training a 3DGS model on the different sets of flare-removed images. Additionally, we compare with BilateralGrid ( Wang et al., 2024 ) and vanilla 3DGS ( Kerbl et al., 2023 ) . While the models are trained on flare-removed images, they are evaluated on actual captured images of the scenes without processing, which happen to not contain flares (see Sec. 5.2 ). Both our models outperform all baselines on average; on plant, Ours-wGlare falls below four of the six baselines on every metric, while Ours remains best. Best results are shown in red bold , second-best in blue underline .

Figure 11. Qualitative comparisons of ablations of our flare removal model, Ours-wGlare, on Flare7K++ benchmark (top), and VFX benchmark (bottom). Left to right, we show the input, ‘wo enc FT’, where we do not finetune the VAE encoder, ‘wo light’, where a light source is added after flare removal, ‘wo VFX data’, where the model is not trained on the VFX dataset, and ‘Ours-wGlare’ and ‘Ours’. As shown, without encoder finetuning, the model struggles to recover color characteristics. Post-processing the output to add a light source does not accurately reproduce it. Without the VFX dataset, the model struggles to remove large flares. Finally, Ours-wGlare and Ours models perform the best on both the benchmarks. Two rows of image strips comparing ablations of the flare removal model on a Flare7K++ scene (top) and a VFX scene (bottom). Each row shows, left to right: the flare-corrupted input, wo enc FT' (VAE encoder not fine-tuned, showing color shifts), wo light' (light source added back after removal, appearing less accurate), `wo VFX data' (large flares only partially removed), the full model with lens glare, and the full model without lens glare. The two rightmost results in each row most closely match a flare-free, correctly colored scene.

Figure 12. Qualitative comparison of results from 3D reconstruction in the presence of lens flares. From left to right, we show the results from 3DGS on input images, trained with BilateralGrid ( Wang et al., 2024 ) , flare removed images from Flare7K++ ( Dai et al., 2023a ) , from FR ( Wu et al., 2021 ) , from ACL-FR ( Zhou et al., 2025 ) , from FlareX ( Lishen et al., 2025 ) , from Ours-wGlare, and from Ours. As shown, our models achieve the best reconstruction, leading to fewer artifacts. This implies that our models are more consistent across different views than existing models. A grid of 3D-scene renderings across five scenes (rows: plant, amp, billboard, guitar, dog), comparing, left to right, vanilla 3DGS trained on raw flare-corrupted input images, 3DGS with BilateralGrid, and 3DGS trained on images pre-processed by Flare7K++, FR, ACL-FR, FlareX, our model with lens glare, and our model without lens glare. The leftmost columns show visible flare-shaped artifacts baked into the 3D reconstruction, while the rightmost two columns (our models) show cleaner reconstructions with fewer flare-related artifacts.

Flare - Decomposed
	Design alternatives									Component ablations
Scene	Ours 2D			Ours sup in-flare			Ours unsup			Ours wo $\delta_{l}$			Ours wo G-SAM2			Ours wo $\delta_{f}$			Ours
Scene	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
hat	31.14	0.850	0.299	28.01	0.659	0.358	30.29	0.720	0.301	37.22	0.873	0.271	35.75	0.862	0.286	36.93	0.867	0.268	37.24	0.876	0.273
park	24.93	0.508	0.355	18.79	0.666	0.307	20.90	0.704	0.285	32.24	0.809	0.269	32.07	0.785	0.276	32.14	0.839	0.244	32.36	0.820	0.258
outtrunk	27.96	0.855	0.291	16.87	0.446	0.516	19.78	0.511	0.427	33.03	0.876	0.294	30.52	0.867	0.303	32.59	0.871	0.294	33.22	0.877	0.293

Table 4. Quantitative comparison on decomposable flare reconstruction. Vanilla 3DGS reconstructs the composite image but produces no flare/scene decomposition, so it appears only in the Scene & Flare sub-table. We group the remaining variants into two families. Design alternatives each replace a core design decision: ‘Ours 2D’ drops the 1D constraint and lets the flare Gaussians move freely on the plane (similar to Deformable 3DGS ( Yang et al., 2023 ) ); ‘Ours sup in-flare’ and ‘Ours unsup’ replace our scene supervision with the input minus the reconstructed flare, and with no decomposition supervision at all, respectively. ‘Ours 2D’ is trained against the same reference as ‘Ours’; it drops the 1D constraint, and with it the finer deformations

\delta_{f}

, which are redundant once the Gaussians move freely on the plane: it matches ‘Ours’ on scene decomposition but loses

5.5

dB on flare, which we attribute to the 1D representation. The two supervision alternatives collapse on both (

11.5

–

14.9

dB on flare,

10.8

–

13.8

dB on scene) because their scene branch retains the flare; this margin is modestly inflated, since unlike the other variants they are not supervised with the reference (Sec. 5.3 ). Component ablations remove one element of the same design and therefore land within

1.1

dB of ‘Ours’: ‘Ours wo

\delta_{l}

’ drops the light-position correction and performs on par with ‘Ours’, since the finer deformations

\delta_{f}

compensate for light-position error; only removing both breaks robustness (Sec. 5.5 ); ‘Ours wo G-SAM2’ replaces detection with a noisy scene centroid, showing the line equation is recovered automatically; ‘Ours wo

\delta_{f}

’ removes the finer deformations and cannot represent asymmetric flares. ‘Ours unsup’ attains the highest Scene & Flare PSNR while decomposing poorly; we treat this as an operating-mode trade-off rather than a deficiency (Sec. 5.3.1 , ‘Operating modes’). Best and second-best are marked across all variants within each sub-table, in red bold and blue underline .

Figure 13. Qualitative comparison ( decos : top, workshop : middle, hat : bottom) of flare representation and decomposition with baselines (3DGS), and ablations. We show the results of 3DGS on the left, followed by our ablation models: ‘Ours 2D’, ‘in-flare’, ‘unsup’, ‘wo

\delta_{l}

’, ‘wo G-SAM2’, ‘wo

\delta_{f}

’, and ‘Ours’. We are interested in a model that provides the best decomposition of lens flares and 3D scenes. 3DGS cannot reconstruct flares because they are not 3D-consistent. ‘Ours 2D’, being unconstrained on the flare locations, struggles to reconstruct them. ‘in-flare’ supervises the scene to reconstruct everything except what the flare model models. ‘unsup’ leads to inconsistent decompositions as the scene is unsupervised. ‘wo

\delta_{l}

’ performs on par with our model, since the finer deformations absorb small light-location errors. ‘wo G-SAM2’ still recovers the line equation, as

\delta_{l}

and

\delta_{f}

absorb the error of the centroid-based light estimate. In ‘Ours wo

\delta_{f}

’, the lens flares lack asymmetric details. Finally, our model leads to better decompositions and flare modeling. ‘Ref.’ denotes the reference images the metrics use, derived from the captured input and our removal output (Sec. 5.3 ). A grid of rendered scene comparisons across three scenes (rows: decos, workshop, hat), each showing 3DGS alongside our ablation variants, left to right: 3DGS, Ours 2D', in-flare', unsup', wo delta-l', wo G-SAM2', wo delta-f', and

Ours', plus a reference column. The 3DGS column shows scenes with reflective flares either missing or baked in as artifacts, while later columns progressively better isolate and reconstruct distinct flare shapes near each scene's light source(s), with the final

Ours' column most closely matching the reference decomposition into scene and flare components.

Figure 14. Lens-flare transfer to a novel scene. We transfer flares reconstructed from three source scenes (left column) – two from our captured reconstruction dataset (Sec. 5.3 ) and one additional captured scene, also used for source flares in Fig. 1 – onto a single target scene that does not appear in either of our captured datasets (Sec. 4 ), rendered from three novel views (top row). Each following row composites the corresponding source flare onto every view of the target scene shown above it. The bottom row additionally shows one of the transferred flares with its Gaussians uniformly scaled

2\times

, demonstrating simple post-hoc artistic control over the flare’s apparent intensity and extent. A grid transferring three source flares onto a target scene of an apartment courtyard with grass and a lamp post, shown from three novel viewpoints (top row, no flare). The three source flares (left column) are a bright deck-railing light with colored ghost artifacts, a sun behind a building corner with a hazy halo, and a headlamp on a desk with a bright starburst. Each subsequent row composites one source flare onto the three target views, adding a bright sun-like glare with radial streaks near the sky in each. A final row shows one transferred flare with its Gaussians scaled up, producing a visibly larger and brighter starburst than the corresponding un-scaled version.

Figure 15. Robustness of our flare representation to error in the input light position

l

, on five scenes. PSNR against the average pixel error injected into

l

, for the composite render (solid) and the flare (dotted); colour indicates which of the light-position correction

\delta_{l}

and finer deformations

\delta_{f}

are enabled. With either correction the model degrades gracefully; with both removed, composite and flare PSNR fall by over

4

dB. A line chart titled ``Light Location Sensitivity Analysis on 5 scenes'', plotting PSNR in decibels (y-axis, 27 to 34) against average pixel error in light location (x-axis, 0 to about 260). Eight lines show four correction settings (with/without delta-l and with/without delta-f) each for Scene+Flare PSNR (solid) and Flare-only PSNR (dotted). The three settings with at least one correction enabled cluster together, declining gently from about 33.5-34 dB to about 31.5-32 dB. The setting with both corrections removed (red) declines much more steeply, from about 32.7 dB to about 27.2-28.2 dB.

Figure 16. Qualitative robustness on the hat sequence to noise in the input light-source location

l

. Top four rows: composite render (scene

+

flare); bottom four rows: isolated flare component. Rows within each group enable, in order, both corrections, only

\delta_{f}

, only

\delta_{l}

, and neither. With at least one correction the flare remains centred on the source and retains its starburst through

\sigma=0.5

, softening at

\sigma=1.0

; with neither, structure is lost from

\sigma=0.1

onward. A table of image strips on the \emph{hat} sequence, with eight rows grouped into two sets of four: the top four rows show the composite scene-plus-flare render, and the bottom four show the isolated flare component. Within each set of four, rows correspond to: both corrections enabled, only the finer deformation, only the light-position correction, and neither. Each row has four columns for increasing injected noise levels (sigma = 0, 0.1, 0.5, 1.0) plus a ground-truth column. Rows with at least one correction retain a sharp, centered starburst flare through moderate noise; the row with neither correction loses its starburst structure starting at the lowest noise level and degrades to a faint blob at the highest noise level.

Figure 17. A limitation of our model. While our model can accurately reconstruct the locations and colors of most flares, it struggles with ring-shaped flares, as Gaussians are inherently more opaque towards the center. Four images of a tree-trunk night scene with a bright light and a small blue secondary flare. The isolated Flare'' render shows a glowing ring with a dark hollow center instead of the expected ring shape. The Scene'', Render'', and GT'' (ground truth) images look nearly identical, showing the tree trunk with a bright light and, in the ground truth, a faint colorful ring artifact near the blue secondary flare that is missing from the reconstructed render.

Lens Flare Removal and Reconstruction

Abstract

Figures & tables

Appendix

Explore similar work

Breaking Spatial Uniformity: Prior-Guided Mamba with Radial Serialization for Lens Flare Removal

Semi-LAR: Semi-supervised Contrastive Learning with Linear Attention for Removal of Nighttime Flares

FujinSplat: Seeing Through Smoke with RAW-Domain Gaussian Splatting