From Synthesis to Removal: Physics-Grounded Reflection Simulation and Diffusion-Based Video Dereflection

Zepeng Wang; Jiagao Hu; Fuhao Li; Yuxuan Chen; Fei Wang; Daiguo Zhou

Abstract

Videos captured through glass often contain reflections that degrade visual quality and interfere with downstream vision tasks. Although single-image reflection removal has been extensively studied, video reflection removal remains largely underexplored due to the lack of paired video data, temporally coherent removal models, and dedicated evaluation benchmarks. We present S2R, a closed-loop framework that unifies physics-grounded reflection simulation, diffusion-based video dereflection, and benchmark evaluation. S2R-Synthesis generates realistic paired videos through structure-space Physics-Grounded Augmentation and a trained video diffusion renderer. S2R-Removal adapts a pretrained video diffusion prior through reflection-aware latent adaptation and one-step pixel-geometric refinement. S2R-Benchmark supports both full-reference evaluation and real-world human perceptual assessment.

Overview of controllable reflection synthesis and temporally coherent reflection removal — **Fig 1.** Overview of the reflection synthesis and removal pipeline. Left: Controllable synthesis across glass roughness, reflectance, and thickness. Right: Qualitative removal results on in-the-wild videos.

S2R-Synthesis

S2R-Synthesis decouples reflection structure from appearance. A structure-guided video renderer learns realistic reflection appearance from FLUX-generated pseudo videos, while Physics-Grounded Augmentation controls reflection geometry and major glass properties in structure space.

Comparison between S2R-Synthesis and a UE5-based reflection synthesis pipeline — **Fig 3.** Comparison of reflection-synthesis trends between S2R-Synthesis (with PGA) and a UE5-based pipeline under controlled variation of three glass properties (roughness, reflectance, thickness).

ID	Setting	PSNR ↑	ΔPSNR	SSIM ↑	ΔSSIM
A: Synthetic data source
A1	Base only	28.05	+0.00	0.9465	+0.0000
A2	+ Trad.	31.98	+3.93	0.9619	+0.0154
A3	+ Ours(P)	32.35	+4.30	0.9643	+0.0178
A4	+ Trad. + Ours(P)	33.26	+5.21	0.9679	+0.0214
B: PGA variants
B1	P	32.35	+0.00	0.9643	+0.0000
B2	P + St	32.81	+0.46	0.9670	+0.0027
B3	P + St + Pa	33.62	+1.27	0.9691	+0.0048
B4	P + St + Pa + Re	33.58	+1.23	0.9695	+0.0052
B5	P + St + Pa + Re + Th	33.70	+1.35	0.9697	+0.0054
B6	Full PGA	34.13	+1.78	0.9704	+0.0061

Table 1. Effectiveness of our synthesis pipeline. RDNet is trained on the OpenRR-1k train split and evaluated on its test split. Trad. = traditional synthesis. P/St/Pa/Re/Th/Ro = Planar/Static/Partial/Reflectance/Thickness/Roughness. Δ is relative to the first row of each panel.

S2R-Removal

S2R-Removal uses two-stage adaptation: Stage I learns reflection-aware latent features with residual-derived intensity supervision, and Stage II performs one-step pixel-geometric refinement with reconstruction, structural, and depth consistency losses.

Comparison with prior video reflection removal methods — **Fig 5.** Qualitative comparison with prior video reflection removal methods.

Comparison with the general-purpose video editing model Kling O1 — **Fig 6.** Qualitative comparison with the general-purpose video editing model Kling O1 on three representative cases.

ID	Stage-I		Stage-II			PSNR ↑	SSIM ↑
ID	ℒ_diff	ℒ_int	ℒ_rec	ℒ_ssim	ℒ_depth	PSNR ↑	SSIM ↑
1	✓	—	—	—	—	27.00	0.8348
2	✓	✓	—	—	—	27.71	0.8430
3	✓	✓	✓	—	—	28.21	0.8513
4	✓	✓	✓	✓	—	28.55	0.8566
5	✓	✓	✓	✓	✓	28.84	0.8594

Table 2. Ablation study of S2R-Removal on S2R-Ref. We progressively enable the Stage-I latent adaptation losses and the Stage-II pixel-geometric refinement losses.

S2R-Benchmark

S2R-Benchmark is the first benchmark dedicated to video reflection removal. It combines paired evaluation with real-world perceptual assessment to measure reconstruction fidelity, temporal consistency, reflection removal quality, and transmission preservation.

S2R-Ref

60 paired videos

Reflected and clean sequences with identical virtual camera motion for full-reference PSNR, SSIM, and temporal consistency evaluation.

S2R-Real

50 real-world videos

In-the-wild reflections spanning diverse glass materials, lighting, reflection strengths, and camera motions for human perceptual evaluation.

Method	Video Bench						Efficiency
	S2R-Ref (60)			S2R-Real (50)			832 × 480
	PSNR ↑	SSIM ↑	TC ↑	Removal ↑	Preserv. ↑	TC ↑	ms/frame ↓
DSRNet (ICCV'23)	23.78	0.857	0.9546	0.163	0.785	0.9790	329.23
DSIT (NeurIPS'24)	24.06	0.870	0.9585	0.297	0.853	0.9781	241.05
RDNet (CVPR'25)	24.71	0.870	0.9564	0.175	0.867	0.9797	145.13
DAI (AAAI'26)	—	—	—	0.332	0.871	0.9777	217.27
GenSIRR (CVPR'26)	27.04	0.846	0.9648	0.654	0.831	0.9747	7214.94
Ours	28.84	0.859	0.9740	0.787	0.980	0.9827	87.09

Table 3. Comparison with state-of-the-art dereflection methods on video benchmark.

Method	Image Bench
	Real (20)		SIR² (454)		Nature (20)		Average
	PSNR ↑	SSIM ↑	PSNR ↑	SSIM ↑	PSNR ↑	SSIM ↑	PSNR ↑	SSIM ↑
DSRNet (ICCV'23)	23.91	0.818	25.71	0.906	25.22	0.832	25.62	0.899
DSIT (NeurIPS'24)	25.22	0.836	26.43	0.911	26.77	0.847	26.40	0.905
RDNet (CVPR'25)	25.71	0.850	26.69	0.908	26.31	0.846	26.63	0.903
DAI (AAAI'26)	25.21	0.841	27.47	0.919	26.81	0.843	27.35	0.913
GenSIRR (CVPR'26)	27.58	0.881	28.08	0.937	27.34	0.840	28.03	0.931
Ours	28.23	0.881	28.83	0.925	27.89	0.883	28.77	0.921

Table 4. Comparison with state-of-the-art dereflection methods on image benchmarks.

Qualitative Comparison

Qualitative comparison on the Real, SIR squared, and Nature image benchmarks — **Fig 9.** Qualitative comparison on the Real, SIR², and Nature image benchmarks.

**Fig 10.** Qualitative comparison between Stage I and Stage II of S2R-Removal.

Citation

@article{wang2026s2r,
  title   = {From Synthesis to Removal: Physics-Grounded Reflection Simulation and Diffusion-Based Video Dereflection},
  author  = {Wang, Zepeng and Hu, Jiagao and Li, Fuhao and Chen, Yuxuan and Wang, Fei and Zhou, Daiguo},
  journal = {arXiv preprint arXiv:2608.11562},
  year    = {2026}
}

views: —