FlashNormal: Detailed Surface Normal Estimation from Flash and No-Flash Images

Ruiyang Chen; Feiran Li; Heng Guo; Zhanyu Ma

cs.CVAug 26, 2026

FlashNormal: Detailed Surface Normal Estimation from Flash and No-Flash Images

Authors: Ruiyang Chen, Feiran Li, Heng Guo, Zhanyu Ma

Organizations: Pattern Recognition and Intelligent System Laboratory, School of Artificial Intelligence, Beijing University of Posts and Telecommunications (BUPT), Beijing 100080, China · independent researcher

Abstract

High-quality surface normal estimation is preferred for detailed surface shape recovery and image editing. Existing single image-based methods, though being a practical setup, often struggle to recover fine surface details and are sensitive to inherent shape-reflectance ambiguity. While photometric stereo achieves high-fidelity surface normal estimation from images under varying lights, its applicability is strictly limited by requiring a multi-illumination capture setup. To this end, we propose FlashNormal, a diffusion-based surface normal estimator from flash/no-flash image pairs. While retaining high practicability on modern smartphones, our proposal takes advantage of flash-induced shading variations, and leverages curvature-guided detail enhancement strategy, improving surface detail recovery and mitigating shape-reflectance ambiguity effectively. To evaluate our proposed method, we further present EvalFlash, the first real-world flash/no-flash evaluation dataset containing 20 objects aligned with ground-truth surface normals for quantitative benchmarking. Extensive experiments demonstrate the effectiveness of FlashNormal over state-of-the-art single image-based methods and show a significant out-performance over flash/no-flash-based normal estimation method on EvalFlash.

Figures & tables

Fig. 2: Pipeline of FlashNormal: Flash/no-flash image pairs and a coarse shape initialization are individually encoded by the VAE and then concatenated into a latent representation, which serves as guidance for normal estimator. The normal estimator denoises the latent representation to generate the latent surface normal, which is then decoded to produce the final estimated surface normal map.

Fortune bag	Pineapple	Lady	Pinkcat	Guanyu

Egyptcat	Shakespeare	Monkey	Tree	Monster

Hugo	Anderson	Happydog	Monkeyhead	Luxun

Sittingbuddha	Goethe	Lion	Persimmon	Toy

Fig. 4: Overview of our EvalFlash benchmark. We illustrate the GT surface normals, flash (top), and no-flash (bottom) images of EvalFlash.

Fig. 5: Capture setup of EvalFlash. (Top) Flash/no-flash capture setup and 3D scanner used for GT shape acquisition. (Bottom) Flash/no-flash image observations and the aligned surface normal map rendered from the mesh.

Panel of Figure 5 — Fig. 6: Four examples from our rendered photorealistic training dataset Flash100K. We show four 3D assets and the corresponding HDR environment maps in the first row and the bottom $4$ rows show the rendering data.

Panel of Figure 6 — Fig. 7: Surface normal estimation of Pineapple and Lady from baseline methods and ours. We show the estimated surface normals and the corresponding angular error distributions with MAE in degrees shown on the top.

	EvalFlash			EvalFlash-synth
Method	MAE	RMSE	MSE	MAE	RMSE	MSE
E2E-FT [ 28 ]	16.02	0.33	0.04	17.50	0.38	0.06
Metric3Dv2 [ 29 ]	16.37	0.34	0.04	16.14	0.35	0.05
StableNormal [ 15 ]	18.53	0.36	0.05	18.48	0.38	0.06
LX18 [ 24 ]	38.83	0.88	0.27	45.39	1.02	0.38
SDM-UniPS [ 13 ]	21.89	0.44	0.07	26.32	0.52	0.11

TABLE I: Comparison On EvalFlash and EvalFlash-synth . We Calculate the Average Values Over the Two Test Sets.

Method	Fortune bag	Pineapple	Lady	Pinkcat	Guanyu	Eygptcat	Shakespeare	Monkey	Tree	Monster
E2E-FT [ 28 ]	15.36 / 0.32	14.45 / 0.29	19.84 / 0.45	17.22 / 0.35	20.02 / 0.41	15.09 / 0.32	16.42 / 0.34	14.52 / 0.29	18.71 / 0.40	12.91 / 0.26
Metric3Dv2 [ 29 ]	14.74 / 0.31	16.03 / 0.32	18.75 / 0.44	20.31 / 0.41	14.88 / 0.31	12.00 / 0.27	18.70 / 0.39	15.54 / 0.31	15.48 / 0.33	16.44 / 0.33
StableNormal [ 15 ]	15.50 / 0.32	18.27 / 0.35	18.15 / 0.42	20.45 / 0.39	19.67 / 0.38	15.19 / 0.31	20.82 / 0.42	17.33 / 0.33	18.10 / 0.35	15.99 / 0.31
LX18 [ 24 ]	31.70 / 0.70	32.10 / 0.74	40.98 / 0.86	41.92 / 0.86	36.90 / 0.77	40.26 / 0.88	33.49 / 0.75	29.53 / 0.70	36.43 / 0.81	35.61 / 0.83
SDM-UniPS [ 13 ]	21.95 / 0.46	18.51 / 0.36	24.01 / 0.51	36.66 / 0.68	29.79 / 0.58	17.73 / 0.38	19.51 / 0.41	16.54 / 0.33	23.00 / 0.46	16.67 / 0.33
MV20 [ 19 ]	25.01 / 0.55	33.19 / 0.69	45.92 / 0.86	44.00 / 0.83	33.42 / 0.68	42.05 / 0.82	33.85 / 0.68	31.76 / 0.67	34.96 / 0.71	38.30 / 0.75

TABLE II: Quantitative Comparison (MAE/RMSE) on Surface Normal Estimation Between Baseline Methods and Ours. The Best and Second-best Results Are Emboldened and Underlined, Respectively. These Are the

10

Objects Results from EvalFlash.

Method	Hugo	Anderson	Happydog	Monkeyhead	Luxun	Sittingbuddha	Goethe	Lion	Persimmon	Toy
E2E-FT [ 28 ]	14.87 / 0.30	17.69 / 0.36	14.00 / 0.29	14.15 / 0.29	15.69 / 0.33	13.15 / 0.27	17.65 / 0.37	18.20 / 0.37	14.58 / 0.32	15.51 / 0.34
Metric3Dv2 [ 29 ]	18.77 / 0.38	19.00 / 0.38	13.28 / 0.28	13.74 / 0.27	18.89 / 0.38	15.69 / 0.31	19.59 / 0.39	15.64 / 0.32	13.12 / 0.29	16.91 / 0.36
StableNormal [ 15 ]	20.41 / 0.39	21.33 / 0.42	19.48 / 0.38	17.45 / 0.34	19.84 / 0.38	17.89 / 0.36	20.89 / 0.41	18.11 / 0.37	17.42 / 0.37	18.44 / 0.39
LX18 [ 24 ]	30.66 / 0.62	31.61 / 0.64	28.17 / 0.61	49.03 / 0.89	28.82 / 0.63	38.09 / 0.73	30.10 / 0.63	43.89 / 1.04	48.35 / 1.12	50.77 / 1.13
SDM-UniPS [ 13 ]	23.62 / 0.46	21.79 / 0.44	21.02 / 0.43	14.55 / 0.29	17.87 / 0.36	22.61 / 0.43	21.70 / 0.43	22.86 / 0.47	17.02 / 0.36	31.34 / 0.61
MV20 [ 19 ]	33.59 / 0.85	41.31 / 0.96	36.74 / 0.90	43.29 / 0.98	32.41 / 0.93	52.38 / 1.11	34.93 / 0.89	33.09 / 0.69	40.67 / 0.81	47.46 / 0.90

TABLE III: Quantitative Comparison (MAE/RMSE) on Surface Normal Estimation Between Baseline Methods and Ours. The Best and Second-best Results Are Emboldened and Underlined, Respectively. These are the other

10

objects results from EvalFlash.

Method	Fortune bag	Pineapple	Lady	Pinkcat	Guanyu	Eygptcat	Shakespeare	Monkey	Tree	Monster
E2E-FT [ 28 ]	0.03	0.03	0.07	0.08	0.05	0.03	0.04	0.03	0.05	0.02
Metric3Dv2 [ 29 ]	0.04	0.03	0.06	0.06	0.03	0.02	0.05	0.03	0.04	0.04
StableNormal [ 15 ]	0.04	0.05	0.10	0.07	0.06	0.03	0.05	0.03	0.06	0.06
LX18 [ 24 ]	0.18	0.27	0.34	0.24	0.29	0.35	0.27	0.22	0.33	0.28
SDM-UniPS [ 13 ]	0.05	0.06	0.09	0.15	0.11	0.06	0.06	0.04	0.06	0.04
MV20 [ 19 ]	0.10	0.14	0.24	0.19	0.12	0.23	0.16	0.14	0.15	0.19

TABLE IV: Quantitative Comparison (MSE) on Surface Normal Estimation Between Baseline Methods and Ours. The Best and Second-best Results Are Emboldened and Underlined, Respectively.These Are the

10

Objects Results from EvalFlash.

Method	Hugo	Anderson	Happydog	Monkeyhead	Luxun	Sittingbuddha	Goethe	Lion	Persimmon	Toy
E2E-FT [ 28 ]	0.03	0.04	0.03	0.03	0.04	0.02	0.04	0.04	0.03	0.04
Metric3Dv2 [ 29 ]	0.04	0.04	0.03	0.02	0.05	0.03	0.05	0.03	0.03	0.04
StableNormal [ 15 ]	0.04	0.05	0.03	0.03	0.04	0.04	0.05	0.07	0.05	0.07
LX18 [ 24 ]	0.21	0.26	0.19	0.24	0.22	0.33	0.24	0.29	0.30	0.33
SDM-UniPS [ 13 ]	0.07	0.08	0.05	0.02	0.06	0.05	0.06	0.07	0.05	0.12
MV20 [ 19 ]	0.24	0.26	0.25	0.26	0.28	0.31	0.29	0.17	0.23	0.27

TABLE V: Quantitative Comparison (MSE) on Surface Normal Estimation Between Baseline Methods and Ours. The Best and Second-best Results Are Emboldened and Underlined, Respectively. These Are the Other

10

Objects Results from EvalFlash.

Panel of Figure 12 — Fig. 8: Qualitative comparison with baseline methods in terms of shape-reflectance ambiguity. Our method can estimate high-quality ambiguous-less surface normal maps, leading to visibly flatter meshes after surface normal integration.

Panel of Figure 13 — Fig. 9: Estimated surface normal maps and reconstructed meshes of detail-rich objects. Compared to the peer methods, our estimated surface normals present richer details, leading to higher-quality structured meshes.

Panel of Figure 14 — Fig. 10: Mesh generation results of Shakespeare (left side) and Monkey (right side) by image-to-3D methods [ 48 , 49 ] and ours. The 3D mesh visualizations are taken as snapshots.

Panel of Figure 15 — Fig. 11: Ablation on geometric detail recovery from different setups.

Panel of Figure 16 — Fig. 12: Ablation study on flash input.

Panel of Figure 17 — Fig. 13: Ablation study on coarse shape initialization. Though the initial surface normal is lack of detail, it can help enhance the accuracy of estimated surface normal from FlashNormal.

	EvalFlash			EvalFlash-synth
Method	MAE	RMSE	MSE	MAE	RMSE	MSE
FlashNormal (Ours)	14.96	0.31	0.03	15.47	0.33	0.04
w/o flash input	15.48	0.32	0.03	15.88	0.35	0.05
w/o curvature-guided detail enhancement	15.50	0.33	0.04	16.07	0.35	0.05
w/o zoomed pixels strategy	15.60	0.32	0.04	17.96	0.38	0.06
w/o global shape initialization	16.33	0.34	0.04	18.23	0.39	0.07

TABLE VI: Ablation study on EvalFlash and EvalFlash-synth . We calculated the average values of two testsets.

$\lambda$	EvalFlash			EvalFlash-synth
	MAE	RMSE	MSE	MAE	RMSE	MSE
10	15.46	0.34	0.11	16.12	0.35	0.05
1000	15.23	0.33	0.05	15.98	0.34	0.05
100 (used in paper)	14.96	0.31	0.03	15.47	0.33	0.04

TABLE VII: Sensitivity study on the loss weight

\lambda

. We report the average values on EvalFlash and EvalFlash-synth.

Prompt variant	EvalFlash			EvalFlash-synth
	MAE	RMSE	MSE	MAE	RMSE	MSE
w/o “DSLR”	15.16	0.31	0.03	15.73	0.34	0.04
w/o “detailed”	15.21	0.32	0.03	15.66	0.34	0.05
full prompt (used in paper)	14.96	0.31	0.03	15.47	0.33	0.04

TABLE VIII: Sensitivity study on prompt variants. We report the average values on EvalFlash and EvalFlash-synth.

Panel of Figure 21 — Fig. 14: Comparison between E2E-FT [ 28 ] and ours about zoomed pixels strategy. Valid regions are resized to the same resolution.

Panel of Figure 22 — Fig. 15: Analysis on the robustness against varying focal length $f$ and camera-to-object distance $z$ . The top and middle four rows show the image observations and estimated surface normal maps. Valid regions are resized to the same resolution. The bottom row summarizes the normal estimation errors measured by MAE under varying $f$ and $z$ .

Fig. 16: Analysis on the robustness w.r.t. varying illuminations & color temperatures. The top three rows show the synthetic experimental results and bottom two rows show the real-world experimental results.

Panel of Figure 24 — Fig. 17: Representative ROI extraction and surface normal prediction results using SAM 2 [ 40 ] . The first three rows are single-object cases, and the last two rows are multi-object cases.

Panel of Figure 25 — Fig. 18: Qualitative result using an in-the-wild image processed by StableDelight [ 52 ] .

Panel of Figure 26 — Fig. 19: 3D reconstruction via integrating multi-view surface normal maps estimated by E2E-FT [ 28 ] and ours. The 3D mesh visualizations are taken as snapshots.

Panel of Figure 27 — Fig. 20: Failure cases from our proposal. Red boxes highlight areas where surface normals are incorrect.

Surface Normal Estimation

Surface Normal Estimation recovers the orientation of visible 3D surfaces from images and other sensor data for shape understanding and reconstruction.

See all 25 papers

Poppy: Polarization-based Plug-and-Play Guidance for Enhancing Monocular Normal Estimation
Mar 29, 2026Irene Kim, Sai Tanmay Reddy Chakkera, Alexandros Graikos +2Surface Normal Estimation Test-Time Optimization
Stony Brook University
GeoStereo: A Unified Stereo Geometry Estimation Framework for Disparity and Surface Normal
Jul 27, 2026Qizhe Wei, Xianda Guo, Shaocong Xu +3Surface Normal Estimation Stereo Matching
Beijing Institute of Technology BAAI Beijing, China · School of Computer Science, Wuhan University Wuhan, China · BAAI Beijing, China +3
Physics-informed Active Polarimetric 3D Imaging for Specular Surfaces
Feb 23, 2026Jiazhang Wang, Hyelim Yang, Tianyi Wang +13D Reconstruction Surface Normal Estimation
Wyant College of Optical Sciences, University of Arizona, Tucson, AZ 85721

Mar 29, 2026J/K move · Enter open · S save

Poppy: Polarization-based Plug-and-Play Guidance for Enhancing Monocular Normal Estimation

Irene Kim, Sai Tanmay Reddy Chakkera, Alexandros Graikos, Dimitris Samaras+1

Stony Brook University

Surface Normal Estimation Test-Time Optimization

Open paper AskPDF arXiv

Monocular surface normal estimators trained on large-scale RGB-normal data often perform poorly in the edge cases of reflective, textureless, and dark surfaces. Polarization encodes surface orientation independently of texture and albedo, offering a physics-based complement for these cases. Existing polarization methods, however, require multi-view capture or specialized training data, limiting generalization. We introduce Poppy, a training-free framework that refines normals from any frozen RGB backbone using single-shot polarization measurements at test time. Keeping backbone weights frozen, Poppy optimizes per-pixel offsets to the input RGB and output normal along with a learned reflectance decomposition. A differentiable rendering layer converts the refined normals into polarization predictions and penalizes mismatches with the observed signal. Across seven benchmarks and three backbone architectures (diffusion, flow, and feed-forward), Poppy reduces mean angular error by 23-26% on synthetic data and 6-16% on real data. These results show that guiding learned RGB-based normal estimators with polarization cues at test time refines normals on challenging surfaces without retraining.

Abstract

Surface Normal Estimation

Explore similar work

Poppy: Polarization-based Plug-and-Play Guidance for Enhancing Monocular Normal Estimation

GeoStereo: A Unified Stereo Geometry Estimation Framework for Disparity and Surface Normal

Physics-informed Active Polarimetric 3D Imaging for Specular Surfaces