A computer-vision method that compares an object with a flash on and off reported lower surface-normal errors than the listed comparison methods on two benchmark tests. Surface normals are estimated directions showing which way different parts of an object’s surface face, a useful clue for recovering shape from images.
The method, called FlashNormal, was tested on rendered images and on a set of real objects. Its reported advantage came with important limits: the authors describe failures on transparent objects and on flat objects covered by complex realistic textures, and the study did not report confidence intervals or other formal uncertainty estimates.
The extra clue comes from the flash
FlashNormal takes a flash/no-flash image pair and a coarse initial normal map. The paired images are encoded separately with a pretrained variational autoencoder, or VAE. The paper describes the system as a diffusion-prior estimator, using a learned visual guide to estimate surface directions.
The system also gives special attention to fine shape. Its detail objective combines angular normal error with supervision from curvature maps, which highlight how sharply a surface bends. Detail regions are selected with a curvature threshold of 0.002, while cropped regions of interest are resized to 512 by 512 pixels with at least 80% of the crop occupied by the object. The loss weight for this detail term is set to 100.
The model was trained by fine-tuning Stable Diffusion v2 on 100,000 flash/no-flash image pairs. The reported setup used AdamW with a learning rate of 3e-5, batches of 256 on eight GPUs, and 10 training epochs over about two days. For comparison, single-image baselines received no-flash images, while the other methods received image pairs. The evaluations used mean angular error, root mean square error and mean square error.
Lower errors on the reported tests
On the real-object EvalFlash benchmark, FlashNormal’s reported average mean angular error was 14.96 degrees, compared with 16.02 for E2E-FT and 36.01 for MV20. Its reported root mean square error was 0.31, versus 0.33 and 0.72, while mean square error was 0.03, versus 0.04 and 0.21. The authors report performance differences of 6.6% against E2E-FT and 58.4% against MV20 on this test.
The synthetic test, EvalFlash-synth, contained 839 high-quality samples rendered from assets kept separate from the training data. FlashNormal had the smallest reported average mean angular error and root mean square error, with reported percentage differences of 4.2% against Metric3Dv2 and 72.1% against MV20.
The real-world EvalFlash set contained 20 objects captured at 8192 by 5464 pixels with the flash on and off. The objects were aligned to scanned meshes, which typically had about 102,000 vertices and 206,000 faces, and the average overlap between the estimated object masks and reference masks was reported as above 99.0% across the set.
Selected visual examples suggested that FlashNormal could produce plausible normal estimates for both shiny, or specular, objects and diffuse objects, while retaining finer geometric detail than several baselines. These examples are qualitative comparisons and were not accompanied by statistical uncertainty estimates.
Ablations point to a combined design
When the researchers removed the flash input, the curvature-guided detail enhancement or the zoomed-pixel regions, the resulting estimates were accompanied by higher mean angular error and lost geometric detail. The paper reports MAE differences of 3.4%, 3.5% and 4.1% on EvalFlash for these ablations, and describes coarse initialization as associated with convergence at 10 epochs.
The authors also report usable estimates under changing ambient-light intensity and color temperature, along with reasonable operation from a single ambient-light image. Automatic region-of-interest extraction was valid in five representative cases, including three single-object and two multi-object examples, although the paper notes lower robustness when masks fail in cluttered scenes.
A small 3D reconstruction demonstration
In a downstream multi-view test, the method was applied to 17-view integrations of two objects, Shakespeare and Monkey. Reported multi-view MAE was 17.62 degrees for FlashNormal versus 19.51 for E2E-FT on Shakespeare, a 9.7% difference, and 12.86 versus 17.83 on Monkey, a 27.8% difference. The FlashNormal meshes were also reported to show richer geometric detail.
That reconstruction result is narrow because it covers only two objects. The real benchmark contains 20 objects, and the authors report failures for transparent objects and for flat objects with complex realistic textures.
What the evidence does not settle
The benchmark reports averages and percentage differences, but gives no confidence intervals or inferential uncertainty for them. The ablation comparisons do not by themselves establish that the flash input or any one module caused the observed differences.
The manuscript is an arXiv version 1 preprint dated 26 August 2026, and journal publication or acceptance is not reported in the supplied text. Its acknowledgment lists support from the Beijing Major Science and Technology Project, the Hebei Natural Science Foundation Project, the Beijing-Tianjin-Hebei Basic Research Funding Program and the National Natural Science Foundation of China.
Paper data and sources
Original title: FlashNormal: Detailed Surface Normal Estimation from Flash and No-Flash Images
Authors: Ruiyang Chen, Feiran Li, Heng Guo, Zhanyu Ma
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text