MONOCULAR SURFACE NORMAL ESTIMATION 2026

TransNormal-2

Geometry-Grounded Rectified Flow with Edge-Aware
Decoding for Precise Normal Estimation

Mingwei Li1,2 Yi Yang1 Hehe Fan1

1 College of Artificial Intelligence, Zhejiang University 2 Zhongguancun Academy

Explore the results Project repository arXiv paper

SEE THE GEOMETRY

One image. A different perspective.

Drag to compare
Input RGB: double-walled glassINPUT RGB
01 / 06Double-walled glassIn the wild
vs. TransNormal-2
TransNormal-2 predicted surface normals for double-walled glass
MoGe-2 predicted surface normals for double-walled glass
MoGe-2TRANSNORMAL-2

Transparent surfaces, recovered from a single RGB image.

Selected qualitative examples from the manuscript and supplementary material. Choose a baseline, input RGB, or available ground truth; TransNormal-2 stays on the right.

1 step

Deterministic rectified-flow inference

7

General and transparent-object benchmarks

122K

Task-specific training samples

4.2° ↓

ClearGrasp MAE vs. strongest prior baseline

METHOD OVERVIEW

A unified approach
to precise normals.

TransNormal-2 combines single-step rectified flow, geometry-aware supervision, and RGB-guided geometric refinement to address VAE reconstruction degradation.

Angular geometry, wavelet edges, and inverse-rendering self-consistency complement latent MSE with supervision in pixel space. Latent correction and the lightweight Geometric Refinement Module (GRM) work with the predictor to improve decoded normals and recover boundary detail.

Single-step predictionGeometry-aware objectivesRGB-guided refinement
TransNormal-2 pipeline showing RGB encoding, single-step latent prediction, VAE decoding and geometric refinement
FLUX.2-based rectified flow with geometry-aware supervision and post-decode correction.

QUANTITATIVE RESULTS

Across scenes. Through glass.

Strong general-scene performance, with the clearest gains on transparent objects.

Selected methods from the manuscript. Mean angular error in degrees; lower is better.
MethodClearGrasp SyntheticTN-Syn SyntheticClearPose Real-world
MoGe-226.66.236.2
FE2E16.921.222.2
Lotus-215.55.523.4
TransNormal16.13.925.5
TransNormal-2 Ours11.33.619.1

MAE improvements over the strongest prior baselines: 4.2° on ClearGrasp and 3.1° on ClearPose. Full comparisons and evaluation protocols are reported in the paper.

FOLLOW THE PROJECT

More is on the way.

The arXiv preprint, inference code, model weights, and visual results are available now. Training code will follow in a later release.

Follow on GitHub Model weights
Project page & visual resultsAvailable
arXiv paperAvailable
Inference codeAvailable
Model weightsAvailable
Training codePlanned

REFERENCE

Cite this work.

Preprint: arXiv:2609.06665.

@misc{li2026transnormal2,
  title = {TransNormal-2: Geometry-Grounded Rectified Flow with Edge-Aware Decoding for Precise Normal Estimation},
  author = {Mingwei Li and Yi Yang and Hehe Fan},
  year = {2026},
  eprint = {2609.06665},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2609.06665}
}