TransNormal-2 leverages geometry‑grounded flow and edge‑aware decoding for precise normal estimation
| Source: HF Papers | Original article
Researchers introduce TransNormal-2, a geometry‑grounded rectified flow model with edge‑aware decoding that improves normal estimation by addressing VAE reconstruction degradation at object boundaries.
A new arXiv pre‑print titled **“TransNormal‑2: Geometry‑Grounded Rectified Flow with Edge‑Aware Decoding for Precise Normal Estimation”** proposes a solution to a long‑standing weakness in diffusion‑based monocular geometry models. The authors show that the 8× spatial compression performed by the VAE encoder‑decoder introduces reconstruction artefacts that blur surface‑normal predictions, especially along object edges. TransNormal‑2 tackles this by adding geometry‑aware training losses that directly penalise boundary errors and by attaching a lightweight RGB‑guided refinement module that restores fine‑grained detail after the diffusion step.
The paper’s core contribution is a single‑step, FLUX.2‑based rectified‑flow predictor that produces an initial normal map, which is then sharpened through the edge‑aware decoder. According to the authors, the approach delivers “strong results with minimal annotations,” suggesting that high‑quality normal fields can be obtained without the dense labeling typically required for supervised depth or normal networks.
Why it matters is twofold. First, precise normal estimation underpins many downstream 3‑D tasks—augmented‑reality rendering, robotic perception, and scene‑understanding pipelines all rely on accurate surface orientation. By mitigating VAE‑induced blur, TransNormal‑2 promises more reliable inputs for those applications. Second, the work builds on the diffusion‑transformer line of research we highlighted in our September 9 report on Marigold V2, which explored diffusion models for monocular depth. TransNormal‑2 extends that trajectory from depth to surface normals, showing that diffusion architectures can be refined for a broader set of geometric outputs.
Looking ahead, the community will watch for benchmark releases that compare TransNormal‑2 against established baselines such as Marigold V2 and FlowBalance. If the authors open‑source the code and pretrained weights, integration into existing 3‑D reconstruction toolkits could follow quickly. Further research may also explore whether the edge‑aware decoding strategy can be generalized to other diffusion‑based vision tasks, potentially reducing the need for heavy annotation regimes across the field.
Sources
Back to AIPULSEN