TransNormal-2: Geometry-Grounded Rectified Flow with Edge-Aware Decoding for Precise Normal Estimation
Mingwei Li, Yi Yang, Hehe Fan
Zhejiang University · Zhongguancun Academy
Published September 6, 2026 · on the feed September 9, 2026 · ▲ 19 · arXiv 2609.06665
The abstract
Diffusion-based models enable monocular geometry estimation, yet their pixel-space precision is limited by a shared, under-studied error source: VAE reconstruction degradation. The 8x spatial compression in the VAE encoder-decoder degrades surface normals at object boundaries; even encoding and decoding ground-truth normals introduces 1.3--8.5° of mean angular error (MAE), with edge MAE reaching 2.8x the global MAE. We present TransNormal-2, a FLUX.2-based rectified-flow framework with single-step deterministic inference that addresses this degradation on both sides of the VAE decoder: in how latent predictions are supervised during training, and in how decoded normals are corrected at inference. First, geometry-aware pixel-space losses, including inverse rendering self-consistency, von~Mises-Fisher angular loss, and wavelet edge-aware regularization, complement latent MSE by enforcing spherical normal geometry and diffuse image-formation cues after VAE decoding. Second, a lightweight Geometric Refinement Module (GRM) applies an RGB-guided residual correction to reduce boundary-localized decoding errors without freely rewriting the coarse prediction. On general-scene benchmarks, TransNormal-2 matches or exceeds MoGe-2 on all eight reported metrics while using only 1.4% as many task-specific normal annotations. The gains are clearest for transparent objects, reducing MAE by 4.2° on ClearGrasp and 3.1° on ClearPose over the strongest prior baselines. Code will be released at https://longxiang-ai.github.io/TransNormal-2.
The Brief
2:34A short podcast summary of the paper. Written and voiced by AI.
Transcript
TransNormal-2. Geometry-Grounded Rectified Flow with Edge-Aware Decoding for Precise Normal Estimation. By Mingwei Li and two others, from Zhejiang University and Zhongguancun Academy. Published on the sixth of September. Its code repository is public but slight: six source files.
Diffusion-based models have made monocular geometry estimation possible, but their precision at the level of pixels is limited by a shared source of error that has received little study: degradation in the VAE's reconstruction. The VAE encoder and decoder compress space by eight times, and that compression degrades surface normals at object boundaries. Even encoding and then decoding ground-truth normals introduces between one point three and eight point five degrees of mean angular error. At edges, that error reaches two point eight times the global figure. Part of the error is baked in before the model predicts anything. A model trained only to get its latent prediction right inherits that loss, however accurate the latent is.
TransNormal-2 is a rectified-flow framework built on FLUX.2, with single-step deterministic inference. It addresses the VAE degradation on both sides of the decoder: in how latent predictions are supervised during training, and in how decoded normals are corrected at inference.
For training, geometry-aware losses in pixel space complement the usual latent MSE. They include inverse rendering self-consistency, a von Mises-Fisher angular loss, and wavelet edge-aware regularization. Together they enforce spherical normal geometry and diffuse image-formation cues after VAE decoding, so the supervision sees what the decoder actually produces. For inference, a lightweight Geometric Refinement Module, or GRM, applies an RGB-guided residual correction. It reduces decoding errors localised at boundaries, without being free to rewrite the coarse prediction. The correction is residual and guided by the input image: it adjusts what the decoder produced, especially along boundaries, rather than replacing it.
On general-scene benchmarks, TransNormal-2 matches or exceeds MoGe-2 on all eight reported metrics, while using only one point four percent as many task- specific normal annotations. The gains are clearest on transparent objects. Mean angular error falls by four point two degrees on ClearGrasp and three point one degrees on ClearPose, compared with the strongest prior baselines.
Everything here follows from where the error was found. If the VAE degrades normals even when the answer is known exactly, then better latent prediction alone cannot fix what the decoder loses. TransNormal-2 supervises after decoding and corrects after decoding, and matches a strong baseline with a small fraction of its annotations. Code is to be released.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
The model is accurate, yet the edges are wrong. Something else is to blame.
You estimate surface normals from a single photo using a big image generator that works in a compressed latent space. Its predictions look good, yet edges of objects come out wrong.
The road ahead
- 1 · The problem ~7 min An image generator's compressor blurs surface normals at object edges, so some error exists before the model predicts anything.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Supervise the normals after decoding with geometry-aware pixel losses, then fix surviving edge errors with a small photo-guided correction.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Predict normal latents in one step, decode them, train with geometry-aware pixel losses, then apply a gated RGB-guided residual.
- 6 · The experiment ~6 min You’ll guess: TransNormal-2 matches or exceeds MoGe-2 on general-scene benchmarks. What share of MoGe-2's task-specific normal annotations does it use to get there?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.