TransNormal-2 Reduces VAE Reconstruction Error in Monocular Normal Estimation
September 5, 2026
TransNormal-2 uses a FLUX.2-based rectified-flow framework to mitigate the 1.3-8.5 degree angular error caused by 8x VAE spatial compression. It employs geometry-aware pixel-space losses and edge-aware decoding to improve surface normal precision at object boundaries.
HOW THIS AFFECTS YOU
●
builderThis provides a more accurate method for generating high-fidelity surface normals in 3D vision pipelines.
●
researcherYou can utilize rectified-flow for more precise geometric latent supervision.