TransNormal-2: Geometry-Grounded Rectified Flow with Edge-Aware Decoding for Precise Normal Estimation
TransNormal-2 targets VAE decoding errors that blur surface normals at object edges.
The paper says 8x VAE compression can add 1.3-8.5° mean angular error even when reconstructing ground-truth normals, with edge errors much worse. Its FLUX.2-based rectified-flow setup adds geometry-aware pixel losses during training and an RGB-guided refinement module at inference. The authors report matching or beating MoGe-2 across eight general-scene metrics while using far fewer task-specific normal annotations. The largest stated gains are on transparent-object benchmarks, cutting MAE by 4.2° on ClearGrasp and 3.1° on ClearPose versus the strongest prior baselines. HF Daily Papers' note
The paper says 8x VAE compression can add 1.3-8.5° mean angular error even when reconstructing ground-truth normals, with edge errors much worse. Its FLUX.2-based rectified-flow setup adds geometry-aware pixel losses during training and an RGB-guided refinement module at inference. The authors report matching or beating MoGe-2 across eight general-scene metrics while using far fewer task-specific normal annotations. The largest stated gains are on transparent-object benchmarks, cutting MAE by 4.2° on ClearGrasp and 3.1° on ClearPose versus the strongest prior baselines. HF Daily Papers' note
score 5