UltraNormal: Scaling Pretrained Diffusion‐Based Surface Normal Estimators to Ultra‐High‐Resolution

Most of the current diffusion‐based surface normal estimators were trained at low resolutions (e.g., ), significantly below the resolution of modern cameras. Applying such models to high‐resolution images causes two issues: (1) downsampling removes high‐frequency geometric details, and (2) direct high‐resolution inference loses global context. To address these challenges, we propose UltraNormal, which augments existing pretrained diffusion‐based surface normal estimators by using overlapping sliding patches with test‐time guidance to inject global context. UltraNormal initializes each patch with the latent of the corresponding low‐resolution estimate enhanced with high‐frequency details from the local estimate and resumes the reverse diffusion process from an intermediate time step. This approach maintains geometric accuracy of the global prediction while adding high‐frequency details without retraining the diffusion model. To the best of our knowledge, this is the first test‐time method to scale low‐resolution diffusion‐based surface normal estimators to 4K resolution without any retraining. We build UltraNormal upon three different base models and evaluate on both real‐world data and a 4K photorealistic synthetic dataset. We perform extensive ablations and demonstrate improved accuracy, sharpness, and surface reconstruction quality compared to the base model.