Photorealistic material replacement in natural images requires accurate estimates of both surface geometry and scene illumination. Existing methods either rely on implicit generative representations, leading to misaligned textures and limited user control, or require explicit 3D reconstruction or class-specific templates. We introduce ImageUnwrap, the first single-view method to estimate dense, per-pixel UV maps for scenes with arbitrary non-developable surfaces, without 3D reconstruction or class-specific templates. This enables users to create or edit textures directly in texture-space—controlling scale, rotation, and tiling—and then apply the result to the surface while preserving local deformations. Our method features a latent diffusion backbone adapted to structured per-pixel regression. We train the model on synthetic data comprising rendered images, UV maps, and segmented UV islands. A multi-scale loss based on spatial gradients guides the prediction of UV maps, explicitly allowing discontinuities at edges. During evaluation, the predicted UV coordinates are realigned with the ground truth using a least-squares method, enabling the computation of performance metrics. The resulting UV maps integrate seamlessly with both intrinsic and PBR-aware pipelines, as well as lightweight 2D compositing, allowing controllable material replacement and decal placement. We demonstrate state-of-the-art performance for monocular UV prediction and show that introducing explicit dense UV parameterization enables more accurate and controllable material editing than prior approaches.