Training-Free Geometric Image Editing with Diffusion Models

Abstract

Geometric image editing means moving, rotating or resizing an object in a photograph while keeping its appearance and the surrounding scene. It is a demanding test for diffusion-based editors, because the object must change its pose without changing its identity. FreeFine is a training-free method built on a frozen Stable Diffusion model. It places the transformed object on an inpainted background and then refines the resulting coarse composite. On the GeoBench benchmark it outperforms earlier editing methods. This thesis reproduces FreeFine, diagnoses where and why it fails, and improves it without training. The reproduction runs the public code on all 5,677 two-dimensional edits of GeoBench-2D. It is deterministic, and it matches the published realism and consistency metrics to within 2%. A metricdriven diagnosis locates the failure in hard rotations and hard resizings, which make up a quarter of the benchmark and carry most of the loss in subject consistency and geometric precision. Six explanations of the failure are then tested in twenty-one controlled experiments under a common protocol. None of the tested reconfigurations explains the failure, and neither does any of the tested mechanisms that supply the object’s identity or the exact geometry of the edit through the entry points of the frozen network. The evidence points instead to how refinement treats the coarse composite, and the right treatment depends on the transformation. For translations and moderate resizings the interior of the composite should be kept; for rotations the released refinement works best; and for severe resizings the reconstruction has to be guided. For this last case the thesis develops EPSREC, a projected guidance with recursive momentum. It lowers the keypoint error by 14.7% on the hard resizings of the development subset, and by 0.34 pixels (1.8%) on all 759 severe resizings of the benchmark. These findings are combined into SGR-EPSREC (Selective Geometric Refinement), a training-free policy that reads only the requested transformation and sends each edit to the treatment that suits it. On all 5,677 edits, SGR-EPSREC improves all seven metrics of GeoBench-2D over FreeFine. It raises subject consistency for 54% of the edits and lowers it for 18%, leaves every rotation exactly as FreeFine produces it, and does not lower realism. The thesis also examines what the benchmark’s metrics can and cannot measure. FreeFine already reaches 99% of the subject consistency of the composite, which is a copy of the source at its target pose, and part of the measured gains follows from how the metrics are built. On this basis, the thesis separates the improvements that reflect better editing from those that do not. All code, executed notebooks and results are publicly available.