Localizing manipulations in fully regenerated (FR) images

Establish image forgery localization techniques that can identify and delineate manipulated regions in fully regenerated images produced by text-guided inpainting, where the entire image is regenerated during editing and traditional forensic traces are weakened or destroyed.

Background

The paper distinguishes between spliced edits and fully regenerated (FR) edits in text-guided inpainting. In FR edits, the generative model may subtly alter the entire image while only semantically changing the masked region, which disrupts conventional forensic traces relied upon by image forgery localization (IFL) methods.

Prior work and the expanded TGIF2 benchmark show that current IFL methods perform well on spliced forgeries but largely fail to localize manipulations in FR images, while synthetic image detection (SID) methods are not designed for localization. The authors emphasize that robust localization in FR scenarios remains unresolved.

References

With new generative inpainting models emerging and the open problem of localization in FR images remaining, updated datasets and benchmarks are needed.

— TGIF2: Extended Text-Guided Inpainting Forgery Dataset & Benchmark  (2603.28613 - Mareen et al., 30 Mar 2026) in Abstract, p. 1

RLDF is deliberately scoped to face swapping: all eight generators transfer identity, so region-level behaviour under reenactment, entire-face synthesis and attribute editing remains open.

— RLDF: A Region-Level Deepfake Dataset and Benchmark for Partial-Face Manipulation Detection  (2610.21438 - Reddy et al., 15 Oct 2026) in Section 5, “Limitations and Future Work”