[Submitted on 28 Jul 2026 (v1), last revised 28 Oct 2026 (this version, v3)]
Abstract:Consumer tools can now keep the subject of a photograph and replace, relight or edit everything behind it. Such edits place genuine people and objects in scenes they were never in, yet involve processing–resampling, harmonization, diffusion autoencoders–that also occurs in benign workflows. We introduce BG-REAL, a benchmark of 25,624 images built from 2,485 Open Images photographs with a dominant subject and real donor scenes. Each recipient is edited by donor compositing, learned harmonization, generative background inpainting, foreground-conditioned relighting and local background edits, and is paired with authentic controls that share the processing but not the semantic edit; all images of a recipient share file format and JPEG quality. Three generators (including FLUX.1 Fill), scene categories and subject classes are held out. Seven detectors published between 2022 and 2026 reach zero-shot AUROCs of at most 0.745 and have no useful operating point: manipulation localisers mark the preserved subject instead of the replaced background, and a detector of AI-generated images flags most authentic images passed through a diffusion autoencoder. After fine-tuning, the strongest detectors reach an AUROC of about 0.95, while re-implemented background-forgery detectors remain below 0.80. Controls show that full-background replacements by the training generators are detected from the edit itself, whereas local edits and SDXL and FLUX edits remain hard, and that full-background localisation largely reduces to subject segmentation. A model-agnostic training objective that uses the controls improves detection on unseen subjects for one detector and on local edits for another without changing in-distribution accuracy. The data are released.
Submission history
From: Buğra Alperen Uluırmak [view email]
[v1] Tue, 28 Jul 2026 20:11:26 UTC (2,601 KB)
[v2] Thu, 15 Oct 2026 18:02:30 UTC (2,602 KB)
[v3] Wed, 28 Oct 2026 10:08:56 UTC (2,402 KB)
