Research map / Detection and segmentation / Weakly supervised detection and segmentation

Less Supervision, Better Generalization: Weakly Supervised Fake Region Localization in Diffusion-Edited Images

NeurIPS 2026 · Detection and segmentation › Weakly supervised detection and segmentation · approach: Weakly supervised object localization · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ReGFLoW, a weakly supervised (image-level real/fake labels only) diffusion-edited fake region localizer that injects SD1.5 VAE reconstruction residuals into feature and score spaces and trains with artifact-centric MIL.

Problem

Localizing diffusion-edited (AI-edited) regions usually requires pixel-level masks, which are hard to obtain at scale and can mislead training by treating out-of-mask diffusion traces as real, hurting cross-generator and cross-dataset generalization.

Approach

Uses frozen SD1.5 VAE reconstruction error as a dense spatial prior, injected via gated residual feature fusion and a zero-initialized logit bias, and trains a 16x16 patch score map with softmax-weighted soft top-k MIL and fake-real pairwise margin loss, using only image-level labels.

Key results

• On OpenSDID P setting, ReGFLoW reaches OOD Avg F1 38.6 vs 37.3-ish for most fully supervised baselines (MaskCLIP 41.7 OOD F1 top baseline in table), with Flux.1 F1 29.2 vs 18.4 for MaskCLIP.
• On OpenSDID P+F setting, ReGFLoW achieves OOD Avg F1 31.3, above MaskCLIP's 29.1.
• On COCO-GLIDE (P) ReGFLoW gets F1 42.01 vs 25.17 for TruFor (best baseline); on DOLOS P+F it gets 21.19 vs 14.65 for MaskCLIP.

Surprising findings

• Image-only baselines generally surpass methods jointly trained with pixel-mask supervision in OOD average image-level performance.
• In the P setting, MaskCLIP has a higher OOD Avg F1 (41.7) than ReGFLoW (38.6), so ReGFLoW is only competitive, not best, in that setting.

Limitations

• Localization in the P setting is only competitive, not best, versus fully supervised MaskCLIP on OOD average F1.
• Reliance on a fixed SD1.5 VAE reconstruction prior, whose effectiveness for generators with different autoencoders is not separately analyzed.
• Evaluated mainly on OpenSDID with a single training source (SD1.5) in main experiments.

Experimental setup

Datasets: OpenSDID COCO-GLIDE DOLOS

Benchmarks: OpenSDID COCO-GLIDE DOLOS

Metrics: pixel-level F1 pixel-level IoU image-level F1 image-level accuracy

Baselines: MaskCLIP TruFor IML-ViT PSCC-Net WSCL DOLOS BoxPromptIML CNNDet GramNet FreqNet NPR

Builds on: multiple instance learning OpenSDI IMDLBenCo SideAdapterNetwork

Closest papers by meaning

• Training-Free Image Manipulation Localization Using Diffusion Models (AAAI 2025 · similarity 0.89)
• ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.86)
• Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images (CVPR 2026 · similarity 0.86)
• DiffForensics: Leveraging Diffusion Prior to Image Forgery Detection and Localization (CVPR 2024 · similarity 0.85)
• Can We Model the Artifacts Explicitly? Disentangle Artifacts via Pairwise Edit Relations for Image Manipulation Localization (NeurIPS 2026 · similarity 0.85)
• A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.85)