Research map / Image generation / Editing and style transfer

Towards Reliable Identification of Diffusion-based Image Manipulations

NeurIPS 2025 · Image generation › Editing and style transfer · approach: Inversion-based diffusion editing · 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 RADAR, a multi-encoder (DINOv2 semantic + Depth Anything V2 geometry) fusion model with a patch-level supervised contrastive loss for localising and detecting diffusion-based inpainting, and introduces the BBC-PAIR benchmark with 28 inpainters.

Problem

Identifying and localising diffusion-based inpainting edits that generalise across many diffusion models, including LoRA-customised and unseen commercial inpainters, where existing IFDL methods trained on few inpainters fail.

Approach

Fuse frozen semantic and geometry transformer encoders via a symmetric cross-attention Fusion Block, and apply a patch-level supervised contrastive loss that separates original, tampered, and affected (non-masked patches from tampered images) distributions to improve cross-inpainter generalisation.

Key results

• RADAR reaches IoU 0.450 and AUC 0.805 on BBC-PAIR-OOD, outperforming all seven baselines on localisation IoU and detection AUC/accuracy.
• On BBC-PAIR-ID RADAR achieves detection AUC 0.945 and accuracy 0.931, versus at most 0.625 AUC for baselines.
• Ablation: removing the contrastive loss drops ID accuracy from 0.893 to 0.849; using K=10 inpainters instead of K=1 raises ID accuracy from 0.793 to 0.893.

Measured effects reported in the paper

Surprising findings

• Baselines such as TruFor achieve higher accuracy on splicing benchmarks like COVERAGE than on inpainting benchmarks, suggesting different detection cues.
• RADAR's localisation F1 on non-inpainting datasets (e.g. 0.944 on CMFD) exceeds its F1 on BBC-PAIR, while its detection AUC there is near chance on most datasets.

Limitations

• Detection performance is weak on non-inpainting datasets (AUC near 0.5 on COVERAGE, CASIA-v2, DSO), attributed to lack of affected/tampered distinction.
• Training uses only open-source inpainters, so commercial and unseen models remain a generalisation challenge; OOD localisation IoU is only 0.450.
• Relies on frozen pretrained encoders and on GPU compute; inference is slower than some baselines.

Experimental setup

Datasets: OpenImages-v7 CocoGlide SID-Set SafireMS-Expert DIV2K DPREVIEW COVERAGE CASIA-v2 IMD2020 CMFD AutoSplice DSO

Benchmarks: BBC-PAIR-ID BBC-PAIR-LoRA BBC-PAIR-OOD COVERAGE CASIA-v2 IMD2020 CMFD AutoSplice DSO

Metrics: AUC accuracy F1 IoU

Baselines: HiFi-Net TruFor MIML AdaIFL Mesorch SAFIRE SIDA

Builds on: DINOv2 Depth Anything V2 Grounded SAM Kosmos-2 supervised contrastive loss

Closest papers by meaning

• Towards Efficient Diffusion-Based Image Editing with Instant Attention Masks (AAAI 2024 · similarity 0.86)
• Diffusion Handles Enabling 3D Edits for Diffusion Models by Lifting Activations to 3D (CVPR 2024 · similarity 0.86)
• PS-Diffusion: Photorealistic Subject-Driven Image Editing with Disentangled Control and Attention (CVPR 2025 · similarity 0.86)
• Training-Free Image Manipulation Localization Using Diffusion Models (AAAI 2025 · similarity 0.86)
• Unsupervised Region-Based Image Editing of Denoising Diffusion Models (AAAI 2025 · similarity 0.86)
• $\text{I}^2\text{AM}$: Interpreting Image-to-Image Latent Diffusion Models via Bi-Attribution Maps (ICLR 2025 · similarity 0.86)