Detective SAM accepted at ICLR 2026
We are thrilled to announce that "Detective SAM: Adaptive AI-Image Forgery Localization" has been accepted at ICLR 2026, one of the top venues in machine learning.
This paper extends our ICML 2025 workshop version with significant new contributions.
What's New
SAM2-Based Architecture
The full version of Detective SAM is built on SAM2, integrating perturbation-driven forensic clues with lightweight feature adapters and a mask adapter. SAM2's decoder is retargeted from object segmentation to forgery localization, with the backbone remaining frozen to enable efficient, lightweight fine-tuning as new editors appear.
AutoEditForge Pipeline
To keep up with the rapidly evolving capabilities of diffusion models, we introduce AutoEditForge: an automated diffusion edit generation pipeline that produces human-like local generative edits with pixel-accurate masks across four edit types:
- Replace: substituting objects or regions
- Remove: erasing elements from the scene
- Add: inserting new elements
- Change Partially: subtle modifications to existing content
This automated pipeline ensures Detective SAM can be continuously trained on realistic, diverse forgeries without requiring manual annotation.
Improved Results
Detective SAM achieves new state-of-the-art performance on common benchmarks as well as custom modern editing models such as NanoBanana and Qwen-Image-Edit, with consistent improvements over the workshop version.
Demo
Coming soon.
Links
This work is a collaboration with Nicolas van Schaik, Chaoyi Zhu, Pin-Yu Chen, Robert Birke, and Lydia Y. Chen.