k-aisi-anti-deepfake/dear-ai-image-detection

DEAR: Dissect and Prune — AI-Generated Image Detection

This Space demonstrates DEAR-c from the paper "Dissect and Prune: Enhancing Robustness in AI-Generated Image Detection" (ICML 2026).

Upload an image to check whether it is a real photograph or AI-generated. The model uses a ResNet-50 backbone with bilateral channel pruning for robustness to post-processing (JPEG compression, resizing, color jitter).

How it works

1. 1.Diagnostic data. Inpaint a random region of each real image with Stable Diffusion 1.5, so real and generated pixels coexist within one image.
2. 2.Dissect. Score each feature channel by Regional Activation Discrepancy (RAD) — the difference between activation inside the inpainted region and the background.
3. 3.Prune & refine. Bilaterally prune channels at both extremes of the RAD distribution (keeping the robust middle band), freeze the backbone, and re-train the linear classifier.

A positive logit ⇒ FAKE (AI-generated), a negative logit ⇒ REAL.

Example images

The first three examples are real photographs (from linoyts/repo-to-space-example-inputs). The last two are AI-generated (LDM) images from the LDMFakeDetect dataset used in the paper.

Links

• —📄 Paper (arXiv)
• —💻 Code
• —🤗 Checkpoints