Overview
Research published on arXiv investigates supervision structure in AI-Generated Image (AIGI) detection, proposing that blending semantic and artifact signals during training limits detector generalization. The authors formalize an approach termed Orthogonal Supervision Disentanglement, which pairs disentangled supervision targets with inductively matched architectures. Using this formulation, the researchers develop AlignGemini, a dual-branch detector that isolates semantic and artifact analysis, and introduce AIGI-Now, an evaluation benchmark designed to diagnose detector capabilities across modern image generators.
Research Context
Prior efforts in AIGI detection have primarily concentrated on developing novel neural network architectures, loss formulations, and training procedures. However, these systems generally train on standard benchmark distributions that conflate semantic content with synthesis artifacts. Controlled experiments reported in the paper demonstrate that implicitly combining these disparate signal types within a single supervision target is counterproductive to cross-generator performance.
Approach
To prevent signal interference, the authors propose Orthogonal Supervision Disentanglement, assigning distinct training objectives to separate, inductively matched model components:
- Semantic Branch: A Vision-Language Model (VLM) receives purely semantic supervision.
- Artifact Branch: A specialized vision expert model receives purely artifact-focused supervision.
This dual-branch configuration, implemented as AlignGemini, relies on a basic training pipeline without diverse generator coverage or complex multi-stage training routines.
To analyze the specific mechanisms underlying detection performance, the study also presents AIGI-Now. This diagnostic benchmark spans eight contemporary generation models, including five closed-source commercial systems. The dataset provides split evaluation subsets focused separately on semantic and pixel-level characteristics, allowing fine-grained assessment of detector behavior.
Findings
The study documents several key experimental results regarding supervision design and model transferability:
- Disentangling supervision targets while pairing them with inductively matched models provides sufficient regularization to achieve strong cross-generator generalization.
- AlignGemini yielded an average performance improvement of 9.5% across five in-the-wild detection benchmarks.
- The 9.5% improvement was achieved after training on output samples from only two image generators, without relying on broad training distributions or specialized training heuristics.