Abstract
Recent advances in generative AI have dramatically improved image and video synthesis capabilities, significantly increasing the risk of misinformation through sophisticated fake content. In response, detection methods have evolved from traditional approaches to multimodal large language models (MLLMs), offering enhanced transparency and interpretability in identifying synthetic media. However, current detection systems remain fundamentally limited by their single-modality design. These approaches analyze images or videos separately, making them ineffective against synthetic content that combines multiple media formats. To address these challenges, we introduce BusterX++, a novel framework designed specifically for cross-modal detection and explanation of synthetic media. Our approach incorporates an advanced reinforcement learning (RL) post-training strategy that eliminates cold-start. Through Multi-stage Training, Thinking Reward, and Hybrid Reasoning, BusterX++ achieves stable and substantial performance improvements. To enable comprehensive evaluation, we also present GenBuster++, a cross-modal benchmark leveraging state-of-the-art image and video generation techniques. This benchmark comprises 4,000 images and video clips, meticulously curated by human experts using a novel filtering methodology to ensure high quality, diversity, and real-world applicability. Extensive experiments demonstrate the effectiveness and generalizability of our approach.
Introduction
As generative AI models such as FLUX and Sora become increasingly sophisticated, the boundary between real and synthetic media has blurred, creating significant opportunities for misinformation. Current detection systems face two primary challenges: modality silos and the "black box" problem. Most detectors are specialized for either images or videos, failing to leverage shared visual and semantic priors across formats. Furthermore, traditional detectors often provide binary labels without explaining the underlying reasoning. While some MLLM-based methods offer explanations, they frequently rely on Supervised Fine-Tuning (SFT) with human-written rationales, which can lead to "pseudo-reasoning" where the model mimics a style without grounding its logic in visual cues.
Methodology
The core of BusterX++ is an SFT-free, multi-stage Reinforcement Learning (RL) pipeline designed to foster genuine visual reasoning.
Unified Training
Rather than training separate models for different media types, BusterX++ utilizes a single base model (Qwen2.5-VL-7B) trained on a mixture of image and video data. This unified approach facilitates cross-modal knowledge transfer, allowing the model to apply high-quality image texture priors to identify forgery traces in video frames.
Direct Reinforcement Learning
BusterX++ bypasses the standard SFT step, moving directly to RL. This allows the model to discover more adaptive decision strategies rather than being constrained by human-authored templates. The training process follows a multi-stage strategy:
- Basic RL: Focuses on classification accuracy (Real vs. Fake) and adherence to output formatting.
- Extended RL: Refines reasoning through a "Thinking Reward" provided by an external judge model (SophiaVL-R1). This reward evaluates the coherence, completeness, and logical flow of the model's step-by-step explanations.
To stabilize this process, the framework employs the Direct Alignment Policy Optimization (DAPO) algorithm, maximizing task-related rewards while maintaining training stability.
The GenBuster++ Benchmark
To evaluate the limits of modern detectors, we introduce GenBuster++, a high-quality unified benchmark consisting of 4,000 samples (1,000 real images, 1,000 fake images, 1,000 real videos, and 1,000 fake videos).
GenBuster++ incorporates content from 13 state-of-the-art video generators (including Sora, Kling 2.0, and Luma Ray2) and 7 image generators (such as FLUX 1.1 and Recraft). Every sample underwent rigorous human filtering to ensure high realism and diversity across nine scenario categories, including Human, Landscape, and Vehicle. Training for BusterX++ was conducted using the So-Fake-Set for images and GenBuster-200K for videos.
Experimental Results
Performance and Synergy
BusterX++ achieved an overall accuracy of 77.5% on the GenBuster++ benchmark, significantly outperforming the previous BusterX (68.3%) and other leading MLLMs. A key finding was the "synergy effect": the unified model outperformed its own image-only and video-only variants, confirming that cross-modal training enhances performance on individual modalities. Furthermore, the RL-direct approach achieved 77.4% accuracy, surpassing the 72.9% achieved by a traditional SFT baseline.
Interpretability and Robustness
In a double-blind evaluation, human experts preferred BusterX++'s explanations over those generated by GPT-5 in 71% of cases, noting superior evidence grounding and specificity. The model also demonstrated zero-shot robustness against common perturbations such as JPEG compression, Gaussian noise, and blur, despite not being explicitly trained on degraded data.
Conclusion
BusterX++ represents a significant step toward unified, interpretable AI-generated content detection. While the rapid evolution of generative models remains a challenge, our SFT-free RL strategy provides a robust foundation for cross-modal reasoning. Future work will focus on improving the spatial localization of forged areas and refining reward functions to further push the boundaries of MLLM-based forensic analysis.
Summary updated