Technical Summary: Ivy-Fake and Ivy-xDetector

Problem Statement

The rapid development of AI-generated content (AIGC), driven by diffusion models such as Sora, DALL-E, and Stable Diffusion, has produced high-quality synthetic media that blurs the line between authentic and fabricated content. This poses significant security risks, including the spread of misinformation, document manipulation, and dataset poisoning.

Current detection approaches face two main limitations:

1. Lack of multidimensional, explainable datasets: Existing open-source datasets (e.g., WildFake, GenVideo) rely predominantly on highly simplified binary annotations (real/fake). This limits the explainability and trustworthiness of trained detectors, as they cannot provide detailed justifications for their predictions. Furthermore, existing resources are fragmented; some focus exclusively on images (e.g., FakeClue), others on videos, with no unified benchmark for multimodal detection.
2. Insufficient interpretability in MLLM-based detectors: Previous detectors based on multimodal large language models (MLLMs, e.g., FakeVLM) often exhibit coarse-grained interpretability. Their step-by-step reasoning lacks the fine-grained depth necessary for reliable artifact localization and comprehensive explanation of synthetic fingerprints.

Methodology

1. Ivy-Fake: A Unified Benchmark

The authors introduce Ivy-Fake, the first large-scale multimodal benchmark designed for explainable AIGC detection for both images and videos.

• Data Scale and Composition: The dataset comprises over 106,000 training samples (61,107 images and 45,272 videos) and 5,000 manually verified evaluation examples.
• Data Sources: Data were curated from diverse public datasets (GenVideo, LOKI, FakeClue, WildFake, Kinetics-400, MMFakeBench, GenImage, DiffusionForensics) and generated using state-of-the-art models (FLUX.1, Lora Flux, Qwen-Image).
• Annotation Strategy: In contrast to binary labels, Ivy-Fake offers rich, fine-grained annotations divided into two main dimensions:
  • Spatial Features: Applicable to both images and videos, covering eight sub-dimensions: Impractical Luminosity, Localized Blur, Illegible Letters, Omitted Components, Spatial Relationships, Distorted Components, Chromatic Irregularities, and Abnormal Textures.
  • Temporal Features: Exclusive to videos, covering four sub-dimensions: Brightness Discrepancy, Unnatural Facial Expressions, Duplicated Components, and Non-spatial Relationships.
• Quality Assurance: A rigorous process involving automated filtering (corrupted files, resolution constraints, removal of near-duplicates) and human verification (10 users, approx. 1,000 hours) ensures data integrity. The final test set contains 2,500 images and 2,500 videos, balanced between real and fake instances.

2. Ivy-xDetector: A Reinforcement Learning Fine-tuned MLLM

To leverage the Ivy-Fake dataset, the authors propose Ivy-xDetector, a multimodal large language model developed for robust detection and explainable reasoning.

• Training Pipeline: The model undergoes a two-stage training process:
  1. Instruction-guided Cold Start (SFT): The model is initialized using detection and explanatory Chain-of-Thought (CoT) data generated by Gemini 2.5 Pro. This stage injects fine-grained artifact perception and reasoning capabilities into the base model (Qwen2.5-VL-3B).
  2. Reinforcement Learning with Verifiable Rewards (RLVR): The model is further optimized via Group Relative Policy Optimization (GRPO). This stage employs a rule-based reward system consisting of:
     • Accuracy Reward: Awards 1 point if the <conclusion> tag matches the Ground Truth (real/fake).
     • Format Reward: Awards 1 point if the output strictly follows structural requirements (reasoning in <thought>, decision in <conclusion>).
• Efficiency: Ivy-xDetector achieves state-of-the-art performance with fewer than 200,000 samples, a significant reduction compared to existing detectors that often rely on millions of data points.

Main Contributions

1. Ivy-Fake Dataset: The first large-scale, unified dataset for explainable AIGC detection covering both images and videos, offering over 106,000 training samples with fine-grained visual annotations and textual reasoning.
2. Ivy-xDetector Model: A reinforcement learning-based MLLM distinguished by its ability to detect generative artifacts (spatial for images; spatial and temporal for videos) and generate human-understandable reasoning chains.
3. Comprehensive Evaluation: Extensive experiments validating the dataset and model, demonstrating their impact on improving the capabilities of multimodal large models in detecting synthetic content.

Experimental Results

The authors conducted extensive evaluations across multiple benchmarks, including GenImage, GenVideo, Chameleon, and the internal Ivy-Fake test set.

• Performance on Ivy-Fake: Ivy-xDetector achieved an overall accuracy of 86.4% and an F1-score of 86.4%, outperforming other open-source models (e.g., InternVL3.5, Qwen2.5-VL) and approaching the performance of closed models like Gemini-2.5-Flash (77.9% accuracy). Notably, Ivy-xDetector demonstrated superior interpretability, achieving an average explanation score of 3.85 (on a scale of 1–5) compared to 3.54 for the best open-source baseline, while significantly surpassing it in detection accuracy.
• Generalization:
  • GenImage (Images): Ivy-xDetector achieved an average accuracy of 96.32% across various generators (Midjourney, Stable Diffusion, etc.), outperforming previous SOTA methods such as AIDE (86.88%) and DRCT (89.50%).
  • GenVideo (Videos): The model achieved an average F1-score of 95.26% and a recall rate of 95.28%, surpassing the previous SOTA DeMamba (F1 = 90.20%).
  • Chameleon (Unseen Domain): On the Chameleon benchmark, Ivy-xDetector achieved an overall accuracy of 73.17% and demonstrated a more balanced performance between real and fake classes compared to AIDE, which tended to overfit to the real category.
• Data Efficiency: The model achieved these results with significantly fewer training samples (approx. 130,000 images and 64,000 videos) compared to competitors that often require millions of samples (e.g., DeMamba used 2.295 million videos).

Significance and Claims

The work claims that Ivy-Fake and Ivy-xDetector address critical gaps in the landscape of AIGC detection by unifying image and video detection under a single, explainable framework. The authors emphasize that:

• Explainability is crucial: Moving beyond binary classification to detailed, multidimensional reasoning strengthens trust and enables better localization of synthetic artifacts.
• Efficiency: High performance can be achieved with smaller, high-quality, richly annotated datasets rather than massive, unlabeled corpora, provided the data contain fine-grained reasoning.
• Unified Approach: A single model can effectively handle both spatial (images) and spatiotemporal (videos) anomalies, overcoming the fragmentation of earlier purely image-based or purely video-based solutions.

The authors conclude that their work provides a robust foundation for transparent, trustworthy multimodal analysis and suggests that such fine-grained explainability-related data can catalyze new research directions for the AIGC detection community. They acknowledge limitations and note that future work could extend these benchmarks to domain-specific scenarios such as DeepFakes and document manipulation, and that the approach shows potential for scaling to larger models (e.g., 32B or 72B parameters).