cs.CVSep 27, 2026
SaveIVT-Guard: All-in-One Reasoning Model for AI-Generated Content Detection
Organizations: Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, Xiamen 361005, P.R. China · Tencent YouTu Lab
Abstract
The rapid proliferation of highly realistic AI-Generated Content (AIGC) necessitates robust and interpretable detection mechanisms. However, existing detectors are predominantly confined to single modalities and provide binary outputs without reasoning. While Multimodal Large Language Models (MLLMs) present a promising solution, their development is constrained by the scarcity of multimodal reasoning data and the reasoning-detection optimization dilemma, where explicit reasoning supervision can compromise detection accuracy. To this end, we introduce IVT-Set, a comprehensive dataset comprising over 152K diverse image, video, and text samples equipped with multi-granularity Chain-of-Thought (CoT) reasoning trajectories. Based on it, we propose IVT-Guard, a pioneering framework for unified and interpretable AIGC detection across image, video, and text modalities. Furthermore, to overcome the aforementioned optimization dilemma, we design a novel three-stage training paradigm: Artifact-Aware Pre-training, Artifact-to-Evidence Supervised Fine-Tuning via artifact-aware injection, and Evidence-Verdict Consistency Group Relative Policy Optimization. Extensive experiments demonstrate that IVT-Guard achieves state-of-the-art detection performance across in-domain, out-of-domain, and cross-dataset settings while delivering faithful reasoning. Code and data will be released.
Figures & tables
Figure 2: Comparison of OOD accuracy (%) under different SFT strategies for three MLLMs on image (I), video (V), and text (T). Colors indicate accuracy changes relative to the standard CoT.
Figure 3: Data annotation pipeline for IVT-Set, consisting of CoT generation, multi-judge quality judgment, and iterative critique–refinement to produce high-quality multi-granularity CoT.
Figure 4: Overview of the three-stage training pipeline for IVT-Guard. Stage 1: AAP enhances sensitivity to modality-specific generation artifacts; Stage 2: A2E-SFT integrates artifact-aware representations into the MLLM for evidence-grounded reasoning; and Stage 3: EVC-GRPO jointly optimizes detection accuracy and evidence–verdict consistency.
| Method | ID | CD | OOD | Avg. | |||||||||
| Chameleon | LOKI | GenImage | GenBuster++ | Real | FLUX.1 Krea | Qwen-Image | Midjourney v6 | Infinity | Janus-Pro-7B | LlamaGen | |||
| Image-specific Detectors | |||||||||||||
| UnivFD ( Ojha et al., 2023 ) | 77.75 | 52.42 | 67.70 | 72.97 | 54.70 | 82.75 | 79.88 | 68.00 | 73.38 | 73.62 | 74.12 | 87.38 | 72.24 |
| DIRE ( Wang et al., 2023 ) | 89.00 | 70.92 | 71.23 | 75.95 | 72.35 | 89.75 | 72.88 | 55.75 | 74.75 | 81.50 | 71.00 | 87.25 | 79.25 |
| NPR ( Tan et al., 2024 ) | 84.62 | 60.15 | 66.07 | 75.70 | 52.90 | 91.25 | 61.12 | 54.62 | 70.12 | 81.00 | 62.62 | 65.25 | 72.58 |
| AIDE ( Yan et al., 2025 ) | 88.62 | 61.81 | 77.74 | 87.76 | 60.15 | 89.25 | 79.75 | 74.12 | 74.25 | 84.12 | 80.62 | 90.00 | 80.74 |
Table 2: Accuracy comparison (%) on the image modality of IVT-Set and four cross-dataset benchmarks. Avg. denotes the mean of ID accuracy, average CD accuracy, and average OOD accuracy. Best and second-best results are shown in bold and underlined , respectively.
| Method | ID | CD | OOD | Avg. | |||||||
| LOKI | AIGVDBench † | GenBuster++ | Real | EasyAnimateV5.1 | HunyuanVideo-I2V | Kling 2.1 | Veo 3 | Vidu Q2 | |||
| Video-specific Detectors | |||||||||||
| ReStraV ( Internò et al., 2025 ) | 69.25 | 36.64 | 44.79 | 52.00 | 77.25 | 48.62 | 67.62 | 58.63 | 50.62 | 54.37 | 57.75 |
| NSG-VD ( Zhang et al., 2025 ) | 50.38 | 57.46 | 78.73 | 50.40 | 18.75 | 46.00 | 45.25 | 50.12 | 48.00 | 47.88 | 51.75 |
| DeMamba ( Chen et al., 2026b ) | 87.62 | 79.25 | 44.22 | 62.45 | 88.25 | 67.00 | 76.00 | 63.88 | 55.88 | 74.38 | 73.50 |
| Generic MLLMs | |||||||||||
Table 3: Accuracy comparison (%) on the video modality of IVT-Set and three cross-dataset benchmarks. † denotes samples from four closed-source generators in AIGVDBench ( Ma et al., 2026 ) . Avg. denotes the mean of ID accuracy, average CD accuracy, and average OOD accuracy. Best and second-best results are shown in bold and underlined , respectively.
Figure 5: Ablation study of visual encoder combinations in Artifact-Aware Pre-training.
Appendix figures & tables25 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | ID | CD | OOD | Avg. | ||||||
| LOKI | BiScope | DetectRL-X | AI Detector Bench | Real | Claude Opus 4.1 | Gemini 2.5 Pro | GPT-4o | |||
| Text-specific Detectors | ||||||||||
| ModernBERT ( Warner et al., 2025 ; Drayson et al., 2025 ) | 51.75 | 78.16 | 63.00 | 54.62 | 62.10 | 89.00 | 64.38 | 62.88 | 66.25 | 62.28 |
| Desklib AI ( Desklib, 2025 ) | 79.75 | 93.75 | 96.61 | 84.28 | 83.40 | 91.00 | 93.50 | 91.62 | 91.88 | 87.09 |
| Generic MLLMs & LLMs | ||||||||||
| Qwen3-VL-8B | 43.25 | 51.81 | 42.14 | 50.93 | 50.70 | 69.50 | 47.62 | 45.38 | 44.25 | 47.94 |
Table 6: Performance comparison (Acc, %) on the text modality of IVT-Set and three cross-dataset benchmarks. Results are reported under in-domain (ID), cross-dataset (CD), and out-of-domain (OOD) settings. Avg. denotes the average of the ID accuracy, the mean CD accuracy, and the mean OOD accuracy. Best results are in bold and the second-best are underlined .
| Method | Image | Video | Text | |||
| Score | Elo | Score | Elo | Score | Elo | |
| GLM-4.6V | 3.09 | 1036 | 2.91 | 964 | 3.33 | 1132 |
| Qwen3-VL-235B-A22B | 3.88 | 1352 | 3.06 | 1024 | 3.68 | 1294 |
| Qwen3-VL-8B-Instruct | 3.91 | 1364 | 2.96 | 984 | 3.41 | 1164 |
| IVT-Guard (Ours) | 4.70 | 1676 | 4.34 | 1564 | 4.28 | 1712 |
Table 7: Explanation reliability evaluation on the OOD test set. We report the absolute score evaluated by Qwen3.5-122B-A10B and the pairwise Elo rating obtained from human evaluation for each modality, with all methods initialized at an Elo rating of 1,000.
| Training Data | ID | OOD | ||||
| Image (Acc/F1) | Video (Acc/F1) | Text (Acc/F1) | Image (Acc/F1) | Video (Acc/F1) | Text (Acc/F1) | |
| Image only | 98.00/98.31 | 51.38/42.62 | 47.00/36.45 | 91.13/91.03 | 48.45/39.05 | 48.21/36.95 |
| Video only | 68.00/67.54 | 91.25/91.25 | 45.38/39.43 | 66.73/63.79 | 75.73/75.14 | 46.54/39.80 |
| Text only | 56.25/51.55 | 49.88/46.38 | 88.63/88.62 | 55.56/50.97 | 49.05/46.77 | 88.33/88.30 |
| Image + Video | 98.75/98.75 | 92.63 / 92.62 | 46.00/39.73 | 91.58/91.52 | 77.23/76.63 | 46.88/39.25 |
| Video + Text | 71.63/71.11 | 91.75/91.75 | 97.75/97.75 | 70.90/68.34 | 78.65 / 78.16 | 96.04/96.12 |
Table 8: Cross-modal evaluation and modality ablation at the A2E-SFT stage. Results are reported as ID/OOD accuracy and F1 (%). Best results are in bold .
| Method | Perturbation | ID | OOD | ||||
| Image (Acc/F1) | Video (Acc/F1) | Text (Acc/F1) | Image (Acc/F1) | Video (Acc/F1) | Text (Acc/F1) | ||
| All | None | 98.75/98.75 | 92.00/91.99 | 98.38/98.37 | 93.48/93.43 | 82.80/82.71 | 98.00/98.00 |
| Image | JPEG Compression (QF=90) | 98.38/98.37 | 92.00/91.99 | 98.38/98.37 | 93.69/93.64 | 82.80/82.71 | 98.00/98.00 |
| Gaussian Blur ( ) | 96.63/96.62 | 92.00/91.99 | 98.38/98.37 | 92.69/92.65 | 82.80/82.71 | 98.00/98.00 | |
| Video | Shuffle frame | 98.75/98.75 | 92.25/92.24 | 98.38/98.37 | 93.48/93.43 | 82.55/82.45 | 98.00/98.00 |
| Resize | 98.75/98.75 | 89.13/89.06 | 98.38/98.37 | 93.48/93.43 | 81.93/81.87 | 98.00/98.00 | |
Table 9: Robustness evaluation of IVT-Guard under modality-specific input perturbations. Results are reported as accuracy/F1 (%) under in-domain (ID) and out-of-domain (OOD) settings for image, video, and text modalities.
Figure 6: OOD accuracy across image, video, and text modalities for three model families at different training stages. Bars denote OOD accuracy (%), while lines denote mean recall (%) across OOD generators, with “fake” treated as the positive category.
Figure 7: Overview of the IVT-Set construction pipeline, including modality-specific data pairing and filtering, diverse source and category coverage, and multi-granularity CoT annotations.
| Method | Model Type | Resolution | Data Scale | Data Split |
| BigGAN ( Brock et al., 2019 ) | GAN | 200 200 / 128 128 | 742 | ID |
| StyleGAN ( Karras et al., 2019 ) | GAN | 200 200 | 1,155 | ID |
| DF-GAN ( Tao et al., 2022 ) | GAN | 256 256 | 198 | ID |
| GALIP ( Tao et al., 2023 ) | GAN | 224 224 | 994 | ID |
| GigaGAN ( Kang et al., 2023 ) | GAN | 512 512 | 497 | ID |
| Stable Diffusion v1.4 ( CompVis, 2022 ) | Diffusion | 1024 704 / 1024 768 704 1024 / 768 1024 | 767 | ID |
Table 10: Details of the image modality generation methods in IVT-Set. The dataset includes 20 generators spanning GAN, diffusion/flow-based, and autoregressive paradigms, including a closed-source system whose architecture is not publicly disclosed. Generators are split into in-domain (ID) and out-of-domain (OOD) to evaluate cross-generator generalization. Image sizes are reported as width height in pixels. For generators with mixed-resolution samples, we report up to the four most common image sizes in descending order.
| Method | Duration (s) | Resolution | FPS | Data Scale | Data Split |
| Pika ( labs, 2022 ) | 3.00–4.00 | 768p / 576p 640p / 1024p | 12–24 | 1,000 | – |
| SVD ( Blattmann et al., 2023 ) | 3.57 | 576p | 7 | 1,000 | – |
| DynamiCrafter ( Xing et al., 2024 ) | 2.00 | 576p | 8 | 1,000 | – |
| OpenSora ( Zheng et al., 2024 ) | 2.00 | 256p / 512p | 8 | 1,000 | – |
| CogVideoX ( Yang et al., 2025b ) | 5.04 | 1024p | 24 | 1,000 | – |
| LTX-Video ( HaCohen et al., 2024 ) | 1.37–10.25 | 672p / 480p 512p / 240p | 6–60 | 4,999 | ID |
Table 11: Details of the video modality generation methods in IVT-Set. The dataset includes 14 generators with diverse resolutions and frame rates. Generators are split into in-domain (ID) and out-of-domain (OOD) categories when applicable to evaluate cross-generator generalization. Video resolutions are reported as output heights in pixels. For generators with mixed-resolution samples, we report up to the four most common output heights in descending order.
| Method | R./F. Ratio | Length | Data Scale | Data Split |
| Llama-3.3-70B-Instruct ( Grattafiori et al., 2024 ) | 1.09 | 20–1,190 | 1,486 | ID |
| gpt-oss-120b ( Agarwal et al., 2025 ) | 0.99 | 27–1,056 | 2,635 | ID |
| GLM-4.5-Air ( Zeng et al., 2025 ) | 1.04 | 20–1,116 | 2,469 | ID |
| DeepSeek-V3.1 ( Liu et al., 2024a ) | 1.09 | 20–1,516 | 2,473 | ID |
| Kimi K2-Instruct-0905 ( Team et al., 2025 ) | 1.05 | 20–787 | 2,190 | ID |
| Qwen3-235B-A22B-Instruct-2507 ( Yang et al., 2025a ) | 0.99 | 21–1,206 | 2,680 | ID |
Table 12: Details of the text modality generation methods in IVT-Set. The dataset includes 10 representative LLMs covering both open-source and closed-source models. Models are split into in-domain (ID) and out-of-domain (OOD) categories to evaluate cross-model generalization. R./F. Ratio denotes the total real-text length divided by the total corresponding generated-text length. Length reports the minimum–maximum generated-text length in words.
Figure 8: Data statistics of the image modality in IVT-Set. (a) Category distribution of all images across 13 semantic categories. (b) Generator distribution of fake images across 20 models spanning GAN, Diffusion, and Autoregressive paradigms. (c) Comparison of the mean quality scores of fake and real images in each category, showing that the quality scores of the two are very close across all categories.
Figure 9: Data statistics of the video modality in IVT-Set. (a) Category distribution of all videos across 11 semantic categories. (b) Generator distribution of fake videos across 14 generation models, covering both open-source and commercial models. (c) Correlation between aesthetic score and temporal consistency score for fake and real videos, showing that the two distributions largely overlap.
Figure 10: Data statistics of the text modality in IVT-Set. (a) Category distribution across 7 domains. (b) Generator distribution of fake texts from 10 LLMs. (c) Stylometric similarity among real-text sources. (d) Stylometric similarity among fake-text generators. For (c) and (d), each source is represented by a category-averaged profile of 20 rule-based stylometric features. Pairwise similarity is computed as the Pearson correlation between standardized source profiles and rescaled to , with higher values indicating more similar profiles.
Figure 11: Prompt template for CoT Generation on real images .
Figure 12: Prompt template for CoT Generation on fake images .
Figure 13: Prompt template for CoT Generation on real videos.
Figure 14: Prompt template for CoT Generation on fake videos .
Figure 15: Prompt template for CoT Generation on real texts .
Figure 16: Prompt template for CoT Generation on fake texts .
Figure 17: Prompt template for CoT Quality Judgment.
Figure 18: Prompt template for Critique Refinement on real images .
Figure 19: Prompt template for Critique Refinement on fake images .
Figure 20: Prompt template for Critique Refinement on real videos .
Figure 21: Prompt template for Critique Refinement on fake videos .
Figure 22: Prompt template for Critique Refinement on real texts .
Figure 23: Prompt template for Critique Refinement on fake texts .