BusterX++: Unified Cross-Modal Detection

• BusterX++ is a unified multimodal framework that classifies images and videos as real or fake while generating integrated, human-readable explanations.
• It employs a cold-start-free, RL-based post-training pipeline with innovative Thinking Mode Fusion, ensuring controlled reasoning and improved detection accuracy.
• Empirical evaluations on benchmarks like GenBuster++ demonstrate significant performance gains over previous models through joint image-video training and robust curation techniques.

BusterX++ is a multimodal LLM framework for unified cross-modal AI-generated content detection and explanation. It is designed to accept either an image or a video, decide whether the input is real or AI-generated, and optionally produce a human-readable rationale for that decision. In the paper introducing the system, “cross-modal” refers specifically to unified image-video detection rather than to text-image retrieval or multimodal generation, and “unified” refers both to inference and to post-training: the same model is trained on mixed image and video data and uses a common reasoning-and-answer interface across both modalities (Wen et al., 19 Jul 2025).

1. Definition and problem formulation

BusterX++ addresses the detection of highly realistic synthetic images and videos produced by contemporary generative models such as FLUX, GPT-4o, Seedance 1.0, and SkyReels V1. The motivating claim is that prior detection systems are fragmented along two axes. Conventional non-MLLM detectors generally perform binary classification with limited interpretability, while recent MLLM-based systems improve explainability but remain predominantly single-modality, handling images or videos separately rather than within a shared policy model (Wen et al., 19 Jul 2025).

The task is framed as unified cross-modal detection and explanation. The operational objective is to identify whether an image or video is real or fake, and, when requested, to explain the evidence. The paper emphasizes that the explanation is not a post-hoc auxiliary module. Instead, it is part of the native output format and later becomes part of the reinforcement-learning objective. This design places BusterX++ in the “reasoning-first” line of MLLM detectors, but with a cross-modal scope absent from its precursor BusterX, which was introduced as a video-only system for authenticity determination and explainable rationale on GenBuster-200K (Wen et al., 19 May 2025).

A central methodological controversy concerns whether explainable detection should begin with a cold-start supervised chain-of-thought stage. BusterX++ rejects that recipe. The paper argues that the cues indicating synthetic media are often subtle, intuitive, multidimensional, and partially non-linguistic—examples include unnatural reflections, inconsistent illumination, odd motion, texture irregularities, and uncanny-valley effects—and that weak chain-of-thought bootstrapping may cap later RL improvements. On this basis, the system is presented as a cold-start-free RL design rather than a conventional SFT-then-RL pipeline (Wen et al., 19 Jul 2025).

2. Architecture, input handling, and inference modes

The base model is Qwen2.5-VL-7B-Instruct. BusterX++ takes either an image or a video as visual input together with a textual instruction asking whether the content is real or fake. For videos, the system samples 16 frames at 4 FPS for video-level detection. The output is structured into two tagged fields:

• > ... </think> for optional reasoning or explanation > > - <answer> ... </answer> for the final classification, where A = real and B = fake (Wen et al., 19 Jul 2025) > > This template supports two inference regimes. In Thinking mode, the model generates non-empty reasoning and then an answer. In Non-thinking mode, the <think> field is left empty and the model directly emits the answer. The behavior is controlled through prompts such as /think and /no_think, and the default behavior is thinking mode. The paper calls this interface Thinking Mode Fusion, inspired by Qwen3 (Wen et al., 19 Jul 2025). > > The corresponding workflow is simple but tightly specified. First, the system receives an image or video plus a modality-specific prompt. Second, the MLLM performs visual analysis over cues including artifacts, realism, motion consistency, lighting, geometry, text, and scene semantics. Third, it predicts real or fake. Fourth, in thinking mode, it explains the evidence before returning the final answer. The paper explicitly distinguishes this from a detector-plus-explainer cascade: explanation is embedded in the model’s own response template and later reinforced as part of post-training (Wen et al., 19 Jul 2025). > > The prompting design in the supplementary material makes the reasoning style explicit. The system prompt requires the form <think> reasoning process here <answer> answer here </answer>. The image and video prompts instruct the model to analyze inconsistencies or signs of forgery, “think as if a human were pondering deeply,” and use phrases such as “let me think,” “wait,” “Hmm,” and “oh, I see,” before answering the multiple-choice question. In non-thinking mode, the prescribed output is:

$$\hat{A}_{i,t} = \frac{r_i - \text{mean}(\{r_i\}_{i=1}^G)}{\text{std}(\{r_i\}_{i=1}^G)} ,$$3

A common misconception is that non-thinking mode constitutes a separate detector. The paper instead presents both modes as prompt-controlled behaviors of the same policy model, with mode controllability established during post-training rather than by maintaining separate architectures (Wen et al., 19 Jul 2025).

3. RL-centric post-training and reward design

BusterX++ uses a three-stage post-training pipeline:

1. Stage-1: Foundation RL
2. Stage-2: Thinking Mode Fusion
3. Stage-3: Advanced RL (Wen et al., 19 Jul 2025)

The reinforcement-learning algorithm is DAPO (Dynamic sAmpling Policy Optimization), inherited from BusterX. The paper gives the per-token policy ratio and normalized advantage as

$$r_{i,t}(\theta)=\frac{\pi_{\theta}(o_{i,t} \mid q, o_{i,<t})}{\pi_{\theta_{\text{old}}}(o_{i,t} \mid q,o_{i,<t})}$$

and

$$\hat{A}_{i,t} = \frac{r_i - \text{mean}(\{r_i\}_{i=1}^G)}{\text{std}(\{r_i\}_{i=1}^G)} ,$$

where outputs are sampled in groups and optimization prefers responses with higher relative reward within the group while clipped policy ratios stabilize updates in a PPO-like manner (Wen et al., 19 Jul 2025).

Stage-1 teaches the basic real-vs-fake classification task. Its reward is

$$R_{\text{stage-1}} = r_{\text{fmt}} + r_{\text{overlong}} + r_{\text{acc}} .$$

The format term assigns $0$ for the required tagged template and $-1$ otherwise. The soft overlong term penalizes excessive response length with a buffer zone before the hard limit. The accuracy reward is $1$ for correct classification and $0$ otherwise. The stated goal is to let the model “quickly grasp the basic characteristics and features of different types of data” before emphasizing reasoning quality (Wen et al., 19 Jul 2025).

Stage-2 is the only explicit SFT stage. It is not a large cold-start chain-of-thought corpus. Instead, the authors collect several hundred samples from the Stage-1 model and use them to teach switching between /think and /no_think. The paper reports that this stage has minimal effect on final accuracy, but it is important for mode controllability (Wen et al., 19 Jul 2025).

Stage-3 further reinforces correct classification, reliable two-mode behavior, and higher-quality reasoning. Its reward is

$$R_{\text{stage-3}} = R_{\text{stage-1}} + r_{\text{hybrid}} + r_{\text{think}} .$$

The Hybrid Reasoning reward enforces proper template behavior in both modes, assigning $0$ if the response matches the prescribed mode and $-1$ if the model skips thinking in Thinking Mode or produces thinking content in Non-Thinking Mode. The Thinking Reward replaces BusterX’s earlier Length Reward with a semantic score from SophiaVL-R1-Thinking-Reward-Model-3B:

$$\hat{A}_{i,t} = \frac{r_i - \text{mean}(\{r_i\}_{i=1}^G)}{\text{std}(\{r_i\}_{i=1}^G)} ,$$0

with $$\hat{A}_{i,t} = \frac{r_i - \text{mean}(\{r_i\}_{i=1}^G)}{\text{std}(\{r_i\}_{i=1}^G)} ,$$1. Because the reward is gated by $$\hat{A}_{i,t} = \frac{r_i - \text{mean}(\{r_i\}_{i=1}^G)}{\text{std}(\{r_i\}_{i=1}^G)} ,$$2, incorrect answers cannot receive positive thinking reward. The paper also notes that applying Thinking Reward too early in Stage-1 destabilizes learning, which motivates the multi-stage schedule (Wen et al., 19 Jul 2025).

Relative to BusterX, this is the principal training change. BusterX used a cold-start SFT stage with 10k short CoT data samples and then RL with format, overlong, and correctness-conditioned length rewards. BusterX++ removes the initial cold-start bootstrapping and replaces explicit length encouragement with a semantically grounded thinking reward model (Wen et al., 19 May 2025).

4. GenBuster++ benchmark and evaluation substrate

To evaluate unified cross-modal detection, the paper introduces GenBuster++, a benchmark of 4,000 samples spanning both images and videos. Its final composition is:

• 1,000 real images
• 1,000 fake images
• 1,000 real videos
• 1,000 fake videos (Wen et al., 19 Jul 2025)

The benchmark is motivated by three criticisms of earlier datasets: single-modality scope, insufficient fine-grained human curation, and lack of realism. GenBuster++ is intended as a cross-modal benchmark aligned with real-world deployment. The paper states that it covers 9 categories of real-world scenarios and includes fake content from 7 state-of-the-art image generators and 13 state-of-the-art video generators. Explicitly named examples are FLUX and GPT-4o for images, and Seedance 1.0 and SkyReels V1 for videos (Wen et al., 19 Jul 2025).

Real images and videos are sourced from OpenVid-1M HD and pre-filtered for scene variety. Fake images and videos are sourced both from MagicArena and from a custom pipeline in which the authors use the Reddit official API to collect social media images, use Qwen-2.5-VL to generate detailed captions, and then use those captions as prompts for multiple generative models. This procedure is intended to ground synthetic content in realistic social-media scenarios rather than in purely fantastical prompts (Wen et al., 19 Jul 2025).

The most distinctive design feature is the filtering methodology. Real content is filtered for resolution, video frame rate, and video bitrate; duplicates, extensive watermarks, anime content, and samples with obviously synthetic backgrounds are removed. Fake content undergoes a two-stage human curation process: a mixed pool of real and fake samples is assembled; experts first select the samples that appear real; those selected samples are then re-examined to isolate the ones that are actually synthetic. The stated effect is to retain hard, realistic fake samples rather than easy ones (Wen et al., 19 Jul 2025).

The benchmark is standardized through post-processing. Images are resized to 1024 × 1024. Videos are standardized to 1920 × 1080, 5 seconds, and 24 FPS, and encoded with HEVC using x265. The stated rationale is elimination of encoding biases and consistency across heterogeneous sources. The paper does not explicitly provide train/validation/test split sizes for GenBuster++ in the available text, although it reports evaluation accuracies by modality and subcategory (Wen et al., 19 Jul 2025).

5. Empirical performance and ablation evidence

BusterX++ is evaluated on single-modality and cross-modal benchmarks. On So-Fake-Set, the paper reports 93.9 ACC / 93.7 F1 for BusterX++, compared with 93.2 ACC / 92.9 F1 for So-Fake-R1. On GenBuster-200K, it reports 88.3 ACC / 88.3 F1 on the test set and 92.4 ACC / 92.3 F1 on the OOD benchmark, improving over BusterX by +2.8 ACC on test and +7.6 ACC on OOD. On GenBuster++, the reported overall accuracy is 77.5, compared with 68.3 for BusterX and 55.4 for Qwen2.5-VL-7B (Wen et al., 19 Jul 2025).

Benchmark	BusterX++	Comparison stated in paper
So-Fake-Set	93.9 ACC / 93.7 F1	Beats So-Fake-R1 by 0.7 ACC and 0.8 F1
GenBuster-200K test	88.3 ACC / 88.3 F1	Beats BusterX by 2.8 ACC
GenBuster++ overall	77.5	Beats BusterX by 9.2 points

The cross-modal benchmark results are especially diagnostic. General-purpose MLLMs often classify real content well but perform poorly on fake detection, especially fake images. The paper highlights the jump in Image Fake accuracy from 8.9 for Qwen2.5-VL-7B to 76.2 for BusterX++. It also reports that non-thinking mode is close to thinking mode on GenBuster++, with 76.8 versus 77.5 overall, indicating that a single model can support both explanation-rich and efficient direct-answer inference with only a small loss in raw accuracy (Wen et al., 19 Jul 2025).

The ablations are central to the paper’s argument. In the cold-start vs non-cold-start comparison, a cold-start configuration with Stage-1 and Stage-3 reaches 72.9 overall, while the non-cold-start version with Stage-1 and Stage-3 reaches 77.4 overall. The improvement is especially pronounced for fake detection, with Image Fake: 76.7 vs 65.9. In the data modality ablation, joint image-video training yields 77.5 overall, compared with 71.4 for image-only training and 72.9 for video-only training, supporting the claim that shared post-training benefits both modalities (Wen et al., 19 Jul 2025).

The training strategy ablation shows that Stage-1 establishes the baseline, Stage-2 contributes little to raw accuracy, and Stage-3 supplies most of the gain. Specifically, Stage-1 only yields 69.4 overall, Stage-1 + Stage-2 yields 69.3, Stage-1 + Stage-3 yields 77.4, and the full pipeline reaches 77.5. The interpretation offered by the paper is that Thinking Mode Fusion is chiefly about controllability rather than discrimination (Wen et al., 19 Jul 2025).

Robustness is evaluated under JPEG compression, Gaussian noise, Gaussian blur, and a degradation cascade inspired by Real-ESRGAN. Starting from 77.5 overall on the original benchmark, performance drops to 74.9 under JPEG, 71.9 under Gaussian noise, 77.4 under blur, and 70.5 under the full cascade. The reported conclusion is that the model is fairly stable under low-level distortions without explicit degraded-data training, although fake detection degrades under the strongest corruption sequence (Wen et al., 19 Jul 2025).

6. Explanatory behavior, relation to BusterX, and limitations

The qualitative analyses describe BusterX++ as relying on a mixture of low-level and semantic cues. In fake-image and fake-video examples, the model mentions overly uniform lighting and shadows, unnatural texture regularity, suspiciously perfect geometry or composition, unrealistic object details, motion or physical inconsistencies, text anomalies, and the absence of expected real-world imperfections. In real examples, it emphasizes natural skin texture, plausible clothing fit and material properties, consistent lighting and shadows, realistic object motion, coherent background context, and natural variability. The paper characterizes its explanations as exhibiting Stable Reasoning, Attention to low-level details, and Advanced Reasoning using pre-training knowledge (Wen et al., 19 Jul 2025).

These explanatory properties distinguish BusterX++ from its immediate predecessor. BusterX was introduced as an MLLM-powered framework for AI-generated video forgery detection and explanation, trained on GenBuster-200K with a cold-start SFT stage and RL, and evaluated primarily on video benchmarks. Its core claim was to reinterpret video authenticity detection as a visual reasoning problem rather than a pure binary classification problem (Wen et al., 19 May 2025). BusterX++ generalizes that program in three ways: it unifies images and videos in one policy model, eliminates cold-start, and replaces length-based reward shaping with a semantic thinking reward model (Wen et al., 19 Jul 2025).

Two limitations are stated explicitly. First, adaptation remains difficult as generators improve: the paper notes that the newest generators used in GenBuster++ are harder than those in GenBuster-200K. Second, the authors suggest that post-training may be nearing a bottleneck for this task and that future work should explore training stages beyond current post-training improvements. Additional directions are described more cautiously: stronger adaptation to unseen generators, improved explanation supervision beyond external reward models, broader and harder benchmarks, and more advanced multimodal reasoning architectures or pretraining schemes. A plausible implication is that BusterX++ should be understood less as a terminal detector than as a reference design for cross-modal RL post-training of forensic MLLMs (Wen et al., 19 Jul 2025).