Skip to content

Semantic Visual Anomaly Detection and Reasoning in AI-Generated Images

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; VLM-based semantic anomaly detection in AI-generated image outputs

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper formalizes semantic anomaly detection in AI-generated images as a structured reasoning task, introducing AnomReason—a 21,539-image benchmark annotated with quadruples (Name, Phenomenon, Reasoning, Severity)—and AnomAgent, a modular multi-agent pipeline that produced ~4.17B GPT-4o tokens of candidate annotations. A LoRA-tuned 7B model (AnomReasonor-7B) trained on AnomReason surpasses GPT-4o on the proposed SemAP metrics, demonstrating that task-aligned supervision substantially closes the gap to closed-source systems.

Problem

Existing AIGC forensics methods focus on low-level surface artifacts (texture statistics, shadow patterns) imperceptible to humans; no benchmark or framework existed to evaluate content-level semantic anomalies—violations of physical laws, commonsense logic, or anatomical plausibility—that are visible to humans and directly affect trust in AIGC media. Generating structured, explainable anomaly descriptions (not just binary labels) at scale is also unsolved.

Method

AnomAgent decomposes anomaly annotation into three sequential stages:

  1. Visual Entity Parsing — ObjectPerceiver runs \(T\) times with varied prompts and takes the union \(O = \bigcup_{t=1}^T O^{(t)}\) to enumerate all semantic entities in the image.

  2. Multi-Perspective Anomaly Mining — For each object \(o_i \in O\):

  3. AttributeAnalyzer identifies intra-object inconsistencies: \(C^{(i)}_{\text{attr}} = \text{AttributeAnalyzer}(o_i)\)

  4. RelationReasoner identifies inter-object implausibilities conditioned on attribute priors: \(C^{(i)}_{\text{rel}} = \text{RelationReasoner}(o_i, O \setminus \{o_i\}, C^{(i)}_{\text{attr}})\)

  5. Anomaly Consolidation and Structuring — AnomalyIntegrator deduplicates candidates; AnomalyFormatter produces structured tuples \(\mathcal{A} = \{(y_i, o_i, r_i, v_i)\}\) (Name, Phenomenon, Reasoning, Severity \(v \in [0,100]\)).

Annotations undergo lightweight human-in-the-loop (HITL) screening: each candidate is accepted/rejected/unsure by a trained annotator; only accepted items survive. The benchmark (AnomReason) is built from 21,539 photorealistic images (Midjourney 9,911 + Flux 6,983 + SD3.5 4,645), yielding ~5.9 valid anomalies per image post-HITL.

Evaluation uses semantic matching via BERTScore at thresholds \(\tau \in \{0.7, 0.8, 0.9\}\), computing per-image Average Precision and F1, then averaging: \(\text{SemAP}_v = \frac{1}{|D|}\sum_I \text{AP}_v(I)\), \(\text{SemF1}_v = \frac{1}{|D|}\sum_I \text{F1}_v(I)\), evaluated over Phenomenon (Phe), Reasoning (Rea), and Full (Phe+Rea) views.

A fine-tuned model, AnomReasonor-7B, is obtained by LoRA fine-tuning Qwen2.5-VL-7B on the train split of AnomReason.

Key Contributions

  • Formalization of semantic visual anomaly detection and reasoning as a structured output task with quadruple annotations (Name, Phenomenon, Reasoning, Severity).
  • AnomAgent: a modular, three-stage multi-agent pipeline for scalable semantic anomaly annotation with HITL verification.
  • AnomReason: first large-scale benchmark (21,539 images, ~174,872 raw candidates, ~5.9 post-HITL anomalies/image) for content-aware semantic anomaly detection in AIGC.
  • SemAP / SemF1: novel structure-aware metrics based on BERTScore matching for evaluating anomaly detection and reasoning quality.
  • AnomReasonor-7B: a LoRA fine-tuned VLM that achieves state-of-the-art on AnomReason, surpassing GPT-4o on SemAP metrics.

Results

  • AnomReasonor-7B achieves SemAP\(_\text{Full}\) = 0.5162 and SemF1\(_\text{Full}\) = 0.5009, outperforming all baselines on SemAP metrics.
  • Best open-source baseline: Qwen2.5-VL-72B (SemAP\(_\text{Full}\) = 0.4568, SemF1\(_\text{Full}\) = 0.4104).
  • Best proprietary baseline: GPT-4o (SemAP\(_\text{Full}\) = 0.4727, SemF1\(_\text{Full}\) = 0.5109); AnomReasonor-7B exceeds GPT-4o on all SemAP metrics but is slightly behind on SemF1\(_\text{Full}\) (0.5009 vs. 0.5109).
  • AnomReasonor-7B surpasses GPT-o3, GPT-5, Gemini-2.5-Pro, and all open-source VLMs tested on SemAP\(_\text{Full}\).
  • Most off-the-shelf VLMs score SemAP\(_\text{Full}\) < 0.42, with larger gaps on reasoning (Rea) than phenomenon (Phe)—e.g., InternVL3-8B: SemAP\(_\text{Phe}\) = 0.4552 vs. SemAP\(_\text{Rea}\) = 0.3676.
  • AnomReasonor-7B shows the most balanced Phe/Rea profile (0.5221 / 0.5130), reflecting benefit of structured supervision.

Limitations

  • Annotations are sourced from a single automated pipeline (GPT-4o-based AnomAgent) with lightweight rather than dense human review; residual biases from GPT-4o's knowledge may propagate into the benchmark.
  • HITL protocol uses a simple accept/reject/unsure binary per annotator with no inter-annotator agreement reported, leaving reliability of the verification unquantified.
  • Benchmark covers only three generative model families (Midjourney, SD3.5, Flux); generalization to other generators (e.g., video diffusion, newer closed-source models) is untested.
  • SemAP/SemF1 rely on BERTScore with a lightweight distilbert model, which may not capture fine-grained semantic distinctions in anomaly descriptions.
  • The paper is truncated before reporting full explainable deepfake detection (Table 2) and generator audit results, so the downstream application numbers are not available for review.

Relevance to Vision-Language Models

This work directly probes the semantic reasoning capabilities of VLMs by constructing a benchmark that requires models to detect and articulate violations of commonsense, physics, and spatial logic in images—tasks that go well beyond standard VQA or captioning. The consistent gap between off-the-shelf VLMs (even 72B-scale models) and the task-aligned 7B fine-tune quantifies a specific reasoning deficit in current VLMs: while models can describe surface appearance, they struggle to reason about semantic plausibility. The structured (Name, Phenomenon, Reasoning, Severity) annotation format and accompanying SemAP/SemF1 metrics provide a reusable evaluation protocol for assessing and improving commonsense-grounded reasoning in VLMs applied to AIGC forensics, deepfake detection, and generative model auditing.