💻 computer science

VIGIL: Tackling Hallucination Detection in Image Recontextualization

This paper introduces VIGIL, a novel benchmark dataset and multi-stage detection framework that categorizes hallucinations in image recontextualization into five fine-grained fidelity types to enable more precise error identification and explanation than existing methods.

Original authors: Joanna Wojciechowicz, Maria Łubniewska, Jakub Antczak, Justyna Baczyńska, Wojciech Gromski, Wojciech Kozłowski, Maciej Zieba

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Joanna Wojciechowicz, Maria Łubniewska, Jakub Antczak, Justyna Baczyńska, Wojciech Gromski, Wojciech Kozłowski, Maciej Zieba

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a director on a movie set, but instead of hiring actors, you are commanding a magical, super-intelligent robot painter. You hand the robot a photo of your favorite red sneaker and a photo of a sunny park, then say, "Put the sneaker in the park." The robot whirs, paints a masterpiece, and hands it back. But sometimes, the robot gets a little too creative. Maybe it turns your red sneaker into a blue boot, or it forgets to paint the sneaker entirely, or it places the shoe floating in mid-air like a ghost. This is the world of multimodal image recontextualization, a fancy term for using Artificial Intelligence (AI) to take an object from one picture and drop it into a new scene based on your instructions.

For a long time, scientists have been trying to figure out how to spot when these AI artists make mistakes, which are called hallucinations. Think of a hallucination not as a dream, but as a lie the computer tells itself. In the past, researchers mostly looked at simple text-to-image tasks (where you type a description and get a picture). They built tools that gave a single "score" to say how good the picture was, kind of like a teacher giving a test a grade of "B-". But that doesn't tell you why the shoe is blue instead of red, or why the shadow is pointing the wrong way. As AI gets better at making realistic images, we need a way to catch these specific, sneaky errors so we can trust the pictures for things like advertising or design.

Enter VIGIL, a new system created by a team of researchers that acts like a super-strict, detail-oriented art critic. Instead of just giving the AI a grade, VIGIL breaks down the picture into five specific categories of mistakes and explains exactly what went wrong. The researchers built a massive dataset of 1,269 examples where they manually checked AI-generated images to see what kinds of errors happened most often. They found that AI struggles in five main ways: it might change the look of the object (Object Visual Fidelity), mess up the background (Background Fidelity), put the object in the wrong spot (Spatial & Instructional Fidelity), make the object look like it doesn't belong physically (Physical & Integration Fidelity), or simply forget to draw the object at all (Object Omission).

The paper introduces a clever pipeline that uses a team of open-source AI tools to check the picture step-by-step. First, it uses a "segmentation" tool (like a digital pair of scissors) to cut out the objects. Then, it compares the cut-out pieces to the original reference photos to see if the shape or color changed. It checks if the background stayed the same, if the lighting and shadows make sense, and if the object is actually there. If the AI made a mistake, VIGIL doesn't just say "Error"; it writes a note, like, "The chair is now a rocking chair, but you asked for a wooden armchair," or "The sofa is floating in the air."

When the researchers tested their system, they found that this step-by-step approach was very good at catching errors, often doing better than other open-source AI detectors. In fact, on furniture images, VIGIL was the best at spotting mistakes. However, the paper notes that the system isn't perfect yet. It sometimes misses tiny details, like a distorted license plate on a car or a blurry logo on an electronic device, because those details are too small for the AI to see clearly after being cropped out. The authors suggest that while their method is a big step forward, the biggest challenge now is making sure the AI can zoom in and check those tiny, high-definition details without losing the big picture. The ultimate goal is to move from just guessing if an image is "good" to knowing exactly why it is broken, so we can fix it and trust what we see.

Technical Summary: VIGIL – Tackling Hallucination Detection in Image Recontextualization

Problem Statement
Conditional image generation, particularly image recontextualization (synthesizing new scenes by combining text with specific visual subjects from reference images), has advanced rapidly with large multimodal models (LMMs). However, these models frequently produce hallucinations—visual artifacts, identity shifts, and semantic inconsistencies unsupported by input prompts or references. While hallucination detection is well-studied in Text-to-Image (T2I) tasks, existing methods for multimodal recontextualization are limited. Current approaches typically report coarse, opaque numerical scores (e.g., 0–10) with unstructured textual descriptions, failing to distinguish between specific error types such as identity distortion versus environmental mismatch. This lack of fine-grained, interpretable evaluation hinders the diagnostic analysis required to deploy these systems in production environments like automated advertising and digital content creation.

Methodology
The authors propose VIGIL (Visual Inconsistency & Generative In-context Lucidity), a framework comprising a benchmark dataset and a multi-stage detection pipeline designed to decompose hallucinations into five distinct categories.

Taxonomy of Hallucinations: The framework defines five non-overlapping categories:
- Object Visual Fidelity: Deviations in the visual appearance (texture, material, color, geometry) of inserted objects relative to references.
- Background Fidelity: Alterations to the surrounding environment (walls, lighting, preexisting furniture) that should remain unchanged.
- Spatial & Instructional Fidelity: Failures in spatial reasoning or compliance with high-level editing instructions (e.g., incorrect placement, failure to replace/remove objects, introduction of unintended objects).
- Physical & Integration Fidelity: Violations of real-world physical consistency, including lighting, shadows, perspective, scale, and occlusion.
- Object Omission: The complete absence of objects explicitly required by the input instructions.
Detection Pipeline: Unlike "all-in-one" detectors, VIGIL employs a coordinated set of open-source models to process images through specialized steps:
- Object Extraction: Qwen3.5 parses the prompt to identify subjects, and Segment Anything Model 3 (SAM 3) generates masks for these entities in both reference and generated images.
- Object Fidelity Verification: DINO v3 generates embeddings for segmented objects. A Linear Sum Assignment algorithm matches reference objects to generated crops based on cosine similarity ( $S_{cos} > \tau$ ). Mismatches or missing matches are flagged.
- Background Fidelity Verification: Objects are masked out (with an extended margin to exclude contact shadows) to isolate the background. The pipeline uses direct comparative analysis and difference-guided localization to identify structural discrepancies, prompting a Vision-Language Model (VLM) to distinguish major errors from natural integration artifacts.
- Physical Fidelity Verification: The VLM evaluates physical plausibility at two granularities: object-level (lighting, shadows, scale relative to context) and scene-level (global incoherence, floating objects, inconsistent light sources).
- Spatial & Instructional Verification: The prompt is distilled into atomic constraints (positional relations or replacement directives). The VLM evaluates each constraint individually against the generated image to detect misplacement, replacement failures, or unwanted objects.

Dataset
The authors introduce the VIGIL dataset, containing 1,269 manually annotated samples across five product domains: clothing, furniture, cosmetics, electronics, and cars. The dataset includes reference images, background images, textual prompts, and the resulting recontextualized images generated by Gemini 2.5 Flash Image. Annotations are provided in JSON format, labeling specific hallucination classes and providing explanatory rationales. The dataset contains 1,024 images with hallucinations (80.7%) and 245 clean samples. Inter-annotator agreement was measured at 0.8129 (Positive Specific Agreement) using an LLM-as-a-Judge framework.

Results
The pipeline was evaluated using a leave-one-domain-out strategy, comparing VIGIL against standalone VLMs (Gemini 2.5 Flash, Qwen3.5-9B, Gemma 3 12B IT) and adapted baselines (VIEScore, CIGEval). Evaluation metrics included macro-F1 for multi-label classification and an LLM-as-a-Judge semantic evaluation.

Performance: VIGIL achieved the best macro-F1 scores among non-Gemini methods in the furniture category (0.5064) and competitive results in clothing, cosmetics, and electronics. It consistently outperformed the adapted baselines (VIEScore and CIGEval), which scored significantly lower across all domains.
Comparison with Proprietary Models: While Gemini 2.5 Flash achieved the highest overall scores, the authors note this may be due to the judge model (also Gemini) favoring similar output styles. Crucially, VIGIL's decomposed pipeline using Qwen3.5-9B often outperformed Qwen3.5-9B when used as a standalone detector, suggesting that solving focused subtasks improves reliability.
Error Analysis: The authors identified that performance gaps in "Cars" and "Electronics" were primarily due to the difficulty of detecting subtle, localized text distortions (e.g., license plates) and fine-grained details (logos, ports) that are often blurred or downsampled in generated crops, indicating a bottleneck in local verification rather than the overall decomposition strategy.

Significance and Claims
The paper claims that VIGIL addresses a critical gap in the field by moving beyond opaque, single-score evaluations to provide structured, fine-grained, and explainable feedback. The primary contributions are:

A benchmark dataset of 1,269 manually annotated samples with ground truth across five hallucination categories.
A decomposed detection method that automatically identifies and describes hallucinations, returning textual explanations for each error rather than isolated scores.
A taxonomy specifically tailored for image recontextualization, supporting the diagnosis of model reliability for practical use cases.

The authors emphasize that their approach provides human-interpretable feedback, which is essential for moving generative models from research to production, particularly in domains where product attributes must be preserved with high fidelity. They explicitly state that the code and dataset will be made publicly available.

More like this