Research map / Detection and segmentation / Open-set and few-shot detection
From Masks to Pixels and Meaning: A New Taxonomy, Benchmark and Metrics for VLM Image Tampering
Xinyi Shang, Yi Tang, Jiacheng Cui, Ahmed Elhagry, Salwa Al Khatib, Sondos Bsharat, Jiacheng Liu, Xiaohan Zhao, Jing-Hao Xue, Hao Li, Salman Khan, Zhiqiang Shen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace mask supervision with pixel-difference supervision (Lite) | SIDA-7B · PIXAR test set | IoU: 11.8 -> 15.0 | Improved |
| replace mask supervision with pixel-difference supervision (Lite) | SIDA-13B · PIXAR test set | IoU: 10.8 -> 15.4 | Improved |
| train on the full PIXAR training set instead of the mask-guided Lite subset | PIXAR-7B · PIXAR test set | IoU: 15.0 -> 18.1 | Improved |
| scale model from 7B to 13B | PIXAR · PIXAR test set | IoU: 18.1 -> 19.3 | Improved |
| use PIXAR instead of FakeShield | FakeShield · PIXAR test set | IoU: 9.3 -> 18.1 | Improved |
| use PIXAR-7B instead of SIDA-7B for real vs tampered detection | SIDA-7B · PIXAR test set | Overall F1-score: 25.1 -> 69.7 | Improved |
| raise training tau from 0.05 to 0.1 (ablation) | PIXAR-7B · PIXAR test set (eval tau=0.05) | IoU: 18.1 -> 12.6 | Hurt |
| raise training tau from 0.05 to 0.2 (ablation) | PIXAR-7B · PIXAR test set (eval tau=0.05) | IoU: 18.1 -> 8.7 | Hurt |
| use tau=0.1 for both training and evaluation (ablation) | PIXAR-7B · PIXAR test set (symmetric tau) | IoU: 18.1 -> 12.0 | Hurt |
| add Dice loss (lambda_dice 1.0 versus 0.0) (ablation) | PIXAR-7B · PIXAR test set | IoU: 10.8 -> 18.1 | Improved |
| add Dice loss (lambda_dice 1.0 versus 0.0) (ablation) | PIXAR-7B · PIXAR test set | Top-1 accuracy: 35.3 -> 36.2 | Improved |
| set lambda_sem 0.5 instead of 0.1 or 1.0 (ablation) | PIXAR-7B · PIXAR test set | Top-1 accuracy: 35.1 -> 36.2 | Mixed |
Closest papers by meaning
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.89)
- Toward Real-world Text Image Forgery Localization: Structured and Interpretable Data Synthesis (NeurIPS 2025 · similarity 0.87)
- Revisiting Tampered Scene Text Detection in the Era of Generative AI (AAAI 2025 · similarity 0.87)
- Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.86)
- SUMI-IFL: An Information-Theoretic Framework for Image Forgery Localization with Sufficiency and Minimality Constraints (AAAI 2025 · similarity 0.85)
- A Hitchhiker's Guide to Fine-Grained Face Forgery Detection Using Common Sense Reasoning (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
