Research map / Image generation / Autoregressive and latent generation
ReAlign: Generalizable Image Forgery Detection via Reasoning-Aligned Representation
Qing Huang, Zhipei Xu, Xuanyu Zhang, Xiangyu Yu, Jian Zhang
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 |
|---|---|---|---|
| use reasoning text instead of image captions as alignment text (ablation) | ReAlign · UltraSynth-10k | mean accuracy: 97.09 -> 91.63 (-5.46%) when replaced by captions | Improved |
| add class-label prefix to reasoning text (ablation) | ReAlign · UltraSynth-10k | mean accuracy: 96.87 -> 97.09 (0.22% decrease without prefix) | Improved |
| reasoning text alone versus caption alone (ablation) | ReAlign · UltraSynth-10k | mean accuracy: 88.32 -> 96.87 | Improved |
| reasoning text alone versus class label alone (ablation) | ReAlign · UltraSynth-10k | mean accuracy: 91.33 -> 96.87 | Improved |
| use joint optimization instead of sequential (LoRA) (ablation) | ReAlign · UltraSynth-10k | mean accuracy: 84.08 -> 97.09 (-13.01%) | Improved |
| use joint optimization instead of sequential (full fine-tuning) (ablation) | ReAlign · UltraSynth-10k | mean accuracy: 79.07 -> 94.69 | Improved |
| use LoRA instead of full fine-tuning of image encoder (ablation) | ReAlign · UltraSynth-10k | mean accuracy: 94.69 -> 97.09 (-2.40%) | Improved |
| add reasoning-alignment contrastive loss to frozen encoder (ablation) | ReAlign · UltraSynth-10k | mean accuracy: 89.15 -> 97.09 (-7.94%) | Improved |
| add reasoning-alignment contrastive loss to LoRA encoder (ablation) | ReAlign · UltraSynth-10k | mean accuracy: 93.68 -> 97.09 (-3.41%) | Improved |
| distill AIGI-R1 reasoning into CLIP (ReAlign vs AIGI-R1) | ReAlign · AIGCDetectBenchmark | mean accuracy: 91.77 -> 96.14 | Improved |
| ReAlign versus AIDE | ReAlign · AIGI-Holmes | mean accuracy: 97.00 -> 99.44 | Improved |
| ReAlign versus AIGI-R1 on unseen proprietary generators | ReAlign · UltraSynth-10k | mean accuracy: 96.42 -> 97.09 | Mixed |
Later papers that compared with this method
2 later accepted papers report ReAlign as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- EasyTune: Efficient Step-Aware Fine-Tuning for Diffusion-Based Motion Generation (ICLR 2026 · better on 1 of 1 shared result)
- MotionRFT: Unified Reinforcement Fine-Tuning for Text-to-Motion Generation (TPAMI 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.89)
- Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.88)
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.88)
- Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation (NeurIPS 2025 · similarity 0.87)
- A Hitchhiker's Guide to Fine-Grained Face Forgery Detection Using Common Sense Reasoning (NeurIPS 2024 · similarity 0.87)
- TextShield-R1: Reinforced Reasoning for Tampered Text Detection (AAAI 2026 · similarity 0.87)
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.
