PAPER_TITLE

FIRST_AUTHOR_LAST, FIRST_AUTHOR_FIRST; SECOND_AUTHOR_LAST, SECOND_AUTHOR_FIRST

DailyBench: A Unified Benchmark for AI-Generated and Manipulated Images from Modern Generative Models

Xin Jiang¹, Hao Tang², Junyao Gao³ Meiqi Cao⁴ Fei Shen⁵ Dongming Zhang¹ Yongdong Zhang⁶

¹State Key Laboratory of Content Cognition, People's Daily Online ²Nanyang Technological University ³Tongji University ⁴Nanjing University of Science and Technology ⁵National University of Singapore ⁶University of Science and Technology of China

arXiv

Dataset

Abstract

Recent advances in generative models have shifted AI-generated image detection from identifying easily distinguishable, fully synthetic images to identifying highly realistic content generated by both modern generation and localized manipulation pipelines. However, existing detection benchmarks are often built with outdated generative models and primarily emphasize full-image synthesis, creating a growing mismatch between benchmark data and the images encountered in real-world generation and editing scenarios. To bridge this gap, we introduce DailyBench, a high-quality unified benchmark for evaluating whether AI-generated image detectors can generalize across both modern full-image synthesis and localized object-level manipulation. DailyBench contains two complementary subsets: FakeBench, which includes high-quality images synthesized by recent open-source and commercial generative models, and ManipulationBench, which introduces challenging object-level edits applied to real images using advanced image-conditional models. This design makes DailyBench a realistic testbed for studying both generator-level generalization and manipulation-aware detection under subtle local edits. In addition to the benchmark, we propose FPD (Fake-Preferences Detector), a diagnostic-driven strong baseline that encourages forgery-sensitive representation learning. FPD combines a dual-pathway architecture with a cascaded compression classifier to progressively extract and compact forgery-relevant cues, and further introduces a fake-preference sampler to reduce excessive bias toward the real class. Experiments on DailyBench reveal substantial robustness gaps in current detectors: methods reporting 91–96% balanced accuracy on GenImage drop to 60–76% on FakeBench and 54–66% on ManipulationBench. These results show that existing detectors remain poorly generalized to realistic synthesis and localized manipulation, highlighting DailyBench as a rigorous testbed for developing robust and manipulation-aware AI-generated image detection methods.

Dataset Sample Examples

FakeBench.

ManipulationBench.

Dataset Statistics

Experimental Results

FakeBench Results

Methods	SD3.5		FLUX.1		FLUX.2		Z-Image		Qwen-Image		N.B. 2		GPT-I. 2		Avg.
Methods	R.Acc	F.Acc	R.Acc	F.Acc	R.Acc	F.Acc	R.Acc	F.Acc	R.Acc	F.Acc	R.Acc	F.Acc	R.Acc	F.Acc	Avg.
UniFD	98.90	2.70	98.72	0.19	98.83	0.50	98.80	1.03	98.63	0.51	99.59	0.00	98.54	0.00	49.78
LOTA	99.83	0.13	99.82	0.02	99.89	0.35	99.81	0.11	99.86	0.03	100.00	0.00	99.64	98.91	57.03
DRCT	85.16	29.05	85.04	21.78	84.72	27.65	85.26	41.82	85.17	30.12	82.45	22.86	86.50	22.26	56.42
FatFormer	99.64	2.12	99.55	0.04	99.54	0.22	99.59	0.94	99.54	0.49	100.00	0.00	99.64	12.04	50.95
NPR	61.85	62.66	61.59	79.09	60.40	66.24	61.55	73.95	60.94	80.88	63.67	59.18	62.04	91.24	67.52
SAFE	94.94	2.48	95.02	3.15	95.23	2.41	94.68	1.88	94.91	3.88	97.14	5.31	94.16	77.74	54.49
AIDE	94.00	13.29	94.29	20.40	93.88	12.15	94.16	21.95	94.22	22.63	95.92	5.71	94.89	15.33	55.34
Effort	38.74	70.04	39.74	74.31	38.30	69.16	39.45	82.49	39.09	88.76	40.00	85.31	40.51	97.08	60.21
ForgeLens	99.96	0.19	99.93	0.66	99.90	0.31	99.92	0.23	99.91	7.37	100.00	0.00	100.00	98.91	57.66
LTD	99.77	2.64	99.74	1.36	99.77	2.48	99.72	5.06	99.77	4.88	100.00	0.41	100.00	68.98	56.04
Aligned	99.81	49.30	99.81	16.27	99.84	3.01	99.86	12.12	99.85	21.39	100.00	42.04	99.27	10.58	60.94
B-Free	89.22	55.74	88.96	8.49	90.01	5.35	89.06	37.28	89.39	2.78	91.02	0.82	89.05	0.73	52.70
RPOBE	87.67	96.65	87.85	19.18	88.46	28.66	87.58	71.62	88.13	77.43	89.80	42.45	86.86	36.13	70.60
OmniAID	88.87	49.04	88.69	37.20	88.95	16.40	88.12	39.57	88.69	45.43	89.83	54.87	89.05	29.19	63.84
DDA	86.19	76.34	85.81	30.97	86.92	22.70	85.99	63.12	86.44	56.47	86.12	45.31	84.31	51.09	70.27
DINO-Det	94.11	73.86	93.96	86.82	94.24	51.48	93.88	57.26	94.23	58.92	94.69	29.39	93.80	54.38	76.50
FPD (Ours)	83.34	84.63	87.67	89.78	89.61	63.21	88.70	68.39	83.68	71.90	94.69	29.39	90.62	53.76	76.78

Table 1. Accuracy (%) comparison on FakeBench. We report Fake Accuracy (F.Acc), Real Accuracy (R.Acc), and Average Accuracy (Avg.) to evaluate detector performance and generalization capability. N.B. 2 and GPT-I. 2 denote Nano Banana 2 and GPT-Image 2, respectively.

ManipulationBench Results

Methods	FLUX-Fill Random	FLUX-Fill Object	FLUX.2-klein	Qwen-Edit	Step-Edit	N.B. 2	GPT-I. 2	Avg.
UniFD	50.16 / 50.16	39.02 / 50.83	38.43 / 50.72	39.61 / 50.39	39.10 / 50.72	37.64 / 50.26	37.45 / 50.23	40.20 / 50.47
LOTA	49.92 / 49.92	37.79 / 49.92	37.31 / 49.92	38.88 / 49.97	37.84 / 49.89	37.32 / 50.00	100.00 / 100.00	48.43 / 57.08
DRCT	51.00 / 51.00	42.51 / 50.86	41.60 / 50.56	44.79 / 52.17	41.31 / 50.13	49.92 / 57.21	57.82 / 63.99	46.99 / 53.70
FatFormer	49.86 / 49.86	37.90 / 49.95	37.56 / 50.11	39.06 / 50.05	38.36 / 50.26	37.32 / 50.00	45.91 / 56.96	40.85 / 51.02
NPR	45.09 / 45.09	40.99 / 46.02	44.98 / 49.81	52.40 / 55.17	44.00 / 48.77	52.70 / 55.53	81.92 / 80.33	51.72 / 54.38
SAFE	52.22 / 52.22	41.63 / 51.94	39.00 / 49.99	38.99 / 48.96	37.40 / 48.26	37.97 / 49.46	84.94 / 86.67	47.45 / 55.35
AIDE	51.13 / 51.13	39.86 / 50.65	38.16 / 49.63	50.45 / 58.63	48.31 / 57.54	43.54 / 54.43	51.65 / 60.43	46.15 / 54.63
Effort	45.60 / 45.60	46.81 / 45.70	52.62 / 50.67	59.98 / 56.98	56.05 / 53.52	64.16 / 60.85	75.90 / 68.99	57.30 / 54.61
ForgeLens	50.01 / 50.01	37.85 / 49.98	37.46 / 50.05	41.25 / 51.89	39.41 / 51.17	41.41 / 53.17	91.82 / 93.49	48.45 / 57.10
LTD	49.88 / 49.88	37.76 / 49.88	37.37 / 49.97	39.00 / 50.04	37.95 / 49.98	37.48 / 50.04	66.57 / 73.40	43.71 / 53.31
Aligned	50.51 / 50.51	38.46 / 50.44	37.94 / 50.42	67.27 / 73.18	38.51 / 50.43	55.16 / 64.14	45.91 / 56.96	47.68 / 56.58
B-Free	62.36 / 62.36	57.67 / 69.97	57.12 / 63.57	77.92 / 79.95	58.59 / 64.61	34.70 / 45.43	42.47 / 51.86	55.83 / 62.53
RPOBE	61.67 / 61.67	56.58 / 62.93	59.58 / 65.46	92.10 / 91.51	55.12 / 61.83	65.47 / 69.61	80.34 / 82.47	67.26 / 70.78
OmniAID	52.66 / 52.66	44.34 / 53.36	47.75 / 56.40	88.10 / 88.50	54.57 / 61.46	67.32 / 71.82	56.55 / 63.08	58.75 / 63.89
DDA	54.54 / 54.54	46.52 / 54.82	55.51 / 62.29	89.44 / 89.20	56.77 / 62.88	56.63 / 62.48	80.20 / 81.96	62.80 / 66.88
DINO-Det	52.22 / 52.22	42.41 / 52.97	43.35 / 53.99	62.95 / 69.01	45.69 / 55.53	40.43 / 52.04	65.14 / 71.55	50.31 / 58.18
FPD (Ours)	52.56 / 52.56	46.01 / 54.09	49.36 / 57.82	79.02 / 80.47	54.17 / 60.70	46.25 / 53.22	73.86 / 77.16	57.32 / 62.29

Table 2. Average / Balanced accuracy (%) comparison on ManipulationBench. FLUX-Fill Random and FLUX-Fill Object denote FLUX-Fill using random and object masks, respectively. Qwen-Edit and Step-Edit denote Qwen-Image-Edit and Step1X-Edit-v1p2, respectively.

For more detailed results, please see the paper .

BibTeX

@article{dailybench,
  title={DailyBench: A Unified Benchmark for AI-Generated and Manipulated Images from Modern Generative Models},
  author={Xin Jiang, Hao Tang, Junyao Gao, Meiqi Cao, Fei Shen, Dongming Zhang, Yongdong Zhang},
  journal={arXiv preprint arXiv:2607.24016},
  year={2026}
}