NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild

Aleksandr Gushchin; Khaled Abud; Ekaterina Shumitskaya; Artem Filippov; Georgii Bychkov; Sergey Lavrushkin; Mikhail Erofeev; Anastasia Antsiferova

cs.CVApr 13, 2026

NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild

Authors: Aleksandr Gushchin, Khaled Abud, Ekaterina Shumitskaya, Artem Filippov, Georgii Bychkov, Sergey Lavrushkin, Mikhail Erofeev, Anastasia Antsiferova, Changsheng Chen, Shunquan Tan, Radu Timofte, Dmitry Vatolin, Chuanbiao Song, Zijian Yu, Hao Tan, Jun Lan, Zhiqiang Yang, Yongwei Tang, Zhiqiang Wu, Jia Wen Seow, Hong Vin Koay, Haodong Ren, Feng Xu, Shuai Chen, Ruiyang Xia, Qi Zhang, Yaowen Xu, Zhaofan Zou, Hao Sun, Dagong Lu, Mufeng Yao, Xinlei Xu, Fei Wu, Fengjun Guo, Cong Luo, Hardik Sharma, Aashish Negi, Prateek Shaily, Jayant Kumar, Sachin Chaudhary, Akshay Dudhane, Praful Hambarde, Amit Shukla, Zhilin Tu, Fengpeng Li, Jiamin Zhang, Jianwei Fei, Kemou Li, Haiwei Wu, Bilel Benjdira, Anas M. Ali, Wadii Boulila, Chenfan Qu, Junchi Li

Abstract

This paper presents an overview of the NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild, held in conjunction with the NTIRE workshop at CVPR 2026. The goal of this challenge was to develop detection models capable of distinguishing real images from generated ones in realistic scenarios: the images are often transformed (cropped, resized, compressed, blurred) for practical usage, and therefore, the detection models should be robust to such transformations. The challenge is based on a novel dataset consisting of 108,750 real and 185,750 AI-generated images from 42 generators comprising a large variety of open-source and closed-source models of various architectures, augmented with 36 image transformations. Methods were evaluated using ROC AUC on the full test set, including both transformed and untransformed images. A total of 511 participants registered, with 20 teams submitting valid final solutions. This report provides a comprehensive overview of the challenge, describes the proposed solutions, and can be used as a valuable reference for researchers and practitioners in increasing the robustness of the detection models to real-world transformations.

Figures & tables

Figure 1: MICV method scheme. DINOv3-based detection framework.

Figure 2: Ant International method scheme.

Figure 3: TeleAI-TeleGuard method scheme. Pipeline of LPT. To achieve robust detection, the framework includes distortion simulation, foundation-model fine-tuning, and pairwise optimization.

Model	Input	Core change	Weight
M1	$256^{2}$	DINOv3-Huge baseline	0.3675
M2	$256^{2}$	Incremental data expansion	0.0735
M3	$256^{2}$	Stronger distortion augmentation	0.0490
M4	$448^{2}$	High-resolution branch	0.2100
M5	$378^{2}$	MetaCLIP2 Giant backbone	0.3000

Table 1: Summary of training paths and ensemble weights.

Dimension	Technical solution	Effect
GPU parallelism	5 independent CUDA streams	Full model overlap
Compute precision	AMP for M1–M4	$\sim$ 30% throughput gain
Data loading	Async I/O + prefetch=4	Reduced wait time
Pipeline	CPU/GPU overlap	Higher utilization

Table 2: Inference-time engineering optimizations.

	Open Test		Hidden Test		Average
Method	ROC AUC	Rob. ROC AUC	ROC AUC	Rob. ROC AUC	ROC AUC	Rob. ROC AUC
MICV	0.9978	0.9738	0.9969	0.9707	0.9974	0.9723
Ant International	0.9973	0.9731	0.9971	0.9711	0.9972	0.9721
TeleAI-TeleGuard	0.9762	0.9215	0.9809	0.9286	0.9786	0.9251
INTSIG	0.9810	0.9090	0.9895	0.9171	0.9897	0.9130
vincentlc	0.9497	0.8633	0.9557	0.8828	0.9527	0.8730

Table 3: Final results and rankings of the top-9 teams. Performance is reported on the open and hidden test sets using both clean ROC-AUC and Robust ROC-AUC. The final ranking is determined by the average Robust ROC-AUC across both splits.

Figure 6: Reagvis Labs method scheme. Overview of RAPID with staged logit fusion.

Model	Time (ms)	Peak VRAM (GB)	Fine-tuning
G4 (CLIP+GAPL)	$\sim$ 37	2.6	LoRA $r{=}32$ (attn+MLP)
SigLIP-v2	$\sim$ 26	1.8	LoRA $r{=}32$ (384px)
SRM ForensicCNN	$\sim$ 5	0.1	full training (FP32)
EVA-02 (original)	$\sim$ 25	1.3	LoRA $r{=}32$ + SwiGLU
EVA-02 (fixed)	$\sim$ 25	1.3	LoRA $r{=}32$ retrain
G4_v2 (CLIP+GAPL)	$\sim$ 37	2.6	LoRA continuation

Table 4: Per-branch inference cost and memory. Sequential loading keeps total peak memory below 4 GB.

Encoder	Paradigm	$d$	Tune	Head
CLIP ViT-L/14	VL-Contrastive	1024	LN	MLP-512
SigLIP ViT-L/16	VL-Contrastive	1024	LN	MLP-512
EVA02-Large	VL-Contrastive	1024	LN	MLP-512
DINOv2 ViT-L/14	Self-Supervised	1024	F	Linear
ConvNeXt-Large	Supervised CNN	1536	F	MLP-512
EfficientNet-V2-L	Supervised CNN	1280	F	MLP-512

Table 5: PRISM encoder pool.

d

: feature dimension. LN: LayerNorm-only tuning. F: frozen encoder.

Figure 9: Shallow Real method scheme. Overall pipeline of the proposed model.

Split	# of Images	Real/Fake Ratio	Labels Provided	Generator Models	Transformations
Train	$\sim$ 277K	$\sim$ 1:1.77	Yes	20 Models: • YOSO PixArt-512 • PixArt- $\alpha$ • PixArt- $\Sigma$ • Kandinsky 2 • Kandinsky 3 • Kolors • OmniGen • OmniGen 2 • Stable Diffusion 1.4 • Stable Diffusion 1.5 • Stable Diffusion 2.1 • Stable Diffusion XL 1.0 • SDXL Lightning • SDXL Turbo • Janus Pro 7B • Infinity 2B • Infinity 8B • Ovis Image • DeepFloyd IF • FLUX.1 Kontext Dev	12 Transformations: (Provided as distortion pipeline) • Gaussian Blur • Lens Blur • Color Shift • Color Saturation • JPEG Compression • White Noise • Impulse Noise • Brightness Increase • Brightness Decrease • Color Jitter • Color Quantization • Linear Contrast Change
Validation	10K	1:1	Yes (After Completion)	9 Models: • FLUX.1 Kontext Dev • SDXL Turbo • FLUX.1 Dev • Playground v2.5 • Lumina Image 2.0 • Qwen Image • Stable Diffusion 3 Medium • Ideogram v3 Turbo † • ImageGen-4 Fast †	19 Transformations: • Gaussian Blur • Lens Blur • Color Shift • Color Saturation • JPEG Compression • White Noise • Impulse Noise • Brightness Increase • Brightness Decrease • Color Jitter • Color Quantization • Linear Contrast Change • Motion Blur • Multiplicative Noise • Pixelation • RGB Channel Shift • Random Crop • Random Aspect Crop • Downscale
Validation Hard	2.5K	1:1	Yes (After Completion)	7 Models: • Playground v2.5 • SDXL Turbo • HiDream • FLUX.1 Schnell • Stable Diffusion 3.5 Large Turbo • Nano Banana † • Seedream 4 †	19 Transformations: • Gaussian Blur • Lens Blur • JPEG Compression • White Noise • Impulse Noise • Color Quantization • Multiplicative Noise • RGB Channel Shift • Random Crop • Random Aspect Crop • Neural Image Compression (JPEG AI) • Random Tone Curve • CLAHE • ISO Noise • Perspective Transform • Multiple Compressions (JPEG) • Multiple Compressions (JPEG + JPEG AI) • Watermark Attack (Adv. Embedding, CLIP/ResNet) • Downscale

Table 6: Detailed information on Train, Validation, and Validation Hard dataset splits. † denotes a proprietary generation model used via API.

Split	# of Images	Real/Fake Ratio	Labels Provided	Generator Models	Transformations
Test (Public)	2.5K	1:1	Yes (After Completion)	10 Models: • HiDream • FLUX.1 Schnell • Stable Diffusion 3.5 Large • FLUX Krea • Z-Image Turbo • Nano Banana Pro † • FLUX-2 Max † • ImageGen-4 Ultra † • Seedream 5 Lite † • Grok Imagine Image †	22 Transformations: • Color Saturation • Brightness Increase • Lens Blur • JPEG Compression • Impulse Noise • RGB Channel Shift • Random Crop • Random Aspect Crop • Neural Image Compression (JPEG AI) • Random Tone Curve • CLAHE • ISO Noise • Perspective Transform • Multiple Compressions (JPEG) • Multiple Compressions (JPEG + JPEG AI) • Watermark Attack (Adv. Embedding, CLIP/ResNet) • JPEG 2000 • Watermark Attack (WMForger) • Neural Image Compression (Cheng2020) • Shot Noise • Downscale • Invisible Watermark Insertion (1 of 6 algorithms)
Test (Private)	2.5K	$\sim$ 1:1	No	10 Models: • HiDream • Stable Diffusion 3.5 Large Turbo • FLUX.1 Dev SRPO • Z-Image Turbo • Kandinsky 5 • Nano Banana 2 † • GPT Image 1.5 † • ImageGen-4 Ultra † • Seedream 5 Lite † • Grok Imagine Image †	24 Transformations: • Color Saturation • Brightness Increase • Lens Blur • JPEG Compression • Impulse Noise • RGB Channel Shift • Random Crop • Random Aspect Crop • Neural Image Compression (JPEG AI) • Random Tone Curve • CLAHE • ISO Noise • Perspective Transform • Multiple Compressions (JPEG) • Multiple Compressions (JPEG + JPEG AI) • Multiple Compressions (JPEG + JPEG 2000) • Watermark Attack (Adv. Embedding, CLIP/ResNet) • JPEG 2000 • Watermark Attack (WMForger) • Neural Image Compression (Cheng2020) • Shot Noise • Glass Blur • Downscale • Invisible Watermark Insertion (1 of 7 algorithms)

Table 7: Detailed information on Test (public) and Test (private) dataset splits. † denotes a proprietary generation model used via API.

AI-Generated Image Detection

AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.

See all 194 papers

HEDGE: Heterogeneous Ensemble for Detection of AI-GEnerated Images in the Wild
Apr 4, 2026Fei Wu, Dagong Lu, Mufeng Yao +2AI-Generated Image Detection Ensemble Learning
Shanghai Jiao Tong University · INTSIG Information
REALIS: A Curated Dataset for Studying the Challenges of AI Image Detection
Sep 26, 2026Aleksandr Gushchin, Khaled Abud, Georgii Bychkov +5AI-Generated Image Detection
MSU Institute for Artificial Intelligence · Trusted AI Research Center RAS · Lomonosov Moscow State University
RAID: Towards Robust AI-Generated Image Detection with Bit-Reversed Images
Jul 31, 2026Renxi Cheng, Jie Gui, Hongsong WangImage Forgery Detection Digital Image Forensics
School of Cyber Science and Engineering, Southeast University, Nanjing 210096, China · Purple Mountain Laboratories, Nanjing 210000, China · Engineering Research Center of Blockchain Application, Supervision And Management (Southeast University), Ministry of Education, China +2

Apr 4, 2026J/K move · Enter open · S save

HEDGE: Heterogeneous Ensemble for Detection of AI-GEnerated Images in the Wild

Fei Wu, Dagong Lu, Mufeng Yao, Xinlei Xu+1

Shanghai Jiao Tong University · INTSIG Information

AI-Generated Image Detection Ensemble Learning

Open paper AskPDF arXiv

Robust detection of AI-generated images in the wild remains challenging due to the rapid evolution of generative models and varied real-world distortions. We argue that relying on a single training regime, resolution, or backbone is insufficient to handle all conditions, and that structured heterogeneity across these dimensions is essential for robust detection. To this end, we propose HEDGE, a Heterogeneous Ensemble for Detection of AI-GEnerated images, that introduces complementary detection routes along three axes: diverse training data with strong augmentation, multi-scale feature extraction, and backbone heterogeneity. Specifically, Route~~A progressively constructs DINOv3-based detectors through staged data expansion and augmentation escalation, Route~~B incorporates a higher-resolution branch for fine-grained forensic cues, and Route~C adds a MetaCLIP2-based branch for backbone diversity. All outputs are fused via logit-space weighted averaging, refined by a lightweight dual-gating mechanism that handles branch-level outliers and majority-dominated fusion errors. HEDGE achieves 4th place in the NTIRE 2026 Robust AI-Generated Image Detection in the Wild Challenge and attains state-of-the-art performance with strong robustness on multiple AIGC image detection benchmarks.

Abstract

AI-Generated Image Detection

Explore similar work

HEDGE: Heterogeneous Ensemble for Detection of AI-GEnerated Images in the Wild

REALIS: A Curated Dataset for Studying the Challenges of AI Image Detection

RAID: Towards Robust AI-Generated Image Detection with Bit-Reversed Images