TeleAI's LPT Beats Top Detectors Spotting AI Images Mangled by Social Media
A LoRA-based pairwise training recipe on EVA-CLIP earns third place at NTIRE 2026 by keeping AI-image detectors reliable under heavy real-world distortions.
Takeaways
- TeleAI's LPT placed third at NTIRE 2026 Robust AI-Generated Image Detection in the Wild.
- Combines EVA-CLIP backbone with LoRA adapters (rank 16) on MHSA and FFN blocks.
- Pairs clean and distorted images per batch, aligned via KL divergence and MSE feature loss.
- Reached 92.15% and 92.50% AUC on public and private hard test sets.
- Beats RINE on ForenSynths under JPEG Q=50 by roughly 9 accuracy points.
- Distortion simulation covers blur, noise, moire, tone curve, JPEG, and quantization.
Detecting AI-generated images looks easy on a clean benchmark and falls apart on the open internet. A team from TeleAI (China Telecom) and Xidian University tackled that gap with a strategy called LoRA-based Pairwise Training (LPT), which combines parameter-efficient fine-tuning of a vision foundation model with an aggressive distortion simulator and a two-headed loss that keeps clean-image accuracy from collapsing. It landed third place in the NTIRE 2026 Robust AI-Generated Image Detection in the Wild challenge.
Why detectors break outside the lab
Most AIGI detectors are trained and evaluated on clean, uncompressed images straight out of a generator. Once those images travel through social platforms, they pick up aggressive JPEG compression, arbitrary scaling, random cropping, blurring, and various noise, which wipes out the signals older detectors relied on. Classic frequency-domain methods, for example, lean on high-frequency spectral peaks that a single JPEG pass acts as a low-pass filter against, effectively erasing them.

The NTIRE benchmark makes this concrete. Training data contains 277,000 images from 20 generators with no distortions, while the private test set uses 35 generators and nine distortion types, and every image is hit with combinations of Gaussian blur, quantization, impulse noise, brightness shifts, and tone curve transformations at high strength. A detector trained on pristine data effectively meets a new distribution at test time.
The three-part recipe
LPT is built around three ideas working together: