Official implementation & Dataset for the paper:
Is this Generated Person Existed in Real-world? Fine-grained Detecting and Calibrating Abnormal Human-body (CVPR 2025).
Zeqing Wang, Qingyang Ma, Wentao Wan, Haojie Li, Keze Wang, Yonghong Tian.
Training Dataset & Ckpt & Log & Baselines
COCO Human-Aware Val is an automated generation split from the COCO 2017 Val split, each image contains an abnormality with the corresponding bounding box.
AIGC Human-Aware 1K is a manually annotated split with 1000 iamges. Due to the uncertainty of anomaly location, we do not annotate the bounding box in the AIGC Human-Aware 1K. You can download them from 🤗Hugging Face.
Download from 'coco_train_replace_person_add_neg.tar.gz'
Following the LLaVa , GroundSamAnything and the SD2.
We provide all the results (with the middle results and the baselines) in GoogleDriver. In detail the results can be categoried in the following three types:
GoogleDriver/new_inpaint_without_SR.tar.gz
-
For the Closed-sorce VLMs (like GPT4o) GoogleDriver/aigc_baselines/close_source
-
For the Open-sorce VLMs (like LLaVa) GoogleDriver/aigc_baselines/open_source
-
For the CLIP like VLMs GoogleDriver/aigc_baselines/clip
The ckpt and the training log of our AHD, trained based on LLaVa1.5 7B, can be downloaded from Google Drive (llava-v1.5-7b-task-train_ver_with_neg_inpainting_1721922413_0392356_2_epoch.tar.gz). It can be inference with the LLaVa codebase.
If you find the repo useful for your work, please star this repo and consider citing:
@inproceedings{wang2025generated,
title={Is this generated person existed in real-world? fine-grained detecting and calibrating abnormal human-body},
author={Wang, Zeqing and Ma, Qingyang and Wan, Wentao and Li, Haojie and Wang, Keze and Tian, Yonghong},
booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
pages={21226--21237},
year={2025}
}