Note: This repository is the code for my final project for UIUC CS 445 Fall 2024.
Recent generative models like Sora have shown promise in producing high quality images and videos given text input, but they can also hallucinate inconsistent, undesirable, and irrelevant features. Existing video hallucination detectors, like Sora Detector, leverage another language model for image captioning, which itself can hallucinate. Recent works have shown promising results in exploiting extracted geometries from static images to detect hallucinations, but have not been applied to video sequences. In this work, we first present a short case study showcasing hallucinated predictions output by language models used for anomaly detection, motivating the need for structure-guided hallucination detectors. We then utilize state-of-the-art models to extract geometric cues from a custom-curated dataset of AI-generated and real videos, and directly apply a recent geometry-based hallucination detector on each static frame. Finally, we extend the static frame hallucination detector to consider the temporal consistency across consecutive frames by developing a Kalman filter that fuses predictions from different models.
Look inside each sub-folder to see how to run that part of the project.
./
│ README.md - This file
│ frame_extraction.py - Saves raw frames from list of videos
| keyframe_selector.py - Extracts key frames from videos
| projective_geometry.yml - Conda environment file
│
└───baseline - Folder with code for pixel-based baseline classifier
│
└───data - Folder where extracted data is saved
│
└───kalman_filter - Folder with code for running Kalman filter
│
└───line_segment - Folder with code for predicting lines, and running line-based classifier
│
└───object_shadow - Folder with code for predicting object and shadow masks, and running object-shadow-based classifier
│
└───perspective_fields - Folder with code for predicting perspective fields, and running perspective fields-based classifier
│
└───results_extractor - Folder with code for generating test splits, and compiling result graphs
│
└───sora_detector - Folder with prompts used for Sora Detector, and output predictions from GPT-4o using Perplexity
We extract 59 AI-generated videos from the T2VHaluBench dataset and 59 real videos from the BDD100K dataset.
- All 59
T2VHaluBenchvideos can be found at https://github.com/TruthAI-Lab/SoraDetector/tree/main/text-to-video%20benchmark/text-to-video - We downloaded
bdd100k_videos_test_00.zipfound underBDD100K→Video Partsat http://bdd-data.berkeley.edu/download.html. Then we chose 59 videos from that set, which are listed in./data/real_video_paths.txt.
We use conda to manage packages for running different parts of the project. Unless otherwise stated, we run our code with the projective-geometry environment. Install the environment with conda env create -f projective_geometry.yml.
- Create and activate the
projective-geometryconda environment - Create
./data/fake_video_paths.txtand./data/real_video_paths.txtwith lists of paths to videos to save frames from - Run
python frame_extraction.pyto save raw frames - Run
python keyframe_selector.pyto extract key frames from raw frames - Follow directions in
./object_shadowto predict masks and run object-shadow-based classifier - Follow directions in
./perspective_fieldsto predict fields and run fields-based classifier - Follow directions in
./line_segmentto predict lines and run lines-based classifier - Follow directions in
./kalman_filterto run Kalman filter - Follow directions in
./baselineto run baseline classifier - Follow directions in
./results_extractorto extract difficulty-based test splits, and compile results - Optional: Prompt GPT-4o with Sora Detector prompts provided in
./sora_detector/prompts.txtand/or look at our case study raw results in./sora_detector/results
We thank the authors of the following projects for making their data/code/models available: