MusapYildiz/aegis-video-detector

AEGIS — AI-Generated Video Detection Checkpoints

Trained weights for AEGIS, a multi-branch model that detects fully AI-generated videos (Sora, Veo, Kling, Runway, Pika, ModelScope, etc.) as opposed to real camera footage. Not a deepfake/face-swap detector — this targets end-to-end synthetic video generation.

Code, architecture details, dataset composition, and benchmark results: https://github.com/MusapYildiz/ai_video_detection_benchmark

Files

• checkpoint_best.pt — Phase 2, Epoch 7/20, Val AUC=0.9980, Val FPR=0.0218. Use this one.
• checkpoint_last.pt — Phase 2, Epoch 20/20 (end of training, not the best epoch).
• epochs/checkpoint_epoch0NN.pt — every epoch checkpoint from the Phase 2 training run (1 through 20), kept for ablation/epoch-selection analysis reproducibility.

Usage

import torch
from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download(
    repo_id="MusapYildiz/aegis-video-detector",
    filename="checkpoint_best.pt",
)
checkpoint = torch.load(ckpt_path, map_location="cpu")

See the GitHub repo's src/branches/inference.py for the full model-loading and inference pipeline (--checkpoint argument accepts this file directly).

License

MIT (matches the source code license). See the GitHub repo for details on which parts of the pipeline are third-party/adapted code under their own licenses.