Deepfake generalization: demo runtime

This repo holds the files the public demo needs to run. It is not a standalone model release. The code lives at github.com/vaibhavshiroorkar/deepfake-detection, and the demo runs in the browser at huggingface.co/spaces/vaibhavshiroorkar/deepfake-generalization.

Contents

  • runs/program-20260906/checkpoints/: six visual detectors, three EfficientNet and three ConvNeXt seeds, with their training histories.
  • runs/probe-20260914/authenticity-density.npz: the familiarity model. The app withholds a verdict when a clip looks unlike the training data.
  • runs/program-20260906/evaluation/: metrics and per-clip predictions for each test set.
  • dist/: the final report as HTML and PDF.
  • configs/webapp.yaml: demo settings. Only the visual ensemble and the best single model are enabled, and uploads are capped at 30 MB.

No datasets are included.

Browser models

browser/ holds ONNX exports for the in-browser demo, made with hosting/browser/export.py in the code repo:

  • visual-best.onnx: the seed 17 EfficientNet-B0 checkpoint above, as one graph from 16 face crops to a logit.
  • pnet.onnx, rnet.onnx, onet.onnx: the MTCNN face detector from facenet-pytorch (MIT).
  • dinov3.onnx: Meta's DINOv3 ViT-S/16, the encoder the familiarity model was fitted on. It is distributed under the DINOv3 License; a copy is in browser/DINOV3_LICENSE.md.
  • gate.json: the familiarity model as plain numbers.

Read the results first

The project studies why deepfake detectors fail on data they were not trained on. The visual ensemble scores ROC AUC 0.999 on its in-domain test set, 0.774 on DFDC and 0.668 on Celeb-DF v2. On Celeb-DF it flags 113 of 155 real videos as fake at the default threshold. Expect the demo to withhold or get wrong most clips from outside its training data.

Do not use these models to judge whether a real video is genuine.

License

The project's own checkpoints are all rights reserved. Ask before reusing them. The DINOv3 and MTCNN files keep their original licenses, listed above.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Space using vaibhavshiroorkar/deepfake-detection 1