Seeking more community models to add to the detection pipeline!
The core strength of this novel way of deepfake detection tool stems from the variety of models and the different & ever-changing datasets that open-source models rely on.
Please, feel free to comment below with detection models that you would like to see integrated and a custom pipeline will be implemented, as long as the model meets some very simple criteron:
- The model must be open-source and its license must allow for academic and research usage
- The model's architecture must be based on a Huggingface supported library (ViT, Swin, etc.). If unsupported, assistance in integration would go a long way, as we are busy with other tasks at the moment.
- Source of origin and/or training dataset would be highly preferred, so proper credit and citations can be given to original authors.
The current models are quantized ONNX models by models found on the Hub. Special thanks to @haywoodsloan @Heem2 @Organika @cmckinle @prithivMLmods @ideepankarsharma2003 @date3k2 for the original contributions as open-source models, which were the basis of the inaugural ensemble team of detection models. And of course, the Community Forensics team.
To the community members listed above, please feel free to check app_optimized.py and app.py (depreciating soon) to ensure that your model was implemented correctly. Also, we're working on making proper model citations and credits more visible, currently trying to figure out the best way UI-wise. In the meantime, if you have a newer trained model, please link below and we can update the model as well.
Thank you 💪 Stronger together!