| license: cc-by-4.0 | |
| tags: | |
| - text-to-video | |
| - membership-inference | |
| # MVDiT Panda-70M Reference | |
| The reference model of *Membership Inference for Generative Video Models* by Raina Panda and Maty Bohacek. Code: [github.com/maty-bohacek/t2v-mia](https://github.com/maty-bohacek/t2v-mia). Data: [matybohacek/t2v-mia](https://huggingface.co/datasets/matybohacek/t2v-mia). | |
| An MVDiT-XL/2 of identical architecture to [MVDiT-VideoUFO](https://huggingface.co/matybohacek/mvdit-videoufo), trained from scratch for 5,000 steps on 2,497 Panda-70M clips whose source videos do not appear in VideoUFO. It uses the training augmentations of the target but no EMA. It has never seen any VideoUFO clip, and so it serves as a control for the attacks (Table 3, Appendix C.5 and D.2). It is not intended as a generative model. Access is reviewed by the research team. | |
| ```shell | |
| huggingface-cli download matybohacek/mvdit-panda70m-reference ref_step5000.pt --local-dir ckpts | |
| ``` | |