| license: other | |
| pipeline_tag: text-to-video | |
| # ViRDM models and training assets | |
| This repository contains the trained causal four-step ViRDM generators and the | |
| compact frozen artifacts required to reproduce training. | |
| - `checkpoints/virdm_causal4_dynamic_step20.pt`: the primary | |
| 20-update model with dynamics regularization (`5e-4`). | |
| - `checkpoints/virdm_causal4_nodynamic_step20.pt`: the matched model | |
| without dynamics regularization. | |
| - `prompt_data/data.mdb`: 6,505 captions in the original ordered | |
| `zhuhz22/Causal-Forcing-data` training split; no video or latent payloads. | |
| - `references/reference_M4096.pt`: the fixed 4,096-landmark joint video-text | |
| reference used by ViRDM. Its publication metadata contains stable upstream | |
| identifiers rather than machine-local paths. | |
| - `references/siglip2_text_fp32.npy`: frozen SigLIP2 features aligned row for | |
| row with the reference and prompt table. | |
| - `asset_receipt.json`: row contract, filenames, and SHA-256 checksums. | |
| 📄 [Paper: ViRDM](https://arxiv.org/abs/2609.28923) | |
| Project page: [https://neu-vi.github.io/ViRDM/](https://neu-vi.github.io/ViRDM/) | |
| Code: [https://github.com/neu-vi/ViRDM](https://github.com/neu-vi/ViRDM) |