---
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)