---
base_model: THUDM/CogVideoX-2b
library_name: peft
tags:
- lora
- dora
- cogvideox
- physics
- video-generation
- warp
---

# PDW — Physics-Corrected CogVideoX-2b World Model (DoRA Adapter)

A **DoRA (Weight-Decomposed Low-Rank Adaptation)** adapter for [CogVideoX-2b](https://huggingface.co/THUDM/CogVideoX-2b), fine-tuned to generate physically accurate videos using **NVIDIA Warp** physics simulation data and **TRD (Temporal Representation Distillation)** with DINOv2-large as teacher.

## Model Details

- **Base model:** THUDM/CogVideoX-2b (1.7B params)
- **Adapter:** DoRA (r=16, lora_alpha=32, use_dora=True)
- **Target modules:** `to_q`, `to_k`, `to_v`, `to_out.0`
- **Trainable params:** 7.6M / 1.7B (0.45%)
- **Physics engine:** NVIDIA Warp (28-scenario 7×4 grid)
- **TRD teacher:** DINOv2-large
- **Hardware:** NVIDIA H100 NVL
- **Training steps:** 400

## Evaluation Results

| Metric | Delta |
|---|---|
| Diffusion MSE | +94.1% |
| Motion score | +1.7% |
| Overall | +47.9% |

## How to Use

```python
from peft import PeftModel
from diffusers import CogVideoXTransformer3DModel

# Load base transformer
base_transformer = CogVideoXTransformer3DModel.from_pretrained(
    "THUDM/CogVideoX-2b", subfolder="transformer"
)

# Load DoRA adapter
model = PeftModel.from_pretrained(base_transformer, "athul020/pdw_final_dora")
```

### Framework versions

- PEFT 0.18.1