File size: 1,200 Bytes
ed91eab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
75dbf0d
ed91eab
 
 
 
 
75dbf0d
ed91eab
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
---
license: mit
tags:
- diffusion
- world-models
- video-generation
- carracing
- reinforcement-learning
- pytorch
---

# Interactive Diffusion: A Playable World Model

This repository hosts the trained model weights and sample data for the project [**Demystifying World Models: A Playable Diffusion Model**](https://github.com/nlml/interactive-diffusion).

- **Blog Post**: [A Playable Diffusion Model](https://nlml.github.io/generative-models/interactive-diffusion-1/)
- **GitHub Repository**: [nlml/interactive-diffusion](https://github.com/nlml/interactive-diffusion)

## Contents

- `model.pt`: 145 MB checkpoint (optimizer stripped) of the `CausalSpatiotemporalDiT` trained with context noise augmentation (50,000 steps).
- `sample_data/`: Sample shard 0 (`shard_00000_*`) containing gameplay frames, actions, and valid episode masks collected from `CarRacing-v3`.

## Quickstart

Clone the GitHub repository and run the live driving server:

```bash
git clone https://github.com/nlml/interactive-diffusion.git
cd interactive-diffusion
pip install -e .

# Run the live interactive driving server:
python webapp/server.py --checkpoint nlml/interactive-diffusion --data_dir ./carracing_data
```