---
license: apache-2.0
library_name: pytorch
pipeline_tag: other
tags:
  - world-model
  - diffusion
  - pong
  - video
  - causal
---

# Diffusion Pong Base

Unconditional Pong world model.

![preview](preview.png)

128x128 at 6 FPS with a 12-frame latent history. Frozen SDXL VAE (`madebyollin/sdxl-vae-fp16-fix`). Needs a CUDA GPU with BF16. ~59M DiT params.

## Run it

```bash
pip install torch numpy pillow safetensors huggingface_hub diffusers

hf download kerzgrr/diffusionpong-base live_infer.py --local-dir .
python live_infer.py --steps 2
```

Or download weights locally:

```bash
hf download kerzgrr/diffusionpong-base --local-dir checkpoints/diffusionpong-base --include "ema.safetensors" --include "config.json" --include "live_infer.py"
python checkpoints/diffusionpong-base/live_infer.py --local-dir checkpoints/diffusionpong-base --steps 2 --window-scale 7 --seed 42
```

No paddle controls. Click the canvas to drop a yellow ball into history for a few frames if you want to poke the dynamics.

For the playable W/S version, see [`kerzgrr/diffusionpong`](https://huggingface.co/kerzgrr/diffusionpong).

## Files

| file | what |
|---|---|
| `ema.safetensors` | weights |
| `config.json` | train-time settings |
| `live_infer.py` | live rollout script |
| `preview.png` | validation strip from step 1050 |