---
tags:
- world-model
- image-generation
- game
- diffusers
library_name: diffusers
---

# Super Mario 64 Oasis Model

An action-conditioned visual world model for experimenting with Super Mario 64-style gameplay in the **Game Oasis Player**. Starting from a compatible gameplay screenshot, the model predicts the next visual frame from player input.

![Model preview](preview.png)

## Controls

The trained controls are:

- **W/A/S/D** — movement
- **Space** — jump

The checkpoint uses format version 4, a 256×144 frame size, and 39 input channels. It was trained through 65 completed epochs.

## Download and play

The easiest option is to open Game Oasis Player, select **Super Mario 64 Oasis Model**, and click **Download & Install Model**. The player downloads and validates this repository automatically.

For a manual installation, download the repository and select its folder using **Add Model from ZIP or Folder**. Keep this layout intact:

```text
action_flow_model_info.json
unet/
  config.json
  diffusion_pytorch_model.safetensors
```

Use a clear 16:9 gameplay screenshot from the same game and visual style as the starting image. A CUDA-capable NVIDIA GPU is strongly recommended.

## Limitations

This is an experimental, short-horizon visual world model. It generates the next view; it is not an emulator, game engine, ROM, or connection to a game server. It does not provide reliable game logic, persistent maps, inventory state, multiplayer, or deterministic physics. Visual drift and inconsistent geometry can accumulate over time.

This is an unofficial fan research project and is not affiliated with or endorsed by Nintendo.