🌍 Worldlet: DreamDoom Video Generative Model
This repository contains the dataset and trained weights for DreamDoom, an experimental prototype exploring Generative Video Models.
Imagine a system that doesn't just play a game, but understands how to generate the visual output of the game frame-by-frame entirely via a neural network. This project provides a sequence-to-sequence video prediction model trained on ViZDoom, allowing you to "play" the hallucinated game interactively.
📁 Repository Contents
data/: Contains the transition datasets collected from random agent gameplay across multiple ViZDoom scenarios.
transitions_doom_diverse.npz: Latest diverse dataset (1.1GB).
models/: Contains the trained PyTorch weights (.safetensors):
vq_vae_doom.safetensors: The VQ-VAE tokenizer.transformer_world_doom.safetensors: The Autoregressive Transformer predictor.
✨ Model Architecture
The repository provides weights for three different generative backend architectures:
- DiT (Diffusion Transformer) [Default & Best Quality]: A continuous diffusion-based world model. It takes continuous latents from the VQ-VAE and learns to iteratively denoise the next frame's latents.
- Autoregressive Transformer (Discrete): A sequence-to-sequence discrete token model. It receives 64 discrete visual tokens and an action token, autoregressively predicting the next 64 tokens.
- VideoPredict CNN (Direct Pixel): A simpler direct pixel-to-pixel generative CNN baseline.
All models utilize the underlying VQ-VAE which encodes 64x64 visual frames into an 8x8 grid of latent representations.
🎮 How to Use
You can use these files directly with the
Worldlet/DreamDoom GitHub repository.
Download the model and run the interactive hallucination:
Controls in DreamDoom
- W / S or Up / Down Arrow: Move Forward / Backward
- A / D or Left / Right Arrow: Move Left / Right
- Spacebar: Attack
This model was trained as a concept demo for exploring interactive neural-rendered environments.
