Game World Model Trained on Self-Play Data
gamajuice1 · reddit · 2026-07-18
The author trained an interactive diffusion world model based on Hollow Knight from scratch, using data from their own recorded gameplay and input controls.
Training Setup
- About 400,000 frames of data, around 7 hours of gameplay
- Trained from random initialization, without using pre-trained weights or LoRA
- Data comes solely from the author's own game recordings and operation logs
Current Capabilities
- Has begun learning some real mechanics
- Can respond to inputs like movement, dash, attack, hit
- The generated clips are not real-time game renders or simple replays, but frames synthesized by the model based on control signals
Current Issues and Plans
- Still makes errors during long runs
- The author will continue to improve consistency and add more diverse game data
- The goal is to create a playable demo that runs in a browser
More from Research
- KANLab: an interactive browser playground for Kolmogorov-Arnold Networks — Lukas Niehaus · 2026-10-27
- CMU's LLM2Cert synthesizes stability certificates verified via SMT and Lean — CarnegieMellonU · 2026-10-27
- NVIDIA open-sources FLORA, a robotics data engine ~125x cheaper than GPT-6 Astra — nvidia · 2026-10-27
- RSNA 2026 Deep Learning Labs return: eight hands-on AI classes over four days in Chicago — PRouzrokh · 2026-10-27
- Steering two orthogonal concept neurons makes a model compose "gnome" — Sauers_ · 2026-10-27
- Paper: data schedules and learning rate schedules should be designed together — alifcoder · 2026-10-27