SyntheticMDProductions/Roblox_Oasis_V4_AI_Model

Roblox Oasis V 2.3.7 V1

An experimental action-conditioned, rectified-flow video world model trained on recorded Roblox gameplay. Given a current visual state and supported input actions, it predicts the next simulated game frame.

This is a research/demo model, not a playable Roblox client. It does not connect to Roblox, its servers, or any Roblox accounts.

Game Previews

Model details

• Resolution: 256 x 144 (16:9)
• Architecture: action-conditioned rectified-flow video model
• Inference weights: unet/diffusion_pytorch_model.safetensors
• Supported demonstrated inputs: movement (W/A/S/D), jump, shift, 1, left/right mouse, and relative mouse movement
• Training result: 29 completed epochs; best recorded counterfactual accuracy: 85.4% on 48 validation samples

Files

• unet/: model architecture configuration and inference weights
• action_flow_model_info.json: input encoding, actions, and training metadata
• gui_previews/: training preview images

Use

Use this model with the companion trainer/player application: Oasis-Game-Trainer. Copy the downloaded folder into that application's model-output directory, then select it from the player interface.

The project uses Python, PyTorch, Diffusers, OpenCV, Pillow, MSS, and Pynput. See the companion repository for installation and runtime instructions.

Limitations

This model is a learned visual simulation. It may drift, become inconsistent over longer rollouts, or react imperfectly to inputs. It was trained on a limited action distribution and should be evaluated as an experimental project rather than a game replacement.

Attribution and rights

Unofficial research project. “Roblox” is used descriptively; this project is not affiliated with, endorsed by, or sponsored by Roblox Corporation. Only use or redistribute data and footage for which you have the necessary rights.