codelion/neural-drive-model · Hugging Face

Log In<br>Sign Up

Neural Drive — SuperTuxKart World Model

An action-conditioned world model for SuperTuxKart : it predicts the next frame from the last<br>7 frames plus your controls, with no game engine underneath . Trained in PyTorch on gameplay<br>(actions extracted automatically), and shipped in two runtimes from the same weights.

▶ Live in-browser demo (WebGPU): https://huggingface.co/spaces/codelion/neural-drive

Resolution 384×192 · Action space 6-dim [accelerate, brake, left, right, nitro, drift]

Codec ConvVAE (z_ch 8, /8 spatial) · Dynamics LatentDiT (dim 768 / depth 12 / patch 3, 130.8M ), rectified flow + per-frame diffusion forcing

Sampling few-step distilled (k = 2 )

Runtimes

runtime<br>where<br>notes

PyTorch<br>CUDA / CPU<br>training + reference (fp32)

ONNX / WebGPU<br>any browser<br>fp16, runs on the visitor's GPU

Files

PyTorch checkpoints (fp32)

file<br>what

dit.pt<br>dynamics, 100k steps (k=8 teacher)

dit_distilled.pt<br>few-step distilled student (k=2)

vae.pt<br>ConvVAE codec

decoder_big.pt<br>larger LPIPS decoder

config.json<br>architecture config

Web / ONNX (fp16 — the in-browser build)

file<br>what

web/genframe_fp16.onnx<br>single-graph frame generator: prefills the 7-frame context once, then runs the k=2 denoise steps internally (KV-cache in the graph). ~262 MB

web/vae.onnx<br>VAE decoder (fp32, to keep decode stable)

web/seeds.bin<br>18 spawn context states, raw float32 [18,7,8,24,48]

web/meta.json<br>shapes + config for the web runtime

Reference

Playing SuperTuxKart in Your Head: Learning Interactive World Models from Mobile Gameplay Videos<br>— https://x.com/latent_node/status/2033769450592485543

Downloads last month -

Inference Providers NEW<br>This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Space using codelion/neural-drive-model 1