Autoregressive video world model for Minecraft. A frozen Stable Diffusion VAE compresses 256×256 frames to 4×32×32 latents; a decoder-only transformer (338M params, RoPE, DWS-style per-block action adapters, SDPA flash attention) predicts the next latent conditioned on discrete action tokens (29 ids: 22 keyboard + 6 camera bins + 1 no-op). Each run directory holds a full resumable state (model + optimizer + scheduler + step/epoch counters), so any file below restarts training bit-identically with the matching config and --resume: Weights derive from public TESS-VLA gameplay data and an sd-vae-ft-mse base. No license is asserted here — check the dataset and base-model terms before commercial…
Independent publisher
Keylhan Paumard--André
keypa
Efficient deep learning, LLM fine-tuning, inference optimization, model compression, distributed training, GPU systems, open-source AI infrastructure
Models in Library1
Datasets in Library0
Models on Hugging Face17
Followers10