blanchon/cs2-world-model

CS2 World Model — action-conditioned latent DiT (OpenCS2)

Action-conditioned latent world model for OpenCS2: given a few recent frames and the player's actions, it predicts the next frame. It never touches pixels — history, target and prediction are all 32×8×8 latents produced by the frozen CS2 RAE tokenizer. Code: github.com/julien-blanchon/world-model.

• Dynamics core: ~205M-param in-context DiT (hidden=1024, depth=16, heads=16, register_tokens=4, always-on within-frame conv-stem).
• Transport: iMeanFlow with plain-Gaussian noise (no OT coupling), schedule-shift α=2, plain MSE. A frame samples in one network evaluation (NFE=1).
• Conditioning: history_frames=4, sequence_frame_gap=1, action-conditioned.
• Training: AdamW lr 2e-4, warmup 2000, EMA 0.9995, DIAMOND scheduled sampling with an nas depth curriculum (1 → 2 → 4) and a context-noise ramp (0.05 → 0.2).

Files

File	Description
world-model.pt	Final world-model checkpoint (runs/world_model/world-model.pt).

Usage

Requires the matching CS2 RAE tokenizer (d32) and the world-model codebase.

Internal research checkpoint. Provided as-is, no warranty.