diamond_csgo(pretrained: bool | str = False, weights: DIAMONDWeights | None = None, overrides: object = {})Construct the Counter-Strike world model.
Model Size
Parameters
pretrainedbool or str= FalseLoad the released CS:GO world model.
True and "CSGO"
both select it; there is only the one.An explicit tag, taking precedence over
pretrained.**overridesobject= {}Optional
DIAMONDConfig field overrides.Returns
DIAMONDModelThe world model and its upsampler. reward_end and
actor_critic are None: this experiment has no
reinforcement learning, so there is no agent to build.
Notes
Reference: Alonso et al., arXiv:2405.12399, Section 6 and Appendix M,
with the architecture taken from the released
config/agent/csgo.yaml — the paper reports parameter counts for
this model but no channel widths or depths.
Frames are 30x56 here and 150x280 after the upsampler, which
is what makes a 3D scene affordable: diffusing at the full resolution
would cost far more for detail a cheaper second network can add.
Examples
>>> from lucid.models import diamond_csgo
>>> config = diamond_csgo().config
>>> config.frame_shape, config.num_actions, config.upsampling_factor
((30, 56), 51, 5)
>>> config.with_agent
False