mstar.model.lingbot_world.config

Configuration for LingBot-World-V2 causal-fast (camera-controlled world model).

The architecture values are facts of the robbyant/lingbot-world-v2-*-causal-fast checkpoints (the reference repo’s wan/configs/wan_i2v_*.py), hardcoded so that constructing the model never touches the network. The DiT is a Wan2.1-layout I2V transformer (16-channel Wan2.1 VAE latents + 4 mask + 16 image-latent channels in) with per-block Plucker camera injection, run block-causally over chunks of latent frames against a rolling self-attention KV cache with attention-sink frames.

The sampling knobs are the reference’s generate.py / run_fast.sh defaults: 4 distilled steps per chunk at fixed indices into a 1000-step shifted schedule, re-noising between steps, then one t=0 pass over the clean chunk to commit its K/V.

Functions

lingbot_world_1_3b()

Classes

LingBotWorldConfig([variant, dim, ffn_dim, ...])

class mstar.model.lingbot_world.config.LingBotWorldConfig(variant: str = '1.3b_causal_fast', dim: int = 1536, ffn_dim: int = 8960, num_heads: int = 12, num_layers: int = 30, freq_dim: int = 256, text_dim: int = 4096, in_dim: int = 36, out_dim: int = 16, patch_size: tuple[int, int, int] = (1, 2, 2), eps: float = 1e-06, text_len: int = 512, control_dim: int = 6, rope_max_frames: int = 1024, vae_z_dim: int = 16, vae_stride: tuple[int, int, int] = (4, 8, 8), chunk_size: int = 4, local_attn_size: int = 18, sink_size: int = 6, timesteps_index: tuple[int, ...] = (0, 250, 500, 750), num_train_timesteps: int = 1000, sample_shift: float = 5.0, default_height: int = 480, default_width: int = 832, max_area: int = 921600, intrinsics_height: int = 480, intrinsics_width: int = 832, video_fps: int = 16, max_chunks: int = 255, vae_decode_fp32: bool = True, vae_decode_channels_last: bool = False)

Bases: object

Parameters:

• variant (str)

• dim (int)

• ffn_dim (int)

• num_heads (int)

• num_layers (int)

• freq_dim (int)

• text_dim (int)

• in_dim (int)

• out_dim (int)

• patch_size (tuple[int, int, int])

• eps (float)

• text_len (int)

• control_dim (int)

• rope_max_frames (int)

• vae_z_dim (int)

• vae_stride (tuple[int, int, int])

• chunk_size (int)

• local_attn_size (int)

• sink_size (int)

• timesteps_index (tuple[int, ...])

• num_train_timesteps (int)

• sample_shift (float)

• default_height (int)

• default_width (int)

• max_area (int)

• intrinsics_height (int)

• intrinsics_width (int)

• video_fps (int)

• max_chunks (int)

• vae_decode_fp32 (bool)

• vae_decode_channels_last (bool)

chunk_size: int = 4

control_dim: int = 6

default_height: int = 480

default_width: int = 832

dim: int = 1536

eps: float = 1e-06

ffn_dim: int = 8960

freq_dim: int = 256

property head_dim: int

in_dim: int = 36

intrinsics_height: int = 480

intrinsics_width: int = 832

local_attn_size: int = 18

max_area: int = 921600

max_chunks: int = 255

num_heads: int = 12

num_layers: int = 30

num_train_timesteps: int = 1000

out_dim: int = 16

patch_size: tuple[int, int, int] = (1, 2, 2)

rope_max_frames: int = 1024

sample_shift: float = 5.0

sink_size: int = 6

text_dim: int = 4096

text_len: int = 512

timesteps_index: tuple[int, ...] = (0, 250, 500, 750)

vae_decode_channels_last: bool = False

vae_decode_fp32: bool = True

vae_stride: tuple[int, int, int] = (4, 8, 8)

vae_z_dim: int = 16

variant: str = '1.3b_causal_fast'

video_fps: int = 16

mstar.model.lingbot_world.config.lingbot_world_1_3b()

Return type:

LingBotWorldConfig