mstar.model.wan22.config

Configuration for Wan2.2-TI2V-5B (dense video DiT + UMT5-XXL + Wan2.2-VAE).

The architecture values are facts of the Wan-AI/Wan2.2-TI2V-5B-Diffusers checkpoint, hardcoded so that constructing the model never touches the network. Wan22Model._refresh_checkpoint_defaults re-reads flow_shift from the checkpoint at load time and hard-fails if the solver order or train-timestep count drift from what the inline UniPC port implements.

Classes

Wan22Config([variant, num_attention_heads, ...])

Wan2.2-TI2V-5B model configuration.

class mstar.model.wan22.config.Wan22Config(variant='ti2v_5b', num_attention_heads=24, attention_head_dim=128, num_layers=30, ffn_dim=14336, in_channels=48, out_channels=48, patch_size=(1, 2, 2), text_dim=4096, freq_dim=256, rope_max_seq_len=1024, qk_norm='rms_norm_across_heads', cross_attn_norm=True, eps=1e-06, text_max_seq_len=512, vae_z_dim=48, vae_scale_factor_spatial=16, vae_scale_factor_temporal=4, flow_shift=5.0, guidance_scale=5.0, default_num_inference_steps=50, default_height=704, default_width=1280, default_num_frames=81, default_negative_prompt='', video_fps=24, max_denoise_steps=100, compile_dit=True, vae_decode_tiling='auto')

Bases: object

Wan2.2-TI2V-5B model configuration.

A single dense video DiT driven by UMT5-XXL embeddings, denoising Wan2.2-VAE latents with UniPC. T2V and I2V share the one transformer: I2V injects the encoded first frame into the latent grid and zeroes its per-token timestep, rather than concatenating on the channel axis.

Parameters:

• variant (str)

• num_attention_heads (int)

• attention_head_dim (int)

• num_layers (int)

• ffn_dim (int)

• in_channels (int)

• out_channels (int)

• patch_size (tuple[int, int, int])

• text_dim (int)

• freq_dim (int)

• rope_max_seq_len (int)

• qk_norm (str)

• cross_attn_norm (bool)

• eps (float)

• text_max_seq_len (int)

• vae_z_dim (int)

• vae_scale_factor_spatial (int)

• vae_scale_factor_temporal (int)

• flow_shift (float)

• guidance_scale (float)

• default_num_inference_steps (int)

• default_height (int)

• default_width (int)

• default_num_frames (int)

• default_negative_prompt (str)

• video_fps (int)

• max_denoise_steps (int)

• compile_dit (bool)

• vae_decode_tiling (str)

attention_head_dim: int = 128

compile_dit: bool = True

cross_attn_norm: bool = True

default_height: int = 704

default_negative_prompt: str = ''

default_num_frames: int = 81

default_num_inference_steps: int = 50

default_width: int = 1280

eps: float = 1e-06

ffn_dim: int = 14336

flow_shift: float = 5.0

freq_dim: int = 256

guidance_scale: float = 5.0

property hidden_size: int

DiT hidden width (num_attention_heads * attention_head_dim).

in_channels: int = 48

max_denoise_steps: int = 100

num_attention_heads: int = 24

num_layers: int = 30

out_channels: int = 48

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

qk_norm: str = 'rms_norm_across_heads'

rope_max_seq_len: int = 1024

property spatial_alignment: tuple[int, int]

(height, width) pixel multiples a request’s size must satisfy — 32 each.

The VAE downsamples a pixel dimension by 16 and the DiT then patchifies by 2, so only exact multiples of the product survive both. An unaligned size has no clean failure inside the model, so it is rejected at the request seam (Wan22Model.process_prompt).

text_dim: int = 4096

text_max_seq_len: int = 512

vae_decode_tiling: str = 'auto'

vae_scale_factor_spatial: int = 16

vae_scale_factor_temporal: int = 4

vae_z_dim: int = 48

variant: str = 'ti2v_5b'

video_fps: int = 24