Mundus

Mundus is a self-contained, provider-based hub for video/image generators, learned world models, representation models, and WAMs. Its SANA integration is bundled in mundus.sana; it does not import or require the sibling sana_sim project at runtime. Providers have one common generation interface; heavyweight models are loaded only when used.

The core has no runtime dependencies. For the bundled SANA provider, install the optional runtime once with pip install -e '.[sana]'.

Standalone SANA video generation

from mundus import Mundus

world = Mundus.from_pretrained(
    "sana-video",
    model_path="checkpoints/SANA_Video_2B_480p.pth",
    vae_path="checkpoints/Wan2.1_VAE.pth",
)
result = world.generate("A robot carefully opens a drawer", num_frames=81, seed=42)
result.save("robot.mp4")

SANA MoT WAM

migma.models.wms.samot.SanaMoTWAM keeps the loaded SANA denoiser as the video branch and adds an action diffusion head. The action head reads SANA's per-layer normalized video tokens (world memory) and its own action tokens; the video branch never reads action tokens. This is the asymmetric Mixture-of-Transformers contract and preserves the normal SANA video output.

from migma.models.wms.samot import SanaMoTWAM

provider = world.model
wam = SanaMoTWAM(provider.backbone, action_dim=7, max_action_horizon=32).cuda()
text, mask = provider.encode_text("Robot reaches for the drawer handle")
video_noise, action_noise = wam(
    noisy_video, video_timestep, text, mask, noisy_actions,
)

Train the two outputs with the appropriate SANA video denoising loss and the action flow/noise loss. SANA can be frozen initially, then selectively unfrozen for joint WAM finetuning.

Extending the hub

Implement the WorldModel protocol (model_id, capabilities, and generate(request)), then register a factory:

from mundus import registry
registry.register("my-world-model", MyWorldModel)

No provider has to share model internals with another provider. A provider that exposes a backbone may additionally be used by a WAM adapter.