Args: base_folder (list of str): Path to the dataset root. num_past_obs (int): Number of past frames. num

Rotates each frame by its pixel shift and converts the equirectangular panorama to a perspective (cube-face) view. If w_per is None, it

Convert numpy array with possible shapes to (T,H,W,C). Accepts: 5D: (B,T,C,H,W) or (B,T,H,W,C) -> collapse to T' 4D: (T,C,H,W) o

Convert numpy array with possible shapes to (T,H,W,C). Accepts: 5D: (B,T,C,H,W) or (B,T,H,W,C) -> collapse to T' 4D: (T,C,H,W) o

Build generation defaults from: - Cosmos-native `args.overrides` (if provided) - worker-level `default_*` fields (fallbacks)

Best-effort lightweight type coercion; falls back to str.

Implementation closely follows k-diffusion. https://github.com/crowsonkb/k-diffusion/blob/6ab5146d4a5ef63901326489f31f1d8e7dd36b48/k_diffusio

Implementation closely follows k-diffusion. https://github.com/crowsonkb/k-diffusion/blob/6ab5146d4a5ef63901326489f31f1d8e7dd36b48/k_diffusio

Implementation closely follows k-diffusion. https://github.com/crowsonkb/k-diffusion/blob/6ab5146d4a5ef63901326489f31f1d8e7dd36b48/k_diffusio

Constructs the noise schedule of Karras et al. (2022).

Constructs the noise schedule of Karras et al. (2022).

Constructs the noise schedule of Karras et al. (2022).

Converts a non-Diffusers LoRA state dict to a Diffusers compatible state dict. Args: state_dict (`dict`): The state dict to convert.

Converts a text encoder LoRA key to a Diffusers compatible key.

Constructs the noise schedule of Karras et al. (2022).

Pick worker with fewest pending tasks and send (client_id, task_id, payload).

Fused `Conv2d()` followed by `downsample_2d()`. Padding is performed only once at the beginning, not between the operations. The fused op is c