Return ``(num_frames, 4)`` intrinsics as ``[fx, fy, cx, cy]``. Accepts ``.npy`` arrays shaped ``(3, 3)``, ``(F, 3, 3)``, ``(4,)``, or ``(F, 4

Load a remote or local dataset description in JSON format.

Compute alpha_t of a given continuous-time label t in [0, T].

Compute sigma_t of a given continuous-time label t in [0, T].

Return the noise prediction model.

Return the noise prediction model.

Compute the KL divergence between two gaussians. Shapes are automatically broadcasted, so batches can be compared to scalars, among other

Input: v: (B, N, H, C) k: (B, N, H, C) Output: vk: (B, H, C+1, C)

Capture the source-sink K/V before the first active refiner block. The sink is just the conditioning latent frame and does not depend on

Relativise poses to frame 0 and build the model-input tensors.

Complex rotary_emb `(1, 1, N, D//2)` → expanded (N, D) cos/sin tables. Encodes the interleaved-pair rotation y[2i] = x[2i]*cos[i] - x[2

Rebase the URLs in a shardlist.

Clear the persistent decoder cache; the next call will be the first chunk.

Reset streaming training state

reset training state

Aspect-preserving resize then center-crop to ``(target_h, target_w)``. Returns ``(cropped_image, src_size, resized_size, crop_offset)`` where

Streaming variant of :meth:`sample` — yields one chunk at a time. After each AR chunk completes (denoising + KV-cache save pass), yields

Calculates the FID of two paths

save git snapshot to the git repository in work_dir use exp/<job_name>_<timestamp> as branch name Args: work_dir: work directory

Helper function for reproducible behavior to set the seed in `random`, `numpy`, `torch`. Args: seed (`int`): The seed to

Sets the discrete timesteps used for the diffusion chain (to be run before inference). Args: num_inference_steps (`int`):

Singlestep solver DPM-Solver-3 from time `s` to time `t`. Args: x: A pytorch tensor. The initial value at time `s`.

Returns the basename and extension of a filename

Take the sum over all non-batch dimensions.

r"""Decode a batch of images using a tiled decoder. Args: z: Input batch of latent vectors [B, C, T, H, W]. temb: Opt

r"""Encode a batch of images using a tiled encoder. Args: x: Input batch of videos [B, C, T, H, W]. causal: Whether t

Create sinusoidal timestep embeddings. :param t: a 1-D Tensor of N indices, one per batch element. These ma

Adjust ``[fx, fy, cx, cy]`` to match a resize-then-center-crop image.

r""" Reconstruct video tensors from patch embeddings. Args: x (List[Tensor]): List of patchified features

x: (N, T, patch_size**2 * C) imgs: (N, H, W, C)

x: (N, T, patch_size**2 * C) imgs: (N, H, W, C)

Validate and update cache mode. Args: causal: Desired causal mode Raises: ValueError: If mode conflicts with

a: (B, H, C+1, D) # C=D b: (B, N, H, D)

Build per-pixel ``ray<-world`` transforms from camera unit rays + C2W poses.

Append a chunk of frames. Args: frames_uint8: ``(T, H, W, 3)`` uint8 RGB tensor with ``H == height`` and ``W == w

Construct a DPM-Solver. We support both DPM-Solver (`algorithm_type="dpmsolver"`) and DPM-Solver++ (`algorithm_type="dpmsolver++"`).