Discretise per-frame translation into WASD key lists. Thresholds are adaptive: ``thresh = max(floor, frac * p95(|delta|))``.

Apply pixel unshuffle to output.

Autoencoder3d adapted from Stable Diffusion 1.x, 2.x and XL.

Generate and save sample videos to monitor training progress.

Override parent dispatcher to always use 10-slot cache logic.

Roll out a ``(N+1, 4, 4)`` camera-to-world trajectory from an action string. The DSL groups segments as ``<keys>-<frames>`` joined by commas; ``"

Diffusion forward corruption process. Input: - clean_latent: the clean latent with shape [B, C, H, W] - noise

Add timestep-dependent noise to the hard-conditioning latents. This helps with motion continuity, especially when conditioned on a si

Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of (1-beta) over time from t = [

Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of (1-beta) over time from t = [

Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of (1-beta) over time from t = [

Apply Gaussian blur to mask similar to diffusers pipeline Args: mask: PIL Image or numpy array mask blur_fac

Calculation of the statistics used by the FID. Params: -- files : List of image files paths -- model : Instance of inception m

Calculates the FID of two paths

Calculates the FID of two paths

Numpy implementation of the Frechet Distance. The Frechet distance between two multivariate Gaussians X_1 ~ N(mu_1, C_1) and X_2 ~ N(mu_2, C_2

Check if the cache miss rate is too high.

Returns binned height and width.

Returns binned height and width.

Returns binned height and width.

Remove old checkpoints if the number exceeds max_checkpoints. Only the main process performs the actual deletion to avoid race conditions

Clear decoder cache (legacy API, delegates to cache manager).

Compute critic loss. Args: chunk: generated chunk chunk_info: chunk metadata Returns: l

Compute the md5sum of a file in chunks. Parameters ---------- fname : Union[str, BinaryIO] Filename or file object chunksize

Compute the generator loss. Args: chunk: generated chunk chunk_info: chunk metadata Returns:

Computes a geometry raymap (directions/moments or origins/directions). Args: intrinsics: (T, 4) tensor [fx, fy, cx, cy] pose

Compute UCPE absolute embedding maps ``(up_map, lat_map)``. ``up_map`` is a 2-channel projected up-direction; ``lat_map`` is a 1-channel lati

Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).

Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).

Compute the mean for the previous step, given a function cond_fn that computes the gradient of a conditional log probability with res

Map a set of canonical control tokens to a target velocity.

Map original LTX-2 connector keys to diffusers LTX2TextConnectors keys.

Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is designed to discretize an

Build chunk boundaries for an autoregressive sweep. Returns a list of frame indices ``[0, c1, c2, ..., total_frames]`` of length ``nu