Fused RMSNorm + ReLU + (K-scale on K) + UCPE 4x4 + RoPE for the cam branch. Args: q_raw, k_raw, v_raw: ``(B, N, H, D)`` contiguous (any f

Bidirectional camera scan using shared chunkwise phases. This is equivalent to ``cam_scan_chunkwise(..., reverse=False) + cam_scan_chunkwise(

Check whether more chunks can be generated

Check that a list of shards is well-formed. This checks that the list is a list of dictionaries, and that each dictionary has a "url" and a "

Convert chunk_size to chunk_index list with a split strategy. Args: T: Number of latent frames. chunk_size: Base chunk size for t

Interpolate large tensors by chunking along the channel dimension. https://discuss.pytorch.org/t/error-using-f-interpolate-for-large-3d-input/207

Returns binned height and width.

Compute actual chunk sizes from chunk_index. Args: chunk_index: List of chunk start indices (e.g., [0, 4, 8, 12]). T: Total numbe

compute DMD loss (consistent with DMD class).

Compute what the p_mean_variance output would have been, should the model's score function be conditioned by cond_fn. See con

Convert the diffusion network's x0 prediction to noise predidction. x0: the predicted clean data with shape [B, C, H, W] xt:

Create a pixel coordinate grid of shape ``(H, W, 3)`` or ``(B, H, W, 3)``.

Return the data prediction model (with corrector).

Return the data prediction model (with corrector).

Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.

Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.

Downloads a pre-trained Sana model from the web.

Downloads a pre-trained Sana model from the web.

Estimate ``(fx, fy, cx, cy)`` for ``image`` using Pi3X. The image is internally resized to a Pi3X-friendly shape; the returned intrinsics are

Forward pass of Sana. x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images) t: (N,) tensor of

Forward pass of Sana. x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images) t: (N,) tensor of

Forward pass of Sana. x: (N, C, T, H, W) tensor of spatial inputs (images or latent representations of images) t: (N,) tensor

Forward pass for self-forcing / chunk-causal AR KV-cache inference. Mirrors :meth:`forward` exactly with two streaming-specific differences:

Forward pass of Sana. x: (N, C, T, H, W) tensor of spatial inputs (images or latent representations of images) t: (N,) tensor

Single-pass Triton fused Q+K inverse-RMS. Replaces ``(_precompute_inv_rms(qkv, 0, C, eps), _precompute_inv_rms(qkv, 1, C, eps))`` with one la

Forward/backward pass using the new StreamingTrainingModel for serialized training

Generate a video. Args: image: First-frame RGB image, already cropped to ``(704, 1280)``. prompt: Text prompt.

same as DMD: unfold generator to get fake video and compute DMD loss.

embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D)

This is the deprecated API for creating beta schedules. See get_named_beta_schedule() for the new library of schedules.

Resolve chunk_index from a config, supporting chunk_size and strategy. Priority: 1) config.model.chunk_index (explicit list) 2) confi

Generates embeddings and an attention mask for the given prompt(s) and image(s), stripping the template part. Args:

Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time. For discrete-time DPMs, we convert `t_continuou

Update the current scheduler with the interface's static method

Generate scribble map from input image Args: input_image: Input image (numpy array, HWC format) det: Detector type ('Scribbl

Return the number of cache accesses and misses.

Compute the intermediate time steps for sampling.

Generate a full video. Args: noise: Gaussian noise latent of shape [B, T, C, H, W] or [B, C, T, H, W]. text_

A piecewise linear function y = f(x), using xp and yp as keypoints. We implement f(x) in a differentiable way (i.e. applicable for autograd).

A piecewise linear function y = f(x), using xp and yp as keypoints. We implement f(x) in a differentiable way (i.e. applicable for autograd).

Interpret the transformations argument. This takes care of transformations specified as string shortcuts and returns a list of callables.

Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.

Iterate over the ranges in a random order.