Causal 3d convolusion.

Causal 3d convolusion.

A sampler that samples in chunks and then shuffles the samples within each chunk. This preserves locality of reference while still shuffling the

Causal 3D convolution for video processing. Like LTXCausalConv3d, but whether causal inference is performed can be specified at runtime via t

MLP as used in Vision Transformer, MLP-Mixer and related networks

Multi-head Flash Attention block with qk norm.

Diffusion model with a Transformer backbone.

Diffusion model with a Transformer backbone.

Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.

The final layer of Sana.

Computes and stores the average and current value.

r"""Lightweight linear attention

`SCMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with non-Markovian guidance.

MLP as used in Vision Transformer, MLP-Mixer and related networks

Diffusion model with a Transformer backbone.

Diffusion model with a Transformer backbone.

Diffusion model with a Transformer backbone.

A sampler wrapper for grouping images with similar aspect ratio into a same batch. Args: sampler (Sampler): Base sampler. dataset

Causal self-attention with a single head.

This debug class helps detect and understand where the model starts getting very large or very small, and more importantly `nan` or `inf` wei

InceptionC block patched for FID computation

Pretrained InceptionV3 network returning feature maps

A 3D ResNet block used in the LTX 2.0 audiovisual model. Args: in_channels: Number of input channels. out_channels: Number of out

Configuration for LoRA (Low-Rank Adaptation) fine-tuning

Model growth initializer

r"""Self Attention with Perturbed Attention & CFG Guidance

r"""Self Attention with Perturbed Attention Guidance

3D Image to Patch Embedding

Per-pixel (per-location) RMS normalization layer. For each element along the chosen dimension, this layer normalizes the tensor by the root-mean-

r"""Self Attention with Lite Linear Attention

A sampler wrapper for grouping images with similar aspect ratio into a same batch. Args: sampler (Sampler): Base sampler. dataset

InceptionA block patched for FID computation

Per-call generation knobs.

3D downsampling layer for spatiotemporal reduction. Uses pixel unshuffle pattern for downsampling with causal temporal handling.

2D Image to Patch Embedding

Embeds scalar timesteps into vector representations.

Convert tensor data type from uint8 to float, divide value by 255.0 and permute the dimensions of clip tensor

Causal self-attention with a single head.

State container for decoder streaming cache. Attributes: feat_map: Per-layer feature cache for causal convolution is_first_chunk:

GLUMBConv with 1x1 Conv replaced by Linear layers. Original GLUMBConv structure: - inverted_conv: Conv2d(in_features, hidden_features*2,