pipeline_ltx2.py

# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import copy
import inspect
from typing import Any, Callable, Dict, List, Optional, Union

import numpy as np
import torch
from transformers import Gemma3ForConditionalGeneration, GemmaTokenizer, GemmaTokenizerFast

from ...callbacks import MultiPipelineCallbacks, PipelineCallback
from ...loaders import FromSingleFileMixin, LTX2LoraLoaderMixin
from ...models.autoencoders import AutoencoderKLLTX2Audio, AutoencoderKLLTX2Video
from ...models.transformers import LTX2VideoTransformer3DModel
from ...schedulers import FlowMatchEulerDiscreteScheduler
from ...utils import is_torch_xla_available, logging, replace_example_docstring
from ...utils.torch_utils import randn_tensor
from ...video_processor import VideoProcessor
from ..pipeline_utils import DiffusionPipeline
from .connectors import LTX2TextConnectors
from .pipeline_output import LTX2PipelineOutput
from .vocoder import LTX2Vocoder


if is_torch_xla_available():
    import torch_xla.core.xla_model as xm

    XLA_AVAILABLE = True
else:
    XLA_AVAILABLE = False

logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from diffusers import LTX2Pipeline
        >>> from diffusers.pipelines.ltx2.export_utils import encode_video

        >>> pipe = LTX2Pipeline.from_pretrained("Lightricks/LTX-2", torch_dtype=torch.bfloat16)
        >>> pipe.enable_model_cpu_offload()

        >>> prompt = "A woman with long brown hair and light skin smiles at another woman with long blonde hair. The woman with brown hair wears a black jacket and has a small, barely noticeable mole on her right cheek. The camera angle is a close-up, focused on the woman with brown hair's face. The lighting is warm and natural, likely from the setting sun, casting a soft glow on the scene. The scene appears to be real-life footage"
        >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"

        >>> frame_rate = 24.0
        >>> video, audio = pipe(
        ...     prompt=prompt,
        ...     negative_prompt=negative_prompt,
        ...     width=768,
        ...     height=512,
        ...     num_frames=121,
        ...     frame_rate=frame_rate,
        ...     num_inference_steps=40,
        ...     guidance_scale=4.0,
        ...     output_type="np",
        ...     return_dict=False,
        ... )
        >>> video = (video * 255).round().astype("uint8")
        >>> video = torch.from_numpy(video)

        >>> encode_video(
        ...     video[0],
        ...     fps=frame_rate,
        ...     audio=audio[0].float().cpu(),
        ...     audio_sample_rate=pipe.vocoder.config.output_sampling_rate,  # should be 24000
        ...     output_path="video.mp4",
        ... )
        ```
"""


# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
def calculate_shift(
    image_seq_len,
    base_seq_len: int = 256,
    max_seq_len: int = 4096,
    base_shift: float = 0.5,
    max_shift: float = 1.15,
):
    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
    b = base_shift - m * base_seq_len
    mu = image_seq_len * m + b
    return mu


# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
def retrieve_timesteps(
    scheduler,
    num_inference_steps: Optional[int] = None,
    device: Optional[Union[str, torch.device]] = None,
    timesteps: Optional[List[int]] = None,
    sigmas: Optional[List[float]] = None,
    **kwargs,
):
    r"""
    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.

    Args:
        scheduler (`SchedulerMixin`):
            The scheduler to get timesteps from.
        num_inference_steps (`int`):
            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
            must be `None`.
        device (`str` or `torch.device`, *optional*):
            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
        timesteps (`List[int]`, *optional*):
            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
            `num_inference_steps` and `sigmas` must be `None`.
        sigmas (`List[float]`, *optional*):
            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
            `num_inference_steps` and `timesteps` must be `None`.

    Returns:
        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
        second element is the number of inference steps.
    """
    if timesteps is not None and sigmas is not None:
        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
    if timesteps is not None:
        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accepts_timesteps:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" timestep schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
    elif sigmas is not None:
        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accept_sigmas:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" sigmas schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
    else:
        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
        timesteps = scheduler.timesteps
    return timesteps, num_inference_steps


# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
    r"""
    Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on
    Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
    Flawed](https://huggingface.co/papers/2305.08891).

    Args:
        noise_cfg (`torch.Tensor`):
            The predicted noise tensor for the guided diffusion process.
        noise_pred_text (`torch.Tensor`):
            The predicted noise tensor for the text-guided diffusion process.
        guidance_rescale (`float`, *optional*, defaults to 0.0):
            A rescale factor applied to the noise predictions.

    Returns:
        noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor.
    """
    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
    # rescale the results from guidance (fixes overexposure)
    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
    return noise_cfg


class LTX2Pipeline(DiffusionPipeline, FromSingleFileMixin, LTX2LoraLoaderMixin):
    r"""
    Pipeline for text-to-video generation.

    Reference: https://github.com/Lightricks/LTX-Video

    Args:
        transformer ([`LTXVideoTransformer3DModel`]):
            Conditional Transformer architecture to denoise the encoded video latents.
        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
        vae ([`AutoencoderKLLTXVideo`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`T5EncoderModel`]):
            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
        tokenizer (`T5TokenizerFast`):
            Second Tokenizer of class
            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
        connectors ([`LTX2TextConnectors`]):
            Text connector stack used to adapt text encoder hidden states for the video and audio branches.
    """

    model_cpu_offload_seq = "text_encoder->connectors->transformer->vae->audio_vae->vocoder"
    _optional_components = []
    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]

    def __init__(
        self,
        scheduler: FlowMatchEulerDiscreteScheduler,
        vae: AutoencoderKLLTX2Video,
        audio_vae: AutoencoderKLLTX2Audio,
        text_encoder: Gemma3ForConditionalGeneration,
        tokenizer: Union[GemmaTokenizer, GemmaTokenizerFast],
        connectors: LTX2TextConnectors,
        transformer: LTX2VideoTransformer3DModel,
        vocoder: LTX2Vocoder,
    ):
        super().__init__()

        self.register_modules(
            vae=vae,
            audio_vae=audio_vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            connectors=connectors,
            transformer=transformer,
            vocoder=vocoder,
            scheduler=scheduler,
        )

        self.vae_spatial_compression_ratio = (
            self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32
        )
        self.vae_temporal_compression_ratio = (
            self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8
        )
        # TODO: check whether the MEL compression ratio logic here is corrct
        self.audio_vae_mel_compression_ratio = (
            self.audio_vae.mel_compression_ratio if getattr(self, "audio_vae", None) is not None else 4
        )
        self.audio_vae_temporal_compression_ratio = (
            self.audio_vae.temporal_compression_ratio if getattr(self, "audio_vae", None) is not None else 4
        )
        self.transformer_spatial_patch_size = (
            self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1
        )
        self.transformer_temporal_patch_size = (
            self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1
        )

        self.audio_sampling_rate = (
            self.audio_vae.config.sample_rate if getattr(self, "audio_vae", None) is not None else 16000
        )
        self.audio_hop_length = (
            self.audio_vae.config.mel_hop_length if getattr(self, "audio_vae", None) is not None else 160
        )

        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio)
        self.tokenizer_max_length = (
            self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 1024
        )

    @staticmethod
    def _pack_text_embeds(
        text_hidden_states: torch.Tensor,
        sequence_lengths: torch.Tensor,
        device: Union[str, torch.device],
        padding_side: str = "left",
        scale_factor: int = 8,
        eps: float = 1e-6,
    ) -> torch.Tensor:
        """
        Packs and normalizes text encoder hidden states, respecting padding. Normalization is performed per-batch and
        per-layer in a masked fashion (only over non-padded positions).

        Args:
            text_hidden_states (`torch.Tensor` of shape `(batch_size, seq_len, hidden_dim, num_layers)`):
                Per-layer hidden_states from a text encoder (e.g. `Gemma3ForConditionalGeneration`).
            sequence_lengths (`torch.Tensor of shape `(batch_size,)`):
                The number of valid (non-padded) tokens for each batch instance.
            device: (`str` or `torch.device`, *optional*):
                torch device to place the resulting embeddings on
            padding_side: (`str`, *optional*, defaults to `"left"`):
                Whether the text tokenizer performs padding on the `"left"` or `"right"`.
            scale_factor (`int`, *optional*, defaults to `8`):
                Scaling factor to multiply the normalized hidden states by.
            eps (`float`, *optional*, defaults to `1e-6`):
                A small positive value for numerical stability when performing normalization.

        Returns:
            `torch.Tensor` of shape `(batch_size, seq_len, hidden_dim * num_layers)`:
                Normed and flattened text encoder hidden states.
        """
        batch_size, seq_len, hidden_dim, num_layers = text_hidden_states.shape
        original_dtype = text_hidden_states.dtype

        # Create padding mask
        token_indices = torch.arange(seq_len, device=device).unsqueeze(0)
        if padding_side == "right":
            # For right padding, valid tokens are from 0 to sequence_length-1
            mask = token_indices < sequence_lengths[:, None]  # [batch_size, seq_len]
        elif padding_side == "left":
            # For left padding, valid tokens are from (T - sequence_length) to T-1
            start_indices = seq_len - sequence_lengths[:, None]  # [batch_size, 1]
            mask = token_indices >= start_indices  # [B, T]
        else:
            raise ValueError(f"padding_side must be 'left' or 'right', got {padding_side}")
        mask = mask[:, :, None, None]  # [batch_size, seq_len] --> [batch_size, seq_len, 1, 1]

        # Compute masked mean over non-padding positions of shape (batch_size, 1, 1, seq_len)
        masked_text_hidden_states = text_hidden_states.masked_fill(~mask, 0.0)
        num_valid_positions = (sequence_lengths * hidden_dim).view(batch_size, 1, 1, 1)
        masked_mean = masked_text_hidden_states.sum(dim=(1, 2), keepdim=True) / (num_valid_positions + eps)

        # Compute min/max over non-padding positions of shape (batch_size, 1, 1 seq_len)
        x_min = text_hidden_states.masked_fill(~mask, float("inf")).amin(dim=(1, 2), keepdim=True)
        x_max = text_hidden_states.masked_fill(~mask, float("-inf")).amax(dim=(1, 2), keepdim=True)

        # Normalization
        normalized_hidden_states = (text_hidden_states - masked_mean) / (x_max - x_min + eps)
        normalized_hidden_states = normalized_hidden_states * scale_factor

        # Pack the hidden states to a 3D tensor (batch_size, seq_len, hidden_dim * num_layers)
        normalized_hidden_states = normalized_hidden_states.flatten(2)
        mask_flat = mask.squeeze(-1).expand(-1, -1, hidden_dim * num_layers)
        normalized_hidden_states = normalized_hidden_states.masked_fill(~mask_flat, 0.0)
        normalized_hidden_states = normalized_hidden_states.to(dtype=original_dtype)
        return normalized_hidden_states

    def _get_gemma_prompt_embeds(
        self,
        prompt: Union[str, List[str]],
        num_videos_per_prompt: int = 1,
        max_sequence_length: int = 1024,
        scale_factor: int = 8,
        device: Optional[torch.device] = None,
        dtype: Optional[torch.dtype] = None,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            device: (`str` or `torch.device`):
                torch device to place the resulting embeddings on
            dtype: (`torch.dtype`):
                torch dtype to cast the prompt embeds to
            max_sequence_length (`int`, defaults to 1024): Maximum sequence length to use for the prompt.
        """
        device = device or self._execution_device
        dtype = dtype or self.text_encoder.dtype

        prompt = [prompt] if isinstance(prompt, str) else prompt
        batch_size = len(prompt)

        if getattr(self, "tokenizer", None) is not None:
            # Gemma expects left padding for chat-style prompts
            self.tokenizer.padding_side = "left"
            if self.tokenizer.pad_token is None:
                self.tokenizer.pad_token = self.tokenizer.eos_token

        prompt = [p.strip() for p in prompt]
        text_inputs = self.tokenizer(
            prompt,
            padding="max_length",
            max_length=max_sequence_length,
            truncation=True,
            add_special_tokens=True,
            return_tensors="pt",
        )
        text_input_ids = text_inputs.input_ids
        prompt_attention_mask = text_inputs.attention_mask
        text_input_ids = text_input_ids.to(device)
        prompt_attention_mask = prompt_attention_mask.to(device)

        text_encoder_outputs = self.text_encoder(
            input_ids=text_input_ids, attention_mask=prompt_attention_mask, output_hidden_states=True
        )
        text_encoder_hidden_states = text_encoder_outputs.hidden_states
        text_encoder_hidden_states = torch.stack(text_encoder_hidden_states, dim=-1)
        sequence_lengths = prompt_attention_mask.sum(dim=-1)

        prompt_embeds = self._pack_text_embeds(
            text_encoder_hidden_states,
            sequence_lengths,
            device=device,
            padding_side=self.tokenizer.padding_side,
            scale_factor=scale_factor,
        )
        prompt_embeds = prompt_embeds.to(dtype=dtype)

        # duplicate text embeddings for each generation per prompt, using mps friendly method
        _, seq_len, _ = prompt_embeds.shape
        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)

        prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
        prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)

        return prompt_embeds, prompt_attention_mask

    def encode_prompt(
        self,
        prompt: Union[str, List[str]],
        negative_prompt: Optional[Union[str, List[str]]] = None,
        do_classifier_free_guidance: bool = True,
        num_videos_per_prompt: int = 1,
        prompt_embeds: Optional[torch.Tensor] = None,
        negative_prompt_embeds: Optional[torch.Tensor] = None,
        prompt_attention_mask: Optional[torch.Tensor] = None,
        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
        max_sequence_length: int = 1024,
        scale_factor: int = 8,
        device: Optional[torch.device] = None,
        dtype: Optional[torch.dtype] = None,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
                Whether to use classifier free guidance or not.
            num_videos_per_prompt (`int`, *optional*, defaults to 1):
                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
            prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            device: (`torch.device`, *optional*):
                torch device
            dtype: (`torch.dtype`, *optional*):
                torch dtype
        """
        device = device or self._execution_device

        prompt = [prompt] if isinstance(prompt, str) else prompt
        if prompt is not None:
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        if prompt_embeds is None:
            prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds(
                prompt=prompt,
                num_videos_per_prompt=num_videos_per_prompt,
                max_sequence_length=max_sequence_length,
                scale_factor=scale_factor,
                device=device,
                dtype=dtype,
            )

        if do_classifier_free_guidance and negative_prompt_embeds is None:
            negative_prompt = negative_prompt or ""
            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt

            if prompt is not None and type(prompt) is not type(negative_prompt):
                raise TypeError(
                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
                    f" {type(prompt)}."
                )
            elif batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                    " the batch size of `prompt`."
                )

            negative_prompt_embeds, negative_prompt_attention_mask = self._get_gemma_prompt_embeds(
                prompt=negative_prompt,
                num_videos_per_prompt=num_videos_per_prompt,
                max_sequence_length=max_sequence_length,
                scale_factor=scale_factor,
                device=device,
                dtype=dtype,
            )

        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask

    def check_inputs(
        self,
        prompt,
        height,
        width,
        callback_on_step_end_tensor_inputs=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
        prompt_attention_mask=None,
        negative_prompt_attention_mask=None,
    ):
        if height % 32 != 0 or width % 32 != 0:
            raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.")

        if callback_on_step_end_tensor_inputs is not None and not all(
            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
        ):
            raise ValueError(
                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
            )

        if prompt is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )
        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")

        if prompt_embeds is not None and prompt_attention_mask is None:
            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")

        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")

        if prompt_embeds is not None and negative_prompt_embeds is not None:
            if prompt_embeds.shape != negative_prompt_embeds.shape:
                raise ValueError(
                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
                    f" {negative_prompt_embeds.shape}."
                )
            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
                raise ValueError(
                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
                    f" {negative_prompt_attention_mask.shape}."
                )

    @staticmethod
    def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor:
        # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p].
        # The patch dimensions are then permuted and collapsed into the channel dimension of shape:
        # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor).
        # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features
        batch_size, num_channels, num_frames, height, width = latents.shape
        post_patch_num_frames = num_frames // patch_size_t
        post_patch_height = height // patch_size
        post_patch_width = width // patch_size
        latents = latents.reshape(
            batch_size,
            -1,
            post_patch_num_frames,
            patch_size_t,
            post_patch_height,
            patch_size,
            post_patch_width,
            patch_size,
        )
        latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3)
        return latents

    @staticmethod
    def _unpack_latents(
        latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1
    ) -> torch.Tensor:
        # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions)
        # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of
        # what happens in the `_pack_latents` method.
        batch_size = latents.size(0)
        latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size)
        latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3)
        return latents

    @staticmethod
    def _denormalize_latents(
        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
    ) -> torch.Tensor:
        # Denormalize latents across the channel dimension [B, C, F, H, W]
        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
        latents = latents * latents_std / scaling_factor + latents_mean
        return latents

    @staticmethod
    def _normalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor):
        latents_mean = latents_mean.to(latents.device, latents.dtype)
        latents_std = latents_std.to(latents.device, latents.dtype)
        return (latents - latents_mean) / latents_std

    @staticmethod
    def _denormalize_audio_latents(latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor):
        latents_mean = latents_mean.to(latents.device, latents.dtype)
        latents_std = latents_std.to(latents.device, latents.dtype)
        return (latents * latents_std) + latents_mean

    @staticmethod
    def _pack_audio_latents(
        latents: torch.Tensor, patch_size: Optional[int] = None, patch_size_t: Optional[int] = None
    ) -> torch.Tensor:
        # Audio latents shape: [B, C, L, M], where L is the latent audio length and M is the number of mel bins
        if patch_size is not None and patch_size_t is not None:
            # Packs the latents into a patch sequence of shape [B, L // p_t * M // p, C * p_t * p] (a ndim=3 tnesor).
            # dim=1 is the effective audio sequence length and dim=2 is the effective audio input feature size.
            batch_size, num_channels, latent_length, latent_mel_bins = latents.shape
            post_patch_latent_length = latent_length / patch_size_t
            post_patch_mel_bins = latent_mel_bins / patch_size
            latents = latents.reshape(
                batch_size, -1, post_patch_latent_length, patch_size_t, post_patch_mel_bins, patch_size
            )
            latents = latents.permute(0, 2, 4, 1, 3, 5).flatten(3, 5).flatten(1, 2)
        else:
            # Packs the latents into a patch sequence of shape [B, L, C * M]. This implicitly assumes a (mel)
            # patch_size of M (all mel bins constitutes a single patch) and a patch_size_t of 1.
            latents = latents.transpose(1, 2).flatten(2, 3)  # [B, C, L, M] --> [B, L, C * M]
        return latents

    @staticmethod
    def _unpack_audio_latents(
        latents: torch.Tensor,
        latent_length: int,
        num_mel_bins: int,
        patch_size: Optional[int] = None,
        patch_size_t: Optional[int] = None,
    ) -> torch.Tensor:
        # Unpacks an audio patch sequence of shape [B, S, D] into a latent spectrogram tensor of shape [B, C, L, M],
        # where L is the latent audio length and M is the number of mel bins.
        if patch_size is not None and patch_size_t is not None:
            batch_size = latents.size(0)
            latents = latents.reshape(batch_size, latent_length, num_mel_bins, -1, patch_size_t, patch_size)
            latents = latents.permute(0, 3, 1, 4, 2, 5).flatten(4, 5).flatten(2, 3)
        else:
            # Assume [B, S, D] = [B, L, C * M], which implies that patch_size = M and patch_size_t = 1.
            latents = latents.unflatten(2, (-1, num_mel_bins)).transpose(1, 2)
        return latents

    def prepare_latents(
        self,
        batch_size: int = 1,
        num_channels_latents: int = 128,
        height: int = 512,
        width: int = 768,
        num_frames: int = 121,
        dtype: Optional[torch.dtype] = None,
        device: Optional[torch.device] = None,
        generator: Optional[torch.Generator] = None,
        latents: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        if latents is not None:
            if latents.ndim == 5:
                # latents are of shape [B, C, F, H, W], need to be packed
                latents = self._pack_latents(
                    latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
                )
            return latents.to(device=device, dtype=dtype)

        height = height // self.vae_spatial_compression_ratio
        width = width // self.vae_spatial_compression_ratio
        num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1

        shape = (batch_size, num_channels_latents, num_frames, height, width)

        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        latents = self._pack_latents(
            latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
        )
        return latents

    def prepare_audio_latents(
        self,
        batch_size: int = 1,
        num_channels_latents: int = 8,
        audio_latent_length: int = 1,  # 1 is just a dummy value
        num_mel_bins: int = 64,
        dtype: Optional[torch.dtype] = None,
        device: Optional[torch.device] = None,
        generator: Optional[torch.Generator] = None,
        latents: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        if latents is not None:
            if latents.ndim == 4:
                # latents are of shape [B, C, L, M], need to be packed
                latents = self._pack_audio_latents(latents)
            latents = self._normalize_audio_latents(latents, self.audio_vae.latents_mean, self.audio_vae.latents_std)
            return latents.to(device=device, dtype=dtype)

        # TODO: confirm whether this logic is correct
        latent_mel_bins = num_mel_bins // self.audio_vae_mel_compression_ratio

        shape = (batch_size, num_channels_latents, audio_latent_length, latent_mel_bins)

        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        latents = self._pack_audio_latents(latents)
        return latents

    @property
    def guidance_scale(self):
        return self._guidance_scale

    @property
    def guidance_rescale(self):
        return self._guidance_rescale

    @property
    def do_classifier_free_guidance(self):
        return self._guidance_scale > 1.0

    @property
    def num_timesteps(self):
        return self._num_timesteps

    @property
    def current_timestep(self):
        return self._current_timestep

    @property
    def attention_kwargs(self):
        return self._attention_kwargs

    @property
    def interrupt(self):
        return self._interrupt

    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        height: int = 512,
        width: int = 768,
        num_frames: int = 121,
        frame_rate: float = 24.0,
        num_inference_steps: int = 40,
        sigmas: Optional[List[float]] = None,
        timesteps: List[int] = None,
        guidance_scale: float = 4.0,
        guidance_rescale: float = 0.0,
        num_videos_per_prompt: Optional[int] = 1,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.Tensor] = None,
        audio_latents: Optional[torch.Tensor] = None,
        prompt_embeds: Optional[torch.Tensor] = None,
        prompt_attention_mask: Optional[torch.Tensor] = None,
        negative_prompt_embeds: Optional[torch.Tensor] = None,
        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
        decode_timestep: Union[float, List[float]] = 0.0,
        decode_noise_scale: Optional[Union[float, List[float]]] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        attention_kwargs: Optional[Dict[str, Any]] = None,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        max_sequence_length: int = 1024,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            height (`int`, *optional*, defaults to `512`):
                The height in pixels of the generated image. This is set to 480 by default for the best results.
            width (`int`, *optional*, defaults to `768`):
                The width in pixels of the generated image. This is set to 848 by default for the best results.
            num_frames (`int`, *optional*, defaults to `121`):
                The number of video frames to generate
            frame_rate (`float`, *optional*, defaults to `24.0`):
                The frames per second (FPS) of the generated video.
            num_inference_steps (`int`, *optional*, defaults to 40):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            sigmas (`List[float]`, *optional*):
                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
                will be used.
            timesteps (`List[int]`, *optional*):
                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
                passed will be used. Must be in descending order.
            guidance_scale (`float`, *optional*, defaults to `4.0`):
                Guidance scale as defined in [Classifier-Free Diffusion
                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
                of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
                `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to
                the text `prompt`, usually at the expense of lower image quality.
            guidance_rescale (`float`, *optional*, defaults to 0.0):
                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
                Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of
                [Common Diffusion Noise Schedules and Sample Steps are
                Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when
                using zero terminal SNR.
            num_videos_per_prompt (`int`, *optional*, defaults to 1):
                The number of videos to generate per prompt.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.Tensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for video
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will be generated by sampling using the supplied random `generator`.
            audio_latents (`torch.Tensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for audio
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will be generated by sampling using the supplied random `generator`.
            prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            prompt_attention_mask (`torch.Tensor`, *optional*):
                Pre-generated attention mask for text embeddings.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
            negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
                Pre-generated attention mask for negative text embeddings.
            decode_timestep (`float`, defaults to `0.0`):
                The timestep at which generated video is decoded.
            decode_noise_scale (`float`, defaults to `None`):
                The interpolation factor between random noise and denoised latents at the decode timestep.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.ltx.LTX2PipelineOutput`] instead of a plain tuple.
            attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            callback_on_step_end (`Callable`, *optional*):
                A function that calls at the end of each denoising steps during the inference. The function is called
                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
                `callback_on_step_end_tensor_inputs`.
            callback_on_step_end_tensor_inputs (`List`, *optional*, defaults to `["latents"]`):
                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                `._callback_tensor_inputs` attribute of your pipeline class.
            max_sequence_length (`int`, *optional*, defaults to `1024`):
                Maximum sequence length to use with the `prompt`.

        Examples:

        Returns:
            [`~pipelines.ltx.LTX2PipelineOutput`] or `tuple`:
                If `return_dict` is `True`, [`~pipelines.ltx.LTX2PipelineOutput`] is returned, otherwise a `tuple` is
                returned where the first element is a list with the generated images.
        """

        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt=prompt,
            height=height,
            width=width,
            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            prompt_attention_mask=prompt_attention_mask,
            negative_prompt_attention_mask=negative_prompt_attention_mask,
        )

        self._guidance_scale = guidance_scale
        self._guidance_rescale = guidance_rescale
        self._attention_kwargs = attention_kwargs
        self._interrupt = False
        self._current_timestep = None

        # 2. Define call parameters
        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        device = self._execution_device

        # 3. Prepare text embeddings
        (
            prompt_embeds,
            prompt_attention_mask,
            negative_prompt_embeds,
            negative_prompt_attention_mask,
        ) = self.encode_prompt(
            prompt=prompt,
            negative_prompt=negative_prompt,
            do_classifier_free_guidance=self.do_classifier_free_guidance,
            num_videos_per_prompt=num_videos_per_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            prompt_attention_mask=prompt_attention_mask,
            negative_prompt_attention_mask=negative_prompt_attention_mask,
            max_sequence_length=max_sequence_length,
            device=device,
        )
        if self.do_classifier_free_guidance:
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)

        additive_attention_mask = (1 - prompt_attention_mask.to(prompt_embeds.dtype)) * -1000000.0
        connector_prompt_embeds, connector_audio_prompt_embeds, connector_attention_mask = self.connectors(
            prompt_embeds, additive_attention_mask, additive_mask=True
        )

        # 4. Prepare latent variables
        latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
        latent_height = height // self.vae_spatial_compression_ratio
        latent_width = width // self.vae_spatial_compression_ratio
        if latents is not None:
            if latents.ndim == 5:
                _, _, latent_num_frames, latent_height, latent_width = latents.shape  # [B, C, F, H, W]
            else:
                logger.warning(
                    f"You have supplied packed `latents` of shape {latents.shape}, so the latent dims cannot be"
                    f" inferred. Make sure the supplied `height`, `width`, and `num_frames` are correct."
                )
        video_sequence_length = latent_num_frames * latent_height * latent_width

        num_channels_latents = self.transformer.config.in_channels
        latents = self.prepare_latents(
            batch_size * num_videos_per_prompt,
            num_channels_latents,
            height,
            width,
            num_frames,
            torch.float32,
            device,
            generator,
            latents,
        )

        duration_s = num_frames / frame_rate
        audio_latents_per_second = (
            self.audio_sampling_rate / self.audio_hop_length / float(self.audio_vae_temporal_compression_ratio)
        )
        audio_num_frames = round(duration_s * audio_latents_per_second)
        if audio_latents is not None:
            if audio_latents.ndim == 4:
                _, _, audio_num_frames, _ = audio_latents.shape  # [B, C, L, M]
            else:
                logger.warning(
                    f"You have supplied packed `audio_latents` of shape {audio_latents.shape}, so the latent dims"
                    f" cannot be inferred. Make sure the supplied `num_frames` is correct."
                )

        num_mel_bins = self.audio_vae.config.mel_bins if getattr(self, "audio_vae", None) is not None else 64
        latent_mel_bins = num_mel_bins // self.audio_vae_mel_compression_ratio
        num_channels_latents_audio = (
            self.audio_vae.config.latent_channels if getattr(self, "audio_vae", None) is not None else 8
        )
        audio_latents = self.prepare_audio_latents(
            batch_size * num_videos_per_prompt,
            num_channels_latents=num_channels_latents_audio,
            audio_latent_length=audio_num_frames,
            num_mel_bins=num_mel_bins,
            dtype=torch.float32,
            device=device,
            generator=generator,
            latents=audio_latents,
        )

        # 5. Prepare timesteps
        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
        mu = calculate_shift(
            video_sequence_length,
            self.scheduler.config.get("base_image_seq_len", 1024),
            self.scheduler.config.get("max_image_seq_len", 4096),
            self.scheduler.config.get("base_shift", 0.95),
            self.scheduler.config.get("max_shift", 2.05),
        )
        # For now, duplicate the scheduler for use with the audio latents
        audio_scheduler = copy.deepcopy(self.scheduler)
        _, _ = retrieve_timesteps(
            audio_scheduler,
            num_inference_steps,
            device,
            timesteps,
            sigmas=sigmas,
            mu=mu,
        )
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler,
            num_inference_steps,
            device,
            timesteps,
            sigmas=sigmas,
            mu=mu,
        )
        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
        self._num_timesteps = len(timesteps)

        # 6. Prepare micro-conditions
        rope_interpolation_scale = (
            self.vae_temporal_compression_ratio / frame_rate,
            self.vae_spatial_compression_ratio,
            self.vae_spatial_compression_ratio,
        )
        # Pre-compute video and audio positional ids as they will be the same at each step of the denoising loop
        video_coords = self.transformer.rope.prepare_video_coords(
            latents.shape[0], latent_num_frames, latent_height, latent_width, latents.device, fps=frame_rate
        )
        audio_coords = self.transformer.audio_rope.prepare_audio_coords(
            audio_latents.shape[0], audio_num_frames, audio_latents.device
        )

        # 7. Denoising loop
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                if self.interrupt:
                    continue

                self._current_timestep = t

                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
                latent_model_input = latent_model_input.to(prompt_embeds.dtype)
                audio_latent_model_input = (
                    torch.cat([audio_latents] * 2) if self.do_classifier_free_guidance else audio_latents
                )
                audio_latent_model_input = audio_latent_model_input.to(prompt_embeds.dtype)

                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                timestep = t.expand(latent_model_input.shape[0])

                with self.transformer.cache_context("cond_uncond"):
                    noise_pred_video, noise_pred_audio = self.transformer(
                        hidden_states=latent_model_input,
                        audio_hidden_states=audio_latent_model_input,
                        encoder_hidden_states=connector_prompt_embeds,
                        audio_encoder_hidden_states=connector_audio_prompt_embeds,
                        timestep=timestep,
                        encoder_attention_mask=connector_attention_mask,
                        audio_encoder_attention_mask=connector_attention_mask,
                        num_frames=latent_num_frames,
                        height=latent_height,
                        width=latent_width,
                        fps=frame_rate,
                        audio_num_frames=audio_num_frames,
                        video_coords=video_coords,
                        audio_coords=audio_coords,
                        # rope_interpolation_scale=rope_interpolation_scale,
                        attention_kwargs=attention_kwargs,
                        return_dict=False,
                    )
                noise_pred_video = noise_pred_video.float()
                noise_pred_audio = noise_pred_audio.float()

                if self.do_classifier_free_guidance:
                    noise_pred_video_uncond, noise_pred_video_text = noise_pred_video.chunk(2)
                    noise_pred_video = noise_pred_video_uncond + self.guidance_scale * (
                        noise_pred_video_text - noise_pred_video_uncond
                    )

                    noise_pred_audio_uncond, noise_pred_audio_text = noise_pred_audio.chunk(2)
                    noise_pred_audio = noise_pred_audio_uncond + self.guidance_scale * (
                        noise_pred_audio_text - noise_pred_audio_uncond
                    )

                    if self.guidance_rescale > 0:
                        # Based on 3.4. in https://huggingface.co/papers/2305.08891
                        noise_pred_video = rescale_noise_cfg(
                            noise_pred_video, noise_pred_video_text, guidance_rescale=self.guidance_rescale
                        )
                        noise_pred_audio = rescale_noise_cfg(
                            noise_pred_audio, noise_pred_audio_text, guidance_rescale=self.guidance_rescale
                        )

                # compute the previous noisy sample x_t -> x_t-1
                latents = self.scheduler.step(noise_pred_video, t, latents, return_dict=False)[0]
                # NOTE: for now duplicate scheduler for audio latents in case self.scheduler sets internal state in
                # the step method (such as _step_index)
                audio_latents = audio_scheduler.step(noise_pred_audio, t, audio_latents, return_dict=False)[0]

                if callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                    latents = callback_outputs.pop("latents", latents)
                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)

                # call the callback, if provided
                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()

                if XLA_AVAILABLE:
                    xm.mark_step()

        latents = self._unpack_latents(
            latents,
            latent_num_frames,
            latent_height,
            latent_width,
            self.transformer_spatial_patch_size,
            self.transformer_temporal_patch_size,
        )
        latents = self._denormalize_latents(
            latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
        )

        audio_latents = self._denormalize_audio_latents(
            audio_latents, self.audio_vae.latents_mean, self.audio_vae.latents_std
        )
        audio_latents = self._unpack_audio_latents(audio_latents, audio_num_frames, num_mel_bins=latent_mel_bins)

        if output_type == "latent":
            video = latents
            audio = audio_latents
        else:
            latents = latents.to(prompt_embeds.dtype)

            if not self.vae.config.timestep_conditioning:
                timestep = None
            else:
                noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype)
                if not isinstance(decode_timestep, list):
                    decode_timestep = [decode_timestep] * batch_size
                if decode_noise_scale is None:
                    decode_noise_scale = decode_timestep
                elif not isinstance(decode_noise_scale, list):
                    decode_noise_scale = [decode_noise_scale] * batch_size

                timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype)
                decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[
                    :, None, None, None, None
                ]
                latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise

            latents = latents.to(self.vae.dtype)
            video = self.vae.decode(latents, timestep, return_dict=False)[0]
            video = self.video_processor.postprocess_video(video, output_type=output_type)

            audio_latents = audio_latents.to(self.audio_vae.dtype)
            generated_mel_spectrograms = self.audio_vae.decode(audio_latents, return_dict=False)[0]
            audio = self.vocoder(generated_mel_spectrograms)

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (video, audio)

        return LTX2PipelineOutput(frames=video, audio=audio)