Upload 2 files

app (5).py

@@ -0,0 +1,301 @@
+import gradio as gr
+import torch
+import numpy as np
+import tempfile
+import os
+
+from diffusers import LTXLatentUpsamplePipeline
+from pipeline_ltx_condition_control import LTXConditionPipeline, LTXVideoCondition
+from diffusers.utils import export_to_video, load_video
+from torchvision import transforms
+import random
+import imageio
+from PIL import Image, ImageOps
+import cv2
+import shutil
+import glob
+from pathlib import Path
+
+import warnings
+import logging
+warnings.filterwarnings("ignore", category=UserWarning)
+warnings.filterwarnings("ignore", category=FutureWarning)
+warnings.filterwarnings("ignore", message=".*")
+from huggingface_hub import logging as ll
+ll.set_verbosity_error()
+ll.set_verbosity_warning()
+ll.set_verbosity_info()
+ll.set_verbosity_debug()
+logger = logging.getLogger("AducDebug")
+logging.basicConfig(level=logging.DEBUG)
+logger.setLevel(logging.DEBUG)
+
+FPS = 24
+dtype = torch.bfloat16
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+base_model_repo = "Lightricks/LTX-Video"
+print(f"Carregando a arquitetura completa da pipeline de {base_model_repo}...")
+pipeline = LTXConditionPipeline.from_pretrained(
+    base_model_repo,
+    torch_dtype=dtype,
+    cache_dir=os.getenv("HF_HOME_CACHE"),
+    token=os.getenv("HF_TOKEN"),
+)
+
+# 2. Definir a URL para o arquivo de pesos FP8 que contém apenas o TRANSFORMER.
+fp8_transformer_weights_url = "https://huggingface.co/Lightricks/LTX-Video/ltxv-13b-0.9.8-distilled-fp8.safetensors"
+print(f"Sobrescrevendo pesos do Transformer com o arquivo FP8 de: {fp8_transformer_weights_url}")
+
+pipeline.load_lora_weights(fp8_transformer_weights_url, from_diffusers=True)
+
+print("Carregando upsampler...")
+pipe_upsample = LTXLatentUpsamplePipeline.from_pretrained(
+    "Lightricks/ltxv-spatial-upscaler-0.9.7",
+    cache_dir=os.getenv("HF_HOME_CACHE"),
+    vae=pipeline.vae, 
+    torch_dtype=dtype
+)
+
+print("Movendo modelos para o dispositivo...")
+pipeline.to(device)
+pipe_upsample.to(device)
+pipeline.vae.enable_tiling()
+
+current_dir = Path(__file__).parent
+
+def cleanup_session_files(request: gr.Request):
+    """Limpa arquivos temporários da sessão quando o usuário se desconecta."""
+    try:
+        session_id = request.session_hash
+        session_dir = os.path.join("/tmp/gradio", session_id)
+        if os.path.exists(session_dir):
+            shutil.rmtree(session_dir)
+            print(f"Limpou o diretório da sessão: {session_dir}")
+    except Exception as e:
+        print(f"Erro durante a limpeza da sessão: {e}")
+
+def read_video(video) -> torch.Tensor:
+    """Lê um arquivo de vídeo e converte para um tensor torch."""
+    to_tensor_transform = transforms.ToTensor()
+    if isinstance(video, str):
+        video_tensor = torch.stack([to_tensor_transform(img) for img in imageio.get_reader(video)])
+    else:
+        video_tensor = torch.stack([to_tensor_transform(img) for img in video])
+    return video_tensor
+
+
+def round_to_nearest_resolution_acceptable_by_vae(height, width, vae_temporal_compression_ratio):
+    """Arredonda a resolução para valores aceitáveis pelo VAE."""
+    height = height - (height % vae_temporal_compression_ratio)
+    width = width - (width % vae_temporal_compression_ratio)
+    return height, width
+
+
+# A assinatura da função volta a aceitar argumentos individuais para compatibilidade com o Gradio
+def generate_video(
+    condition_image_1,
+    condition_strength_1,
+    condition_frame_index_1,
+    condition_image_2,
+    condition_strength_2,
+    condition_frame_index_2,
+    prompt,
+    duration=3.0,
+    negative_prompt="worst quality, inconsistent motion, blurry, jittery, distorted",
+    height=768,
+    width=1152,
+    num_inference_steps=7,
+    guidance_scale=1.0,
+    seed=0,
+    randomize_seed=False,
+    progress=gr.Progress(track_tqdm=True)
+):
+    try:
+        # Lógica para agrupar as condições *dentro* da função
+        # Cálculo de frames e resolução
+        num_frames = int(duration * FPS) + 1
+        temporal_compression = pipeline.vae_temporal_compression_ratio
+        num_frames = ((num_frames - 1) // temporal_compression) * temporal_compression + 1
+        
+        downscale_factor = 2 / 3
+        downscaled_height = int(height * downscale_factor)
+        downscaled_width = int(width * downscale_factor)
+        downscaled_height, downscaled_width = round_to_nearest_resolution_acceptable_by_vae(
+            downscaled_height, downscaled_width, pipeline.vae_temporal_compression_ratio
+        )
+        
+        
+        
+        conditions = []
+        if condition_image_1 is not None:
+            condition_image_1 = ImageOps.fit(condition_image_1, (downscaled_width, downscaled_height), Image.LANCZOS)
+            conditions.append(LTXVideoCondition(
+                image=condition_image_1,
+                strength=condition_strength_1,
+                frame_index=int(condition_frame_index_1)
+            ))
+        if condition_image_2 is not None:
+            condition_image_2 = ImageOps.fit(condition_image_2, (downscaled_width, downscaled_height), Image.LANCZOS)
+            conditions.append(LTXVideoCondition(
+                image=condition_image_2,
+                strength=condition_strength_2,
+                frame_index=int(condition_frame_index_2)
+            ))
+
+        pipeline_args = {}
+        if conditions:
+            pipeline_args["conditions"] = conditions
+
+        # Manipulação da seed
+        if randomize_seed:
+            seed = random.randint(0, 2**32 - 1)
+
+                
+        # ETAPA 1: Geração do vídeo em baixa resolução
+        latents = pipeline(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            width=downscaled_width,
+            height=downscaled_height,
+            num_frames=num_frames,
+            timesteps=[1000, 993, 987, 981, 975, 909, 725, 0.03],
+            decode_timestep=0.05,
+            decode_noise_scale=0.025,
+            image_cond_noise_scale=0.0,
+            guidance_scale=guidance_scale,
+            guidance_rescale=0.7,
+            generator=torch.Generator().manual_seed(seed),
+            output_type="latent",
+            **pipeline_args
+        ).frames
+
+        # ETAPA 2: Upscale dos latentes
+        #upscaled_height, upscaled_width = downscaled_height * 2, downscaled_width * 2
+        #upscaled_latents = pipe_upsample(
+        #    latents=latents,
+        #    output_type="latent"
+        #).frames       
+
+        print(f"ETAPA 1 latents {latents.shape}")
+        
+        
+
+        # ETAPA 3: Denoise final em alta resolução
+        final_video_frames_np = pipeline(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            width=downscaled_width,
+            height=downscaled_height,
+            num_frames=num_frames,
+            denoise_strength=0.999,
+            timesteps=[1000, 909, 725, 421, 0],
+            latents=latents,
+            decode_timestep=0.05,
+            decode_noise_scale=0.025,
+            image_cond_noise_scale=0.0,
+            guidance_scale=guidance_scale,
+            guidance_rescale=0.7,
+            generator=torch.Generator(device="cuda").manual_seed(seed),
+            output_type="np",
+            **pipeline_args
+        ).frames[0]
+
+        print(f"ETAPA 3 final_video_frames_np {final_video_frames_np.shape}")
+
+        # Exportação para arquivo MP4
+        video_uint8_frames = [(frame * 255).astype(np.uint8) for frame in final_video_frames_np]
+        output_filename = "output.mp4"
+        with imageio.get_writer(output_filename, fps=FPS, quality=8, macro_block_size=1) as writer:
+             for frame_idx, frame_data in enumerate(video_uint8_frames):
+                progress((frame_idx + 1) / len(video_uint8_frames), desc="Codificando frames do vídeo...")
+                writer.append_data(frame_data)
+
+        return output_filename, seed
+        
+    except Exception as e:
+        print(f"Ocorreu um erro: {e}")
+        return None, seed
+
+# Interface Gráfica com Gradio
+with gr.Blocks(theme=gr.themes.Ocean(font=[gr.themes.GoogleFont("Lexend Deca"), "sans-serif"]), delete_cache=(60, 900)) as demo:
+    gr.Markdown(
+        """
+        # Geração de Vídeo com LTX
+        **Crie vídeos a partir de texto e imagens de condição usando o modelo LTX-Video.**
+        """
+    )
+    
+    with gr.Row():
+        with gr.Column(scale=1):
+    
+            prompt = gr.Textbox(
+                label="Prompt",
+                placeholder="Descreva o vídeo que você quer gerar...",
+                lines=3,
+                value="O Coringa em seu icônico terno roxo e cabelo verde, dançando sozinho em um quarto escuro e decadente. Seus movimentos são erráticos e imprevisíveis, alternando entre graciosos e caóticos enquanto ele se perde no momento. A câmera captura seus gestos teatrais, sua dança refletindo sua personalidade desequilibrada. Iluminação temperamental com sombras dançando pelas paredes, criando uma atmosfera de bela loucura."
+            )
+            
+            with gr.Accordion("Imagem de Condição 1", open=True):
+                condition_image_1 = gr.Image(label="Imagem de Condição 1", type="pil")
+                with gr.Row():
+                    condition_strength_1 = gr.Slider(label="Peso (Strength)", minimum=0.0, maximum=1.0, step=0.05, value=1.0)
+                    condition_frame_index_1 = gr.Number(label="Frame", value=0, precision=0)
+
+            with gr.Accordion("Imagem de Condição 2", open=False):
+                condition_image_2 = gr.Image(label="Imagem de Condição 2", type="pil")
+                with gr.Row():
+                    condition_strength_2 = gr.Slider(label="Peso (Strength)", minimum=0.0, maximum=1.0, step=0.05, value=1.0)
+                    condition_frame_index_2 = gr.Number(label="Frame", value=0, precision=0)
+            
+            duration = gr.Slider(label="Duração (segundos)", minimum=1.0, maximum=10.0, step=0.5, value=2)
+            
+            with gr.Accordion("Configurações Avançadas", open=False):
+                negative_prompt = gr.Textbox(label="Prompt Negativo", placeholder="O que você não quer no vídeo...", lines=2, value="pior qualidade, movimento inconsistente, embaçado, tremido, distorcido")
+                with gr.Row():
+                    height = gr.Slider(label="Altura", minimum=256, maximum=1536, step=32, value=768)
+                    width = gr.Slider(label="Largura", minimum=256, maximum=1536, step=32, value=1152)
+                
+                num_inference_steps = gr.Slider(label="Passos de Inferência", minimum=5, maximum=10, step=1, value=7, visible=False)
+                
+                with gr.Row():
+                    guidance_scale = gr.Slider(label="Escala de Orientação (Guidance)", minimum=1.0, maximum=5.0, step=0.1, value=1.0)
+                    
+                with gr.Row():
+                    randomize_seed = gr.Checkbox(label="Seed Aleatória", value=True)
+                    seed = gr.Number(label="Seed", value=0, precision=0)
+            
+            generate_btn = gr.Button("Gerar Vídeo", variant="primary", size="lg")
+        
+        with gr.Column(scale=1):          
+            output_video = gr.Video(label="Vídeo Gerado", height=400)
+    
+    # CORREÇÃO: A lista de inputs agora é "plana", contendo apenas componentes do Gradio
+    generate_btn.click(
+        fn=generate_video,
+        inputs=[
+            condition_image_1,
+            condition_strength_1,
+            condition_frame_index_1,
+            condition_image_2,
+            condition_strength_2,
+            condition_frame_index_2,
+            prompt,
+            duration,
+            negative_prompt,
+            height,
+            width,
+            num_inference_steps,
+            guidance_scale,
+            seed,
+            randomize_seed,
+        ],
+        outputs=[output_video, seed],
+        show_progress=True
+    )
+
+    demo.unload(cleanup_session_files)
+
+
+if __name__ == "__main__":
+    demo.queue().launch(server_name="0.0.0.0", server_port=7860, debug=True, show_error=True)

pipeline_ltx_condition_control (1).py

@@ -0,0 +1,1506 @@
+# 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 inspect
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import PIL.Image
+import torch
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.image_processor import PipelineImageInput
+from diffusers.loaders import FromSingleFileMixin, LTXVideoLoraLoaderMixin
+from diffusers.models.autoencoders import AutoencoderKLLTXVideo
+from diffusers.models.transformers import LTXVideoTransformer3DModel
+from diffusers.pipelines.ltx.pipeline_output import LTXPipelineOutput
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+from transformers import T5EncoderModel, T5TokenizerFast
+from torchvision.transforms.functional import center_crop, resize
+
+import warnings
+import logging
+from debug_utils import log_function_io
+warnings.filterwarnings("ignore", category=UserWarning)
+warnings.filterwarnings("ignore", category=FutureWarning)
+warnings.filterwarnings("ignore", message=".*")
+from huggingface_hub import logging as ll
+ll.set_verbosity_error()
+ll.set_verbosity_warning()
+ll.set_verbosity_info()
+ll.set_verbosity_debug()
+logger = logging.getLogger("AducDebug")
+logging.basicConfig(level=logging.DEBUG)
+logger.setLevel(logging.DEBUG)
+
+XLA_AVAILABLE= False
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers.pipelines.ltx.pipeline_ltx_condition import LTXConditionPipeline, LTXVideoCondition
+        >>> from diffusers.utils import export_to_video, load_video, load_image
+
+        >>> pipe = LTXConditionPipeline.from_pretrained("Lightricks/LTX-Video-0.9.5", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> # Load input image and video
+        >>> video = load_video(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cosmos/cosmos-video2world-input-vid.mp4"
+        ... )
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cosmos/cosmos-video2world-input.jpg"
+        ... )
+
+        >>> # Create conditioning objects
+        >>> condition1 = LTXVideoCondition(
+        ...     image=image,
+        ...     frame_index=0,
+        ... )
+        >>> condition2 = LTXVideoCondition(
+        ...     video=video,
+        ...     frame_index=80,
+        ... )
+
+        >>> prompt = "The video depicts a long, straight highway stretching into the distance, flanked by metal guardrails. The road is divided into multiple lanes, with a few vehicles visible in the far distance. The surrounding landscape features dry, grassy fields on one side and rolling hills on the other. The sky is mostly clear with a few scattered clouds, suggesting a bright, sunny day. And then the camera switch to a winding mountain road covered in snow, with a single vehicle traveling along it. The road is flanked by steep, rocky cliffs and sparse vegetation. The landscape is characterized by rugged terrain and a river visible in the distance. The scene captures the solitude and beauty of a winter drive through a mountainous region."
+        >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
+
+        >>> # Generate video
+        >>> generator = torch.Generator("cuda").manual_seed(0)
+        >>> # Text-only conditioning is also supported without the need to pass `conditions`
+        >>> video = pipe(
+        ...     conditions=[condition1, condition2],
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     width=768,
+        ...     height=512,
+        ...     num_frames=161,
+        ...     num_inference_steps=40,
+        ...     generator=generator,
+        ... ).frames[0]
+
+        >>> export_to_video(video, "output.mp4", fps=24)
+        ```
+"""
+
+
+@dataclass
+class LTXVideoCondition:
+    """
+    Defines a single frame-conditioning item for LTX Video - a single frame or a sequence of frames.
+
+    Attributes:
+        image (`PIL.Image.Image`):
+            The image to condition the video on.
+        video (`List[PIL.Image.Image]`):
+            The video to condition the video on.
+        frame_index (`int`):
+            The frame index at which the image or video will conditionally effect the video generation.
+        strength (`float`, defaults to `1.0`):
+            The strength of the conditioning effect. A value of `1.0` means the conditioning effect is fully applied.
+    """
+
+    image: Optional[PIL.Image.Image] = None
+    video: Optional[List[PIL.Image.Image]] = None
+    frame_index: int = 0
+    strength: float = 1.0
+
+
+# from LTX-Video/ltx_video/schedulers/rf.py
+def linear_quadratic_schedule(num_steps, threshold_noise=0.025, linear_steps=None):
+    if linear_steps is None:
+        linear_steps = num_steps // 2
+    if num_steps < 2:
+        return torch.tensor([1.0])
+    linear_sigma_schedule = [i * threshold_noise / linear_steps for i in range(linear_steps)]
+    threshold_noise_step_diff = linear_steps - threshold_noise * num_steps
+    quadratic_steps = num_steps - linear_steps
+    quadratic_coef = threshold_noise_step_diff / (linear_steps * quadratic_steps**2)
+    linear_coef = threshold_noise / linear_steps - 2 * threshold_noise_step_diff / (quadratic_steps**2)
+    const = quadratic_coef * (linear_steps**2)
+    quadratic_sigma_schedule = [
+        quadratic_coef * (i**2) + linear_coef * i + const for i in range(linear_steps, num_steps)
+    ]
+    sigma_schedule = linear_sigma_schedule + quadratic_sigma_schedule + [1.0]
+    sigma_schedule = [1.0 - x for x in sigma_schedule]
+    return torch.tensor(sigma_schedule[:-1])
+
+
+# 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_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# 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 LTXConditionPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin):
+    r"""
+    Pipeline for text/image/video-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).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+    
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLLTXVideo,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: LTXVideoTransformer3DModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            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
+        )
+        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.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 128
+        )
+
+        self.default_height = 512
+        self.default_width = 704
+        self.default_frames = 121
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 256,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        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)
+
+        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
+        prompt_attention_mask = prompt_attention_mask.bool().to(device)
+
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # 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
+
+    # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt
+    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 = 256,
+        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_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                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_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    def check_inputs(
+        self,
+        prompt,
+        conditions,
+        image,
+        video,
+        frame_index,
+        strength,
+        denoise_strength,
+        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,
+        reference_video=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}."
+                )
+
+        if conditions is not None and (image is not None or video is not None):
+            raise ValueError("If `conditions` is provided, `image` and `video` must not be provided.")
+
+        if conditions is None:
+            if isinstance(image, list) and isinstance(frame_index, list) and len(image) != len(frame_index):
+                raise ValueError(
+                    "If `conditions` is not provided, `image` and `frame_index` must be of the same length."
+                )
+            elif isinstance(image, list) and isinstance(strength, list) and len(image) != len(strength):
+                raise ValueError("If `conditions` is not provided, `image` and `strength` must be of the same length.")
+            elif isinstance(video, list) and isinstance(frame_index, list) and len(video) != len(frame_index):
+                raise ValueError(
+                    "If `conditions` is not provided, `video` and `frame_index` must be of the same length."
+                )
+            elif isinstance(video, list) and isinstance(strength, list) and len(video) != len(strength):
+                raise ValueError("If `conditions` is not provided, `video` and `strength` must be of the same length.")
+
+        if denoise_strength < 0 or denoise_strength > 1:
+            raise ValueError(f"The value of strength should in [0.0, 1.0] but is {denoise_strength}")
+
+        if reference_video is not None:
+            if not isinstance(reference_video, torch.Tensor):
+                raise ValueError(
+                    "`reference_video` must be a torch.Tensor with shape [F, C, H, W] as returned by read_video()."
+                )
+            if reference_video.ndim != 4:
+                raise ValueError(
+                    f"`reference_video` must be a 4D tensor with shape [F, C, H, W], but got shape {reference_video.shape}."
+                )
+
+    @staticmethod
+    def _prepare_video_ids(
+        batch_size: int,
+        num_frames: int,
+        height: int,
+        width: int,
+        patch_size: int = 1,
+        patch_size_t: int = 1,
+        device: torch.device = None,
+    ) -> torch.Tensor:
+        latent_sample_coords = torch.meshgrid(
+            torch.arange(0, num_frames, patch_size_t, device=device),
+            torch.arange(0, height, patch_size, device=device),
+            torch.arange(0, width, patch_size, device=device),
+            indexing="ij",
+        )
+        latent_sample_coords = torch.stack(latent_sample_coords, dim=0)
+        latent_coords = latent_sample_coords.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
+        latent_coords = latent_coords.reshape(batch_size, -1, num_frames * height * width)
+
+        return latent_coords
+
+    @staticmethod
+    def _scale_video_ids(
+        video_ids: torch.Tensor,
+        scale_factor: int = 32,
+        scale_factor_t: int = 8,
+        frame_index: int = 0,
+        device: torch.device = None,
+    ) -> torch.Tensor:
+        scaled_latent_coords = (
+            video_ids
+            * torch.tensor([scale_factor_t, scale_factor, scale_factor], device=video_ids.device)[None, :, None]
+        )
+        scaled_latent_coords[:, 0] = (scaled_latent_coords[:, 0] + 1 - scale_factor_t).clamp(min=0)
+        scaled_latent_coords[:, 0] += frame_index
+
+        return scaled_latent_coords
+
+
+    
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._pack_latents
+    @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
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._unpack_latents
+    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
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._normalize_latents
+    def _normalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Normalize 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_mean) * scaling_factor / latents_std
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._denormalize_latents
+    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
+
+    def trim_conditioning_sequence(self, start_frame: int, sequence_num_frames: int, target_num_frames: int):
+        """
+        Trim a conditioning sequence to the allowed number of frames.
+
+        Args:
+            start_frame (int): The target frame number of the first frame in the sequence.
+            sequence_num_frames (int): The number of frames in the sequence.
+            target_num_frames (int): The target number of frames in the generated video.
+        Returns:
+            int: updated sequence length
+        """
+        scale_factor = self.vae_temporal_compression_ratio
+        num_frames = min(sequence_num_frames, target_num_frames - start_frame)
+        # Trim down to a multiple of temporal_scale_factor frames plus 1
+        num_frames = (num_frames - 1) // scale_factor * scale_factor + 1
+        return num_frames
+
+    @staticmethod
+    def add_noise_to_image_conditioning_latents(
+        t: float,
+        init_latents: torch.Tensor,
+        latents: torch.Tensor,
+        noise_scale: float,
+        conditioning_mask: torch.Tensor,
+        generator,
+        eps=1e-6,
+    ):
+        """
+        Add timestep-dependent noise to the hard-conditioning latents. This helps with motion continuity, especially
+        when conditioned on a single frame.
+        """
+        noise = randn_tensor(
+            latents.shape,
+            generator=generator,
+            device=latents.device,
+            dtype=latents.dtype,
+        )
+        # Add noise only to hard-conditioning latents (conditioning_mask = 1.0)
+        need_to_noise = (conditioning_mask > 1.0 - eps).unsqueeze(-1)
+        noised_latents = init_latents + noise_scale * noise * (t**2)
+        latents = torch.where(need_to_noise, noised_latents, latents)
+        return latents
+
+    def prepare_latents(
+        self,
+        conditions: Optional[List[torch.Tensor]] = None,
+        condition_strength: Optional[List[float]] = None,
+        condition_frame_index: Optional[List[int]] = None,
+        batch_size: int = 1,
+        num_channels_latents: int = 128,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        num_prefix_latent_frames: int = 2,
+        sigma: Optional[torch.Tensor] = None,
+        latents: Optional[torch.Tensor] = None,
+        generator: Optional[torch.Generator] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, int]:
+        num_latent_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
+
+        shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width)
+
+        noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        if latents is not None and sigma is not None:
+            if latents.shape != shape:
+                raise ValueError(
+                    f"Latents shape {latents.shape} does not match expected shape {shape}. Please check the input."
+                )
+            latents = latents.to(device=device, dtype=dtype)
+            sigma = sigma.to(device=device, dtype=dtype)
+            latents = sigma * noise + (1 - sigma) * latents
+        else:
+            latents = noise
+
+        if len(conditions) > 0:
+            condition_latent_frames_mask = torch.zeros(
+                (batch_size, num_latent_frames), device=device, dtype=torch.float32
+            )
+
+            extra_conditioning_latents = []
+            extra_conditioning_video_ids = []
+            extra_conditioning_mask = []
+            extra_conditioning_num_latents = 0
+            for data, strength, frame_index in zip(conditions, condition_strength, condition_frame_index, strict=False):
+                condition_latents = retrieve_latents(self.vae.encode(data), generator=generator)
+                condition_latents = self._normalize_latents(
+                    condition_latents, self.vae.latents_mean, self.vae.latents_std
+                ).to(device, dtype=dtype)
+
+                num_data_frames = data.size(2)
+                num_cond_frames = condition_latents.size(2)
+
+                if frame_index == 0:
+                    latents[:, :, :num_cond_frames] = torch.lerp(
+                        latents[:, :, :num_cond_frames], condition_latents, strength
+                    )
+                    condition_latent_frames_mask[:, :num_cond_frames] = strength
+
+                else:
+                    if num_data_frames > 1:
+                        if num_cond_frames < num_prefix_latent_frames:
+                            raise ValueError(
+                                f"Number of latent frames must be at least {num_prefix_latent_frames} but got {num_data_frames}."
+                            )
+
+                        if num_cond_frames > num_prefix_latent_frames:
+                            start_frame = frame_index // self.vae_temporal_compression_ratio + num_prefix_latent_frames
+                            end_frame = start_frame + num_cond_frames - num_prefix_latent_frames
+                            latents[:, :, start_frame:end_frame] = torch.lerp(
+                                latents[:, :, start_frame:end_frame],
+                                condition_latents[:, :, num_prefix_latent_frames:],
+                                strength,
+                            )
+                            condition_latent_frames_mask[:, start_frame:end_frame] = strength
+                            condition_latents = condition_latents[:, :, :num_prefix_latent_frames]
+
+                    noise = randn_tensor(condition_latents.shape, generator=generator, device=device, dtype=dtype)
+                    condition_latents = torch.lerp(noise, condition_latents, strength)
+
+                    condition_video_ids = self._prepare_video_ids(
+                        batch_size,
+                        condition_latents.size(2),
+                        latent_height,
+                        latent_width,
+                        patch_size=self.transformer_spatial_patch_size,
+                        patch_size_t=self.transformer_temporal_patch_size,
+                        device=device,
+                    )
+                    condition_video_ids = self._scale_video_ids(
+                        condition_video_ids,
+                        scale_factor=self.vae_spatial_compression_ratio,
+                        scale_factor_t=self.vae_temporal_compression_ratio,
+                        frame_index=frame_index,
+                        device=device,
+                    )
+                    condition_latents = self._pack_latents(
+                        condition_latents,
+                        self.transformer_spatial_patch_size,
+                        self.transformer_temporal_patch_size,
+                    )
+                    condition_conditioning_mask = torch.full(
+                        condition_latents.shape[:2], strength, device=device, dtype=dtype
+                    )
+
+                    extra_conditioning_latents.append(condition_latents)
+                    extra_conditioning_video_ids.append(condition_video_ids)
+                    extra_conditioning_mask.append(condition_conditioning_mask)
+                    extra_conditioning_num_latents += condition_latents.size(1)
+
+        video_ids = self._prepare_video_ids(
+            batch_size,
+            num_latent_frames,
+            latent_height,
+            latent_width,
+            patch_size_t=self.transformer_temporal_patch_size,
+            patch_size=self.transformer_spatial_patch_size,
+            device=device,
+        )
+        if len(conditions) > 0:
+            conditioning_mask = condition_latent_frames_mask.gather(1, video_ids[:, 0])
+        else:
+            conditioning_mask, extra_conditioning_num_latents = None, 0
+        video_ids = self._scale_video_ids(
+            video_ids,
+            scale_factor=self.vae_spatial_compression_ratio,
+            scale_factor_t=self.vae_temporal_compression_ratio,
+            frame_index=0,
+            device=device,
+        )
+        latents = self._pack_latents(
+            latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+        )
+
+        if len(conditions) > 0 and len(extra_conditioning_latents) > 0:
+            latents = torch.cat([*extra_conditioning_latents, latents], dim=1)
+            video_ids = torch.cat([*extra_conditioning_video_ids, video_ids], dim=2)
+            conditioning_mask = torch.cat([*extra_conditioning_mask, conditioning_mask], dim=1)
+
+        return latents, conditioning_mask, video_ids, extra_conditioning_num_latents
+
+    def get_timesteps(self, sigmas, timesteps, num_inference_steps, strength):
+        num_steps = min(int(num_inference_steps * strength), num_inference_steps)
+        start_index = max(num_inference_steps - num_steps, 0)
+        sigmas = sigmas[start_index:]
+        timesteps = timesteps[start_index:]
+        return sigmas, timesteps, num_inference_steps - start_index
+
+    @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,
+        conditions: Union[LTXVideoCondition, List[LTXVideoCondition]] = None,
+        image: Union[PipelineImageInput, List[PipelineImageInput]] = None,
+        video: List[PipelineImageInput] = None,
+        frame_index: Union[int, List[int]] = 0,
+        strength: Union[float, List[float]] = 1.0,
+        denoise_strength: float = 1.0,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        frame_rate: int = 25,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 3,
+        guidance_rescale: float = 0.0,
+        image_cond_noise_scale: float = 0.15,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        reference_video: Optional[torch.Tensor] = None,
+        output_reference_comparison: bool = False,
+        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 = 256,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            conditions (`List[LTXVideoCondition], *optional*`):
+                The list of frame-conditioning items for the video generation.If not provided, conditions will be
+                created using `image`, `video`, `frame_index` and `strength`.
+            image (`PipelineImageInput` or `List[PipelineImageInput]`, *optional*):
+                The image or images to condition the video generation. If not provided, one has to pass `video` or
+                `conditions`.
+            video (`List[PipelineImageInput]`, *optional*):
+                The video to condition the video generation. If not provided, one has to pass `image` or `conditions`.
+            frame_index (`int` or `List[int]`, *optional*):
+                The frame index or frame indices at which the image or video will conditionally effect the video
+                generation. If not provided, one has to pass `conditions`.
+            strength (`float` or `List[float]`, *optional*):
+                The strength or strengths of the conditioning effect. If not provided, one has to pass `conditions`.
+            denoise_strength (`float`, defaults to `1.0`):
+                The strength of the noise added to the latents for editing. Higher strength leads to more noise added
+                to the latents, therefore leading to more differences between original video and generated video. This
+                is useful for video-to-video editing.
+            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`, defaults to `512`):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, defaults to `704`):
+                The width in pixels of the generated image. This is set to 848 by default for the best results.
+            num_frames (`int`, defaults to `161`):
+                The number of video frames to generate
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            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`, defaults to `3 `):
+                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://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                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 image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            reference_video (`torch.Tensor`, *optional*):
+                An optional reference video to guide the generation process. Should be a tensor with shape
+                [F, C, H, W] in range [0, 1] as returned by `read_video()` from video_utils. The reference video
+                will be encoded and concatenated to the latent sequence, providing global guidance while remaining
+                unchanged during denoising. The reference video can be of any size and will be automatically
+                resized and cropped to match the target dimensions.
+            output_reference_comparison (`bool`, defaults to `False`):
+                Whether to output a side-by-side comparison showing both the reference video (if provided) and the
+                generated video. If `False`, only the generated video is returned. Only applies when `reference_video`
+                is provided.
+            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.LTXPipelineOutput`] 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*):
+                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` defaults to `128 `):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] 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
+        # if latents is not None:
+        #     raise ValueError("Passing latents is not yet supported.")
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt=prompt,
+            conditions=conditions,
+            image=image,
+            video=video,
+            frame_index=frame_index,
+            strength=strength,
+            denoise_strength=denoise_strength,
+            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,
+            reference_video=reference_video,
+        )
+
+        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]
+
+        if conditions is not None:
+            if not isinstance(conditions, list):
+                conditions = [conditions]
+
+            strength = [condition.strength for condition in conditions]
+            frame_index = [condition.frame_index for condition in conditions]
+            image = [condition.image for condition in conditions]
+            video = [condition.video for condition in conditions]
+        elif image is not None or video is not None:
+            if not isinstance(image, list):
+                image = [image]
+                num_conditions = 1
+            elif isinstance(image, list):
+                num_conditions = len(image)
+            if not isinstance(video, list):
+                video = [video]
+                num_conditions = 1
+            elif isinstance(video, list):
+                num_conditions = len(video)
+
+            if not isinstance(frame_index, list):
+                frame_index = [frame_index] * num_conditions
+            if not isinstance(strength, list):
+                strength = [strength] * num_conditions
+
+        device = self._execution_device
+        vae_dtype = self.vae.dtype
+
+        # 3. Prepare text embeddings & conditioning image/video
+        (
+            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)
+
+        conditioning_tensors = []
+        is_conditioning_image_or_video = image is not None or video is not None
+        if is_conditioning_image_or_video:
+            for condition_image, condition_video, condition_frame_index, condition_strength in zip(
+                image, video, frame_index, strength, strict=False
+            ):
+                if condition_image is not None:
+                    condition_tensor = (
+                        self.video_processor.preprocess(condition_image, height, width)
+                        .unsqueeze(2)
+                        .to(device, dtype=vae_dtype)
+                    )
+                elif condition_video is not None:
+                    condition_tensor = self.video_processor.preprocess_video(condition_video, height, width)
+                    num_frames_input = condition_tensor.size(2)
+                    num_frames_output = self.trim_conditioning_sequence(
+                        condition_frame_index, num_frames_input, num_frames
+                    )
+                    condition_tensor = condition_tensor[:, :, :num_frames_output]
+                    condition_tensor = condition_tensor.to(device, dtype=vae_dtype)
+                else:
+                    raise ValueError("Either `image` or `video` must be provided for conditioning.")
+
+                if condition_tensor.size(2) % self.vae_temporal_compression_ratio != 1:
+                    raise ValueError(
+                        f"Number of frames in the video must be of the form (k * {self.vae_temporal_compression_ratio} + 1) "
+                        f"but got {condition_tensor.size(2)} frames."
+                    )
+                conditioning_tensors.append(condition_tensor)
+
+        # 4. Prepare timesteps
+        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 timesteps is None:
+            sigmas = linear_quadratic_schedule(num_inference_steps)
+            timesteps = sigmas * 1000
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+        sigmas = self.scheduler.sigmas
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        latent_sigma = None
+        if denoise_strength < 1:
+            sigmas, timesteps, num_inference_steps = self.get_timesteps(
+                sigmas, timesteps, num_inference_steps, denoise_strength
+            )
+            latent_sigma = sigmas[:1].repeat(batch_size * num_videos_per_prompt)
+
+        self._num_timesteps = len(timesteps)
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents, conditioning_mask, video_coords, extra_conditioning_num_latents = self.prepare_latents(
+            conditioning_tensors,
+            strength,
+            frame_index,
+            batch_size=batch_size * num_videos_per_prompt,
+            num_channels_latents=num_channels_latents,
+            height=height,
+            width=width,
+            num_frames=num_frames,
+            sigma=latent_sigma,
+            latents=latents,
+            generator=generator,
+            device=device,
+            dtype=torch.float32,
+        )
+
+        # 4.5. Process reference video (if provided) and concatenate at the beginning
+        reference_latents = None
+        reference_num_latents = 0
+        if reference_video is not None:
+            # Work with the original tensor format [F, C, H, W]
+            ref_frames = reference_video  # [F, C, H, W]
+
+            # Resize maintaining aspect ratio (resize all frames)
+            current_height, current_width = ref_frames.shape[2:]
+            aspect_ratio = current_width / current_height
+            target_aspect_ratio = width / height
+
+            if aspect_ratio > target_aspect_ratio:
+                # Width is relatively larger, resize based on height
+                resize_height = height
+                resize_width = int(resize_height * aspect_ratio)
+            else:
+                # Height is relatively larger, resize based on width
+                resize_width = width
+                resize_height = int(resize_width / aspect_ratio)
+
+            ref_frames = resize(ref_frames, [resize_height, resize_width], antialias=True)
+
+            # Center crop to target dimensions
+            ref_frames = center_crop(ref_frames, [height, width])
+
+            # Convert to VAE input format: [1, C, F, H, W] and proper range [-1, 1]
+            reference_tensor = ref_frames.unsqueeze(0).permute(0, 2, 1, 3, 4)  # [1, F, C, H, W] -> [1, C, F, H, W]
+            reference_tensor = reference_tensor * 2.0 - 1.0  # [0, 1] -> [-1, 1]
+
+            # Trim reference video to proper frame count for temporal compression
+            ref_num_frames_input = reference_tensor.size(2)
+            ref_num_frames_output = self.trim_conditioning_sequence(0, ref_num_frames_input, num_frames)
+            reference_tensor = reference_tensor[:, :, :ref_num_frames_output]
+            reference_tensor = reference_tensor.to(device, dtype=vae_dtype)
+
+            # Ensure proper frame count for VAE temporal compression
+            if reference_tensor.size(2) % self.vae_temporal_compression_ratio != 1:
+                # Trim to make it compatible with temporal compression
+                ref_frames_to_keep = (
+                    (reference_tensor.size(2) - 1) // self.vae_temporal_compression_ratio
+                ) * self.vae_temporal_compression_ratio + 1
+                reference_tensor = reference_tensor[:, :, :ref_frames_to_keep]
+
+            # Expand reference tensor for batch and num_videos_per_prompt
+            reference_tensor = reference_tensor.repeat(batch_size * num_videos_per_prompt, 1, 1, 1, 1)
+
+            # Encode reference video to latents
+            reference_latents = retrieve_latents(self.vae.encode(reference_tensor), generator=generator)
+            reference_latents = self._normalize_latents(
+                reference_latents, self.vae.latents_mean, self.vae.latents_std
+            ).to(device, dtype=torch.float32)
+
+            # Create "clean" coordinates for reference video (as if no frame conditioning applied)
+            ref_latent_frames = reference_latents.size(2)
+            ref_latent_height = reference_latents.size(3)
+            ref_latent_width = reference_latents.size(4)
+
+            reference_video_coords = self._prepare_video_ids(
+                batch_size * num_videos_per_prompt,
+                ref_latent_frames,
+                ref_latent_height,
+                ref_latent_width,
+                patch_size_t=self.transformer_temporal_patch_size,
+                patch_size=self.transformer_spatial_patch_size,
+                device=device,
+            )
+            reference_video_coords = self._scale_video_ids(
+                reference_video_coords,
+                scale_factor=self.vae_spatial_compression_ratio,
+                scale_factor_t=self.vae_temporal_compression_ratio,
+                frame_index=0,  # Reference video starts at frame 0
+                device=device,
+            )
+
+            # Pack reference latents
+            reference_latents = self._pack_latents(
+                reference_latents,
+                self.transformer_spatial_patch_size,
+                self.transformer_temporal_patch_size,
+            )
+            reference_num_latents = reference_latents.size(1)
+
+            # Concatenate reference latents at the beginning: [reference_latents, frame_conditions, target_latents]
+            latents = torch.cat([reference_latents, latents], dim=1)
+
+            # Update video coordinates: [reference_coords, existing_coords]
+            reference_video_coords = reference_video_coords.float()
+            video_coords = torch.cat([reference_video_coords, video_coords], dim=2)
+            video_coords[:, 0] = video_coords[:, 0] * (1.0 / frame_rate)
+
+            # Update conditioning mask to include reference (frozen = strength 1.0)
+            if conditioning_mask is not None:
+                reference_conditioning_mask = torch.ones(
+                    (batch_size * num_videos_per_prompt, reference_num_latents), device=device, dtype=torch.float32
+                )
+                conditioning_mask = torch.cat([reference_conditioning_mask, conditioning_mask], dim=1)
+            else:
+                # If no frame conditioning, still create mask for reference
+                conditioning_mask = torch.ones(
+                    (batch_size * num_videos_per_prompt, reference_num_latents), device=device, dtype=torch.float32
+                )
+                # Add zeros for target latents
+                target_conditioning_mask = torch.zeros(
+                    (batch_size * num_videos_per_prompt, latents.size(1) - reference_num_latents),
+                    device=device,
+                    dtype=torch.float32,
+                )
+                conditioning_mask = torch.cat([conditioning_mask, target_conditioning_mask], dim=1)
+
+        video_coords = video_coords.float()
+        if reference_video is None:
+            video_coords[:, 0] = video_coords[:, 0] * (1.0 / frame_rate)
+
+        init_latents = latents.clone() if is_conditioning_image_or_video or reference_video is not None else None
+
+        if self.do_classifier_free_guidance:
+            video_coords = torch.cat([video_coords, video_coords], dim=0)
+
+        # 6. 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
+
+                if image_cond_noise_scale > 0 and init_latents is not None:
+                    # Add timestep-dependent noise to the hard-conditioning latents
+                    # This helps with motion continuity, especially when conditioned on a single frame
+                    latents = self.add_noise_to_image_conditioning_latents(
+                        t / 1000.0,
+                        init_latents,
+                        latents,
+                        image_cond_noise_scale,
+                        conditioning_mask,
+                        generator,
+                    )
+
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                if is_conditioning_image_or_video or reference_video is not None:
+                    conditioning_mask_model_input = (
+                        torch.cat([conditioning_mask, conditioning_mask])
+                        if self.do_classifier_free_guidance
+                        else conditioning_mask
+                    )
+                latent_model_input = 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]).unsqueeze(-1).float()
+                if is_conditioning_image_or_video or reference_video is not None:
+                    timestep = torch.min(timestep, (1 - conditioning_mask_model_input) * 1000.0)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    encoder_attention_mask=prompt_attention_mask,
+                    video_coords=video_coords,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                    timestep, _ = timestep.chunk(2)
+
+                    if self.guidance_rescale > 0:
+                        # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                        noise_pred = rescale_noise_cfg(
+                            noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale
+                        )
+
+                denoised_latents = self.scheduler.step(
+                    -noise_pred, t, latents, per_token_timesteps=timestep, return_dict=False
+                )[0]
+                if is_conditioning_image_or_video or reference_video is not None:
+                    tokens_to_denoise_mask = (t / 1000 - 1e-6 < (1.0 - conditioning_mask)).unsqueeze(-1)
+                    latents = torch.where(tokens_to_denoise_mask, denoised_latents, latents)
+                else:
+                    latents = denoised_latents
+
+                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()
+
+        # Handle reference video output processing
+        if reference_video is not None and output_reference_comparison:
+            # Split latents: [reference_latents, frame_conditions, target_latents]
+            reference_latents_out = latents[:, :reference_num_latents]
+            remaining_latents = latents[:, reference_num_latents:]
+
+            # Remove frame conditioning from remaining latents if needed
+            if is_conditioning_image_or_video:
+                target_latents_out = remaining_latents[:, extra_conditioning_num_latents:]
+            else:
+                target_latents_out = remaining_latents
+
+            # Process both reference and target latents
+            videos = []
+            for curr_latents in [reference_latents_out, target_latents_out]:
+                if output_type == "latent":
+                    curr_video = curr_latents
+                else:
+                    curr_latents = self._unpack_latents(
+                        curr_latents,
+                        latent_num_frames,
+                        latent_height,
+                        latent_width,
+                        self.transformer_spatial_patch_size,
+                        self.transformer_temporal_patch_size,
+                    )
+                    curr_latents = self._denormalize_latents(
+                        curr_latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
+                    )
+                    curr_latents = curr_latents.to(prompt_embeds.dtype)
+
+                    if not self.vae.config.timestep_conditioning:
+                        timestep = None
+                    else:
+                        noise = torch.randn(
+                            curr_latents.shape, generator=generator, device=device, dtype=curr_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=curr_latents.dtype)
+                        decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=curr_latents.dtype)[
+                            :, None, None, None, None
+                        ]
+                        curr_latents = (1 - decode_noise_scale) * curr_latents + decode_noise_scale * noise
+
+                    curr_video = self.vae.decode(curr_latents, timestep, return_dict=False)[0]
+                    curr_video = self.video_processor.postprocess_video(curr_video, output_type=output_type)
+                videos.append(curr_video)
+
+            # Concatenate videos side-by-side (along width dimension for visual output)
+            if output_type == "latent":
+                video = torch.cat(videos, dim=0)
+            # For video tensors, shape is [B, C, F, H, W] or list of PIL images
+            elif isinstance(videos[0], list):
+                # Handle PIL images case - concatenate each frame side by side
+                video = []
+                for batch_idx in range(len(videos[0])):
+                    combined_video = []
+                    for frame_idx in range(len(videos[0][batch_idx])):
+                        ref_frame = videos[0][batch_idx][frame_idx]
+                        gen_frame = videos[1][batch_idx][frame_idx]
+                        # Create side-by-side comparison
+                        import PIL.Image
+
+                        if isinstance(ref_frame, PIL.Image.Image) and isinstance(gen_frame, PIL.Image.Image):
+                            combined_width = ref_frame.width + gen_frame.width
+                            combined_height = max(ref_frame.height, gen_frame.height)
+                            combined_frame = PIL.Image.new("RGB", (combined_width, combined_height))
+                            combined_frame.paste(ref_frame, (0, 0))
+                            combined_frame.paste(gen_frame, (ref_frame.width, 0))
+                            combined_video.append(combined_frame)
+                        else:
+                            combined_video.append(gen_frame)  # Fallback to generated only
+                    video.append(combined_video)
+            else:
+                # Handle tensor case - concatenate along width dimension (dim=4)
+                video = torch.cat(videos, dim=4)
+        else:
+            # Regular processing - just remove conditioning parts and output generated video
+            if reference_video is not None:
+                # Remove reference latents
+                latents = latents[:, reference_num_latents:]
+
+            if is_conditioning_image_or_video:
+                latents = latents[:, extra_conditioning_num_latents:]
+
+            latents = self._unpack_latents(
+                latents,
+                latent_num_frames,
+                latent_height,
+                latent_width,
+                self.transformer_spatial_patch_size,
+                self.transformer_temporal_patch_size,
+            )
+
+            if output_type == "latent":
+                video = latents
+            else:
+                latents = self._denormalize_latents(
+                    latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
+                )
+                latents = latents.to(prompt_embeds.dtype)
+
+                if not self.vae.config.timestep_conditioning:
+                    timestep = None
+                else:
+                    noise = torch.randn(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
+
+                video = self.vae.decode(latents, timestep, return_dict=False)[0]
+                video = self.video_processor.postprocess_video(video, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return LTXPipelineOutput(frames=video)