app(draft2).py

# =============================================================================
# Installation and Setup
# =============================================================================
import os
import subprocess
import sys

# Disable torch.compile / dynamo before any torch import
# This prevents CUDA initialization issues in the Space environment
os.environ["TORCH_COMPILE_DISABLE"] = "1"
os.environ["TORCHDYNAMO_DISABLE"] = "1"

# Clone LTX-2 repo at specific commit for reproducibility
# The commit ensures we have the exact pipeline code matching our analysis
LTX_REPO_URL = "https://github.com/Lightricks/LTX-2.git"
LTX_REPO_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "LTX-2")
# Using specific commit for stability - can be updated to main later
LTX_COMMIT_SHA = "a2c3f24078eb918171967f74b6f66b756b29ee45"

if not os.path.exists(LTX_REPO_DIR):
    print(f"Cloning {LTX_REPO_URL} at commit {LTX_COMMIT_SHA}...")
    os.makedirs(LTX_REPO_DIR)
    subprocess.run(["git", "init", LTX_REPO_DIR], check=True)
    subprocess.run(["git", "remote", "add", "origin", LTX_REPO_URL], cwd=LTX_REPO_DIR, check=True)
    subprocess.run(["git", "fetch", "--depth", "1", "origin", LTX_COMMIT_SHA], cwd=LTX_REPO_DIR, check=True)
    subprocess.run(["git", "checkout", LTX_COMMIT_SHA], cwd=LTX_REPO_DIR, check=True)

# Add repo packages to Python path
# This allows us to import from ltx-core and ltx-pipelines
sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines", "src"))
sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-core", "src"))

# =============================================================================
# Imports
# =============================================================================
import logging
import random
import tempfile
from pathlib import Path

import torch
# Disable torch.compile/dynamo at runtime level
torch._dynamo.config.suppress_errors = True
torch._dynamo.config.disable = True

import gradio as gr
import spaces
import numpy as np
from huggingface_hub import hf_hub_download, snapshot_download

# Import from the cloned LTX-2 pipeline
# These imports come from ti2vid_two_stages_hq.py
from ltx_core.model.video_vae import TilingConfig, get_video_chunks_number
from ltx_core.quantization import QuantizationPolicy
from ltx_core.loader import LoraPathStrengthAndSDOps
from ltx_pipelines.ti2vid_two_stages_hq import TI2VidTwoStagesHQPipeline
from ltx_pipelines.utils.args import ImageConditioningInput
from ltx_pipelines.utils.media_io import encode_video
from ltx_pipelines.utils.constants import LTX_2_3_HQ_PARAMS
from ltx_core.components.guiders import MultiModalGuiderParams

# =============================================================================
# Constants and Configuration
# =============================================================================

# Model repository on Hugging Face
LTX_MODEL_REPO = "Lightricks/LTX-2.3"
GEMMA_REPO = "Lightricks/gemma-3-12b-it-qat-q4_0-unquantized"

# Default parameters from LTX_2_3_HQ_PARAMS
DEFAULT_FRAME_RATE = 24.0

# Resolution constraints (must be divisible by 64 for two-stage pipeline)
# The pipeline generates at half-resolution in Stage 1, so input must be divisible by 2
MIN_DIM = 256
MAX_DIM = 1280
STEP = 64  # Both width and height must be divisible by 64

# Duration constraints (frames must be 8*K + 1)
MIN_FRAMES = 9   # 8*1 + 1
MAX_FRAMES = 257  # 8*32 + 1

# Seed range
MAX_SEED = np.iinfo(np.int32).max

# Default prompts
DEFAULT_PROMPT = (
    "A majestic eagle soaring over mountain peaks at sunset, "
    "wings spread wide against the orange sky, feathers catching the light, "
    "wind currents visible in the motion blur, cinematic slow motion, 4K quality"
)
DEFAULT_NEGATIVE_PROMPT = (
    "worst quality, inconsistent motion, blurry, jittery, distorted, "
    "deformed, artifacts, text, watermark, logo, frame, border, "
    "low resolution, pixelated, unnatural, fake, CGI, cartoon"
)

# =============================================================================
# Model Download and Initialization
# =============================================================================

print("=" * 80)
print("Downloading LTX-2.3 models...")
print("=" * 80)

# Download all required model files
# 1. Dev checkpoint - full trainable 22B model
checkpoint_path = hf_hub_download(
    repo_id=LTX_MODEL_REPO, 
    filename="ltx-2.3-22b-dev.safetensors"
)
print(f"Dev checkpoint: {checkpoint_path}")

# 2. Spatial upscaler - x2 upscaler for latent space
spatial_upsampler_path = hf_hub_download(
    repo_id=LTX_MODEL_REPO, 
    filename="ltx-2.3-spatial-upscaler-x2-1.1.safetensors"
)
print(f"Spatial upsampler: {spatial_upsampler_path}")

# 3. Distilled LoRA - distilled knowledge in LoRA format (rank 384)
# This LoRA is specifically trained to work with the dev model
distilled_lora_path = hf_hub_download(
    repo_id=LTX_MODEL_REPO, 
    filename="ltx-2.3-22b-distilled-lora-384.safetensors"
)
print(f"Distilled LoRA: {distilled_lora_path}")

# 4. Gemma text encoder - required for prompt encoding
gemma_root = snapshot_download(repo_id=GEMMA_REPO)
print(f"Gemma root: {gemma_root}")

print("=" * 80)
print("All models downloaded!")
print("=" * 80)

# =============================================================================
# Pipeline Initialization
# =============================================================================

# Create the LoraPathStrengthAndSDOps for distilled LoRA
# The sd_ops parameter uses the ComfyUI renaming map for compatibility
from ltx_core.loader import LTXV_LORA_COMFY_RENAMING_MAP

distilled_lora = [
    LoraPathStrengthAndSDOps(
        path=distilled_lora_path,
        strength=1.0,  # Will be set per-stage (0.25 for stage 1, 0.5 for stage 2)
        sd_ops=LTXV_LORA_COMFY_RENAMING_MAP,
    )
]

# Initialize the Two-Stage HQ Pipeline
# Key parameters:
# - checkpoint_path: Full dev model (trainable)
# - distilled_lora: LoRA containing distilled knowledge
# - distilled_lora_strength_stage_1: 0.25 (lighter application at half-res)
# - distilled_lora_strength_stage_2: 0.5 (stronger application after upscaling)
# - spatial_upsampler_path: Required for two-stage upscaling
# - gemma_root: Gemma text encoder for prompt encoding
print("Initializing LTX-2.3 Two-Stage HQ Pipeline...")

pipeline = TI2VidTwoStagesHQPipeline(
    checkpoint_path=checkpoint_path,
    distilled_lora=distilled_lora,
    distilled_lora_strength_stage_1=0.25,  # From HQ params
    distilled_lora_strength_stage_2=0.50,  # From HQ params
    spatial_upsampler_path=spatial_upsampler_path,
    gemma_root=gemma_root,
    loras=(),  # No additional custom LoRAs for this Space
    quantization=QuantizationPolicy.fp8_cast(),  # FP8 for memory efficiency
    torch_compile=False,  # Disable for Space compatibility
)

print("Pipeline initialized successfully!")
print("=" * 80)

# =============================================================================
# ZeroGPU Tensor Preloading - CPU Tensor Approach
# =============================================================================
# ZeroGPU should pack any tensors in memory, not just GPU tensors.
# We load model weights to CPU as proxy tensors to trigger packing.
# During actual generation, ZeroGPU will move them to GPU.

print("Creating CPU proxy tensors for ZeroGPU tensor packing...")
print("This may take a few minutes (loading to CPU only)...")

import gc

# Create small proxy tensors for each model component
# These don't need to be the actual weights - just tensors to trigger packing
# ZeroGPU will pack whatever tensors exist when it runs

_proxy_tensors = []

def create_proxy(name, shape, dtype=torch.float32):
    """Create a proxy tensor and ensure ZeroGPU sees it."""
    print(f"  Creating proxy for {name}: {shape}")
    t = torch.zeros(shape, dtype=dtype)
    _proxy_tensors.append(t)
    return t

# Create proxies for various model components
# These are just to ensure tensors exist in memory for ZeroGPU to pack
create_proxy("transformer_stage1", (1, 1024, 512))
create_proxy("transformer_stage2", (1, 1024, 512))
create_proxy("video_encoder", (1, 768, 512))
create_proxy("video_decoder", (1, 512, 512))
create_proxy("audio_decoder", (1, 256, 512))
create_proxy("spatial_upsampler", (1, 256, 512))
create_proxy("text_encoder", (1, 2048, 256))
create_proxy("vocoder", (1, 128, 256))

# Keep proxies alive by storing in module globals
proxy_stage1 = _proxy_tensors[0]
proxy_stage2 = _proxy_tensors[1]
proxy_venc = _proxy_tensors[2]
proxy_vdec = _proxy_tensors[3]
proxy_adec = _proxy_tensors[4]
proxy_upsamp = _proxy_tensors[5]
proxy_tenc = _proxy_tensors[6]
proxy_voc = _proxy_tensors[7]

# Clean up the temporary list
del _proxy_tensors

# Now trigger the actual model loading but catch GPU errors
print("\nAttempting model initialization (GPU errors expected)...")

try:
    # Try to access components - this will trigger loading but fail on GPU
    _ = pipeline.stage_1._transformer_ctx
    _ = pipeline.prompt_encoder._text_encoder_ctx
    print("  Model contexts accessed")
except Exception as e:
    print(f"  Context access: {type(e).__name__}")

print("\n" + "=" * 80)
print("Startup complete. Models will load to GPU during first generation.")
print("=" * 80)

# =============================================================================
# Helper Functions
# =============================================================================

def log_memory(tag: str):
    """Log current GPU memory usage for debugging."""
    if torch.cuda.is_available():
        allocated = torch.cuda.memory_allocated() / 1024**3
        peak = torch.cuda.max_memory_allocated() / 1024**3
        free, total = torch.cuda.mem_get_info()
        print(f"[VRAM {tag}] allocated={allocated:.2f}GB peak={peak:.2f}GB free={free / 1024**3:.2f}GB total={total / 1024**3:.2f}GB")


def calculate_frames(duration: float, frame_rate: float = DEFAULT_FRAME_RATE) -> int:
    """
    Calculate number of frames from duration.
    
    Frame count must be 8*K + 1 (K is a non-negative integer) for the LTX model.
    This constraint comes from the temporal upsampling architecture.
    
    Args:
        duration: Duration in seconds
        frame_rate: Frames per second
    
    Returns:
        Frame count that satisfies the 8*K + 1 constraint
    """
    ideal_frames = int(duration * frame_rate)
    # Ensure it's at least MIN_FRAMES
    ideal_frames = max(ideal_frames, MIN_FRAMES)
    # Round to nearest 8*K + 1
    k = round((ideal_frames - 1) / 8)
    frames = k * 8 + 1
    # Clamp to max
    return min(frames, MAX_FRAMES)


def validate_resolution(height: int, width: int) -> tuple[int, int]:
    """
    Ensure resolution is valid for two-stage pipeline.
    
    The two-stage pipeline requires:
    - Both dimensions divisible by 64 (for final resolution)
    - Stage 1 operates at half resolution (divisible by 32)
    
    Args:
        height: Target height
        width: Target width
    
    Returns:
        Validated (height, width) tuple
    """
    # Round to nearest multiple of 64
    height = round(height / STEP) * STEP
    width = round(width / STEP) * STEP
    
    # Clamp to valid range
    height = max(MIN_DIM, min(height, MAX_DIM))
    width = max(MIN_DIM, min(width, MAX_DIM))
    
    return height, width


def detect_aspect_ratio(image) -> str:
    """Detect the closest aspect ratio from an image for resolution presets."""
    if image is None:
        return "16:9"
    
    if hasattr(image, "size"):
        w, h = image.size
    elif hasattr(image, "shape"):
        h, w = image.shape[:2]
    else:
        return "16:9"
    
    ratio = w / h
    candidates = {"16:9": 16/9, "9:16": 9/16, "1:1": 1.0}
    return min(candidates, key=lambda k: abs(ratio - candidates[k]))


# Resolution presets based on aspect ratio
RESOLUTIONS = {
    "16:9": {"width": 1280, "height": 704},   # 960x540 * 1.33 = 1280x720, halved = 640x360 -> 1280x720
    "9:16": {"width": 704, "height": 1280},
    "1:1": {"width": 960, "height": 960},
}


def get_duration(
    prompt: str,
    negative_prompt: str,
    input_image,
    duration: float,
    seed: int,
    randomize_seed: bool,
    height: int,
    width: int,
    enhance_prompt: bool,
    video_cfg_scale: float,
    video_stg_scale: float,
    video_rescale_scale: float,
    video_a2v_scale: float,
    audio_cfg_scale: float,
    audio_stg_scale: float,
    audio_rescale_scale: float,
    audio_v2a_scale: float,
    progress,
) -> int:
    """
    Dynamically calculate GPU duration based on generation parameters.
    
    This is used by @spaces.GPU to set the appropriate time limit.
    Longer videos and higher resolution require more time.
    
    Args:
        duration: Video duration in seconds
        height, width: Resolution
        num_frames: Number of frames (indicates complexity)
    
    Returns:
        Duration in seconds for the GPU allocation
    """
    base = 60
    
    # Longer videos need more time
    if duration > 4:
        base += 15
    if duration > 6:
        base += 15
    
    # Higher resolution needs more time
    if height > 700 or width > 1000:
        base += 15
    
    # More frames means more processing
    # Calculate num_frames inside get_duration since it's no longer a parameter
    frames_from_duration = int(duration * DEFAULT_FRAME_RATE)
    if frames_from_duration > 81:
        base += 10


@spaces.GPU(duration=get_duration)
@torch.inference_mode()
def generate_video(
    prompt: str,
    negative_prompt: str,
    input_image,
    duration: float,
    seed: int,
    randomize_seed: bool,
    height: int,
    width: int,
    enhance_prompt: bool,
    # Guidance parameters
    video_cfg_scale: float,
    video_stg_scale: float,
    video_rescale_scale: float,
    video_a2v_scale: float,
    audio_cfg_scale: float,
    audio_stg_scale: float,
    audio_rescale_scale: float,
    audio_v2a_scale: float,
    progress=gr.Progress(track_tqdm=True),
):
    """
    Generate high-quality video using the Two-Stage HQ Pipeline.
    
    This function implements a two-stage generation process:
    
    Stage 1 (Half Resolution + CFG):
    - Generates video at half the target resolution
    - Uses GuidedDenoiser with CFG (positive + negative prompts)
    - Applies distilled LoRA at strength 0.25
    - Res2s sampler for efficient second-order denoising
    
    Stage 2 (Upscale + Refine):
    - Upscales latent representation 2x using spatial upsampler
    - Refines using SimpleDenoiser (no CFG, distilled approach)
    - Applies distilled LoRA at strength 0.5
    - 4-step refined denoising schedule
    
    Args:
        prompt: Text description of desired video content
        negative_prompt: What to avoid in the video
        input_image: Optional input image for image-to-video
        duration: Video duration in seconds
        seed: Random seed for reproducibility
        randomize_seed: Whether to use a random seed
        height, width: Target resolution (must be divisible by 64)
        enhance_prompt: Whether to use prompt enhancement
        video_cfg_scale: Video CFG (prompt adherence)
        video_stg_scale: Video STG (spatio-temporal guidance)
        video_rescale_scale: Video rescaling factor
        video_a2v_scale: Audio-to-video cross-attention scale
        audio_cfg_scale: Audio CFG (prompt adherence)
        audio_stg_scale: Audio STG (spatio-temporal guidance)
        audio_rescale_scale: Audio rescaling factor
        audio_v2a_scale: Video-to-audio cross-attention scale
    
    Returns:
        Tuple of (output_video_path, used_seed)
    """
    try:
        torch.cuda.reset_peak_memory_stats()
        log_memory("start")

        # Handle random seed
        current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
        print(f"Using seed: {current_seed}")

        # Validate and adjust resolution
        height, width = validate_resolution(int(height), int(width))
        print(f"Resolution: {width}x{height}")

        # Calculate frames (must be 8*K + 1)
        num_frames = calculate_frames(duration, DEFAULT_FRAME_RATE)
        print(f"Frames: {num_frames} ({duration}s @ {DEFAULT_FRAME_RATE}fps)")

        # Prepare image conditioning if provided
        images = []
        if input_image is not None:
            # Save input image temporarily
            output_dir = Path("outputs")
            output_dir.mkdir(exist_ok=True)
            temp_image_path = output_dir / f"temp_input_{current_seed}.jpg"
            
            if hasattr(input_image, "save"):
                input_image.save(temp_image_path)
            else:
                import shutil
                shutil.copy(input_image, temp_image_path)
            
            # Create ImageConditioningInput
            # path: image file path
            # frame_idx: target frame to condition on (0 = first frame)
            # strength: conditioning strength (1.0 = full influence)
            images = [ImageConditioningInput(
                path=str(temp_image_path),
                frame_idx=0,
                strength=1.0
            )]

        # Create tiling config for VAE decoding
        # Tiling is necessary to avoid OOM errors during decoding
        tiling_config = TilingConfig.default()
        video_chunks_number = get_video_chunks_number(num_frames, tiling_config)

        # Configure MultiModalGuider parameters
        # These control how the model adheres to prompts and handles modality guidance
        
        # Video guider parameters
        # cfg_scale: Classifier-free guidance scale (higher = stronger prompt adherence)
        # stg_scale: Spatio-temporal guidance scale (0 = disabled)
        # rescale_scale: Rescaling factor for oversaturation prevention
        # modality_scale: Cross-attention scale (audio-to-video)
        # skip_step: Step skipping for faster inference (0 = no skipping)
        # stg_blocks: Which transformer blocks to perturb for STG
        video_guider_params = MultiModalGuiderParams(
            cfg_scale=video_cfg_scale,
            stg_scale=video_stg_scale,
            rescale_scale=video_rescale_scale,
            modality_scale=video_a2v_scale,
            skip_step=0,
            stg_blocks=[],  # Empty for LTX 2.3 HQ
        )

        # Audio guider parameters
        audio_guider_params = MultiModalGuiderParams(
            cfg_scale=audio_cfg_scale,
            stg_scale=audio_stg_scale,
            rescale_scale=audio_rescale_scale,
            modality_scale=audio_v2a_scale,
            skip_step=0,
            stg_blocks=[],  # Empty for LTX 2.3 HQ
        )

        log_memory("before pipeline call")

        # Call the pipeline
        # The pipeline uses Res2sDiffusionStep for second-order sampling
        # Stage 1: num_inference_steps from LTX_2_3_HQ_PARAMS (15 steps)
        # Stage 2: Fixed 4-step schedule from STAGE_2_DISTILLED_SIGMAS
        video, audio = pipeline(
            prompt=prompt,
            negative_prompt=negative_prompt,
            seed=current_seed,
            height=height,
            width=width,
            num_frames=num_frames,
            frame_rate=DEFAULT_FRAME_RATE,
            num_inference_steps=LTX_2_3_HQ_PARAMS.num_inference_steps,  # 15 steps
            video_guider_params=video_guider_params,
            audio_guider_params=audio_guider_params,
            images=images,
            tiling_config=tiling_config,
            enhance_prompt=enhance_prompt,
        )

        log_memory("after pipeline call")

        # Encode video with audio
        output_path = tempfile.mktemp(suffix=".mp4")
        encode_video(
            video=video,
            fps=DEFAULT_FRAME_RATE,
            audio=audio,
            output_path=output_path,
            video_chunks_number=video_chunks_number,
        )

        log_memory("after encode_video")
        return str(output_path), current_seed

    except Exception as e:
        import traceback
        log_memory("on error")
        print(f"Error: {str(e)}\n{traceback.format_exc()}")
        return None, current_seed


# =============================================================================
# Gradio UI
# =============================================================================

css = """
/* Custom styling for LTX-2.3 Space */
.fillable {max-width: 1200px !important}
.progress-text {color: white}
"""

with gr.Blocks(title="LTX-2.3 Two-Stage HQ Video Generation") as demo:
    gr.Markdown("# LTX-2.3 Two-Stage HQ Video Generation")
    gr.Markdown(
        "High-quality text/image-to-video generation using the dev model + distilled LoRA. "
        "[[Model]](https://huggingface.co/Lightricks/LTX-2.3) "
        "[[GitHub]](https://github.com/Lightricks/LTX-2)"
    )

    with gr.Row():
        # Input Column
        with gr.Column():
            # Input image (optional)
            input_image = gr.Image(
                label="Input Image (Optional - for image-to-video)",
                type="pil",
                sources=["upload", "webcam", "clipboard"]
            )
            
            # Prompt inputs
            prompt = gr.Textbox(
                label="Prompt",
                info="Describe the video you want to generate",
                value=DEFAULT_PROMPT,
                lines=3,
                placeholder="Enter your prompt here..."
            )
            
            negative_prompt = gr.Textbox(
                label="Negative Prompt",
                info="What to avoid in the generated video",
                value=DEFAULT_NEGATIVE_PROMPT,
                lines=2,
                placeholder="Enter negative prompt here..."
            )
            
            # Duration slider
            duration = gr.Slider(
                label="Duration (seconds)",
                minimum=0.5,
                maximum=8.0,
                value=2.0,
                step=0.1,
                info="Video duration (clamped to 8K+1 frames)"
            )
            
            # Enhance prompt toggle
            enhance_prompt = gr.Checkbox(
                label="Enhance Prompt",
                value=False,
                info="Use Gemma to enhance the prompt for better results"
            )

            # Generate button
            generate_btn = gr.Button("Generate Video", variant="primary", size="lg")

        # Output Column
        with gr.Column():
            output_video = gr.Video(
                label="Generated Video",
                autoplay=True,
                interactive=False
            )

    # Advanced Settings Accordion
    with gr.Accordion("Advanced Settings", open=False):
        with gr.Row():
            # Resolution inputs
            width = gr.Number(
                label="Width",
                value=1280,
                precision=0,
                info="Must be divisible by 64"
            )
            height = gr.Number(
                label="Height",
                value=704,
                precision=0,
                info="Must be divisible by 64"
            )
        
        with gr.Row():
            # Seed controls
            seed = gr.Number(
                label="Seed",
                value=42,
                precision=0,
                minimum=0,
                maximum=MAX_SEED
            )
            randomize_seed = gr.Checkbox(
                label="Randomize Seed",
                value=True
            )
        
        gr.Markdown("### Video Guidance Parameters")
        gr.Markdown("Control how strongly the model follows the video prompt and handles guidance.")
        
        with gr.Row():
            video_cfg_scale = gr.Slider(
                label="Video CFG Scale",
                minimum=1.0,
                maximum=10.0,
                value=LTX_2_3_HQ_PARAMS.video_guider_params.cfg_scale,
                step=0.1,
                info="Classifier-free guidance for video (higher = stronger prompt adherence)"
            )
            video_stg_scale = gr.Slider(
                label="Video STG Scale",
                minimum=0.0,
                maximum=2.0,
                value=0.0,
                step=0.1,
                info="Spatio-temporal guidance (0 = disabled)"
            )
        
        with gr.Row():
            video_rescale_scale = gr.Slider(
                label="Video Rescale",
                minimum=0.0,
                maximum=2.0,
                value=0.45,
                step=0.1,
                info="Rescaling factor for oversaturation prevention"
            )
            video_a2v_scale = gr.Slider(
                label="A2V Scale",
                minimum=0.0,
                maximum=5.0,
                value=3.0,
                step=0.1,
                info="Audio-to-video cross-attention scale"
            )
        
        gr.Markdown("### Audio Guidance Parameters")
        gr.Markdown("Control audio generation quality and sync.")
        
        with gr.Row():
            audio_cfg_scale = gr.Slider(
                label="Audio CFG Scale",
                minimum=1.0,
                maximum=15.0,
                value=LTX_2_3_HQ_PARAMS.audio_guider_params.cfg_scale,
                step=0.1,
                info="Classifier-free guidance for audio"
            )
            audio_stg_scale = gr.Slider(
                label="Audio STG Scale",
                minimum=0.0,
                maximum=2.0,
                value=0.0,
                step=0.1,
                info="Spatio-temporal guidance for audio (0 = disabled)"
            )
        
        with gr.Row():
            audio_rescale_scale = gr.Slider(
                label="Audio Rescale",
                minimum=0.0,
                maximum=2.0,
                value=1.0,
                step=0.1,
                info="Audio rescaling factor"
            )
            audio_v2a_scale = gr.Slider(
                label="V2A Scale",
                minimum=0.0,
                maximum=5.0,
                value=3.0,
                step=0.1,
                info="Video-to-audio cross-attention scale"
            )

    # Event handlers
    def on_image_upload(image, current_h, current_w):
        """Update resolution based on uploaded image aspect ratio."""
        if image is None:
            return gr.update(), gr.update()
        
        aspect = detect_aspect_ratio(image)
        if aspect in RESOLUTIONS:
            return (
                gr.update(value=RESOLUTIONS[aspect]["width"]),
                gr.update(value=RESOLUTIONS[aspect]["height"])
            )
        return gr.update(), gr.update()

    input_image.change(
        fn=on_image_upload,
        inputs=[input_image, height, width],
        outputs=[width, height],
    )

    # Generate button click handler
    generate_btn.click(
        fn=generate_video,
        inputs=[
            prompt,
            negative_prompt,
            input_image,
            duration,
            seed,
            randomize_seed,
            height,
            width,
            enhance_prompt,
            video_cfg_scale,
            video_stg_scale,
            video_rescale_scale,
            video_a2v_scale,
            audio_cfg_scale,
            audio_stg_scale,
            audio_rescale_scale,
            audio_v2a_scale,
        ],
        outputs=[output_video, seed],
    )


# =============================================================================
# Main Entry Point
# =============================================================================

if __name__ == "__main__":
    demo.queue().launch(
        theme=gr.themes.Citrus(),
        css=css,
        mcp_server=True,
        share=True,
    )