app.py

# PyTorch 2.8 (temporary hack)
import os

os.system(
    'pip install --upgrade --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu126 "torch<2.9" spaces'
)

# Actual demo code
import spaces
import torch
from diffusers import LTXConditionPipeline
from diffusers.pipelines.ltx.pipeline_ltx_condition import LTXVideoCondition
from diffusers.utils import export_to_video, load_video
import gradio as gr
import tempfile
import numpy as np
from PIL import Image
import random
from optimization import optimize_pipeline_

MODEL_ID = "Lightricks/LTX-Video-0.9.8-13B-distilled"

LANDSCAPE_WIDTH = 480
LANDSCAPE_HEIGHT = 832
MAX_SEED = np.iinfo(np.int32).max

FIXED_FPS = 24
MIN_FRAMES_MODEL = 8
MAX_FRAMES_MODEL = 96

MIN_DURATION = round(MIN_FRAMES_MODEL / FIXED_FPS, 1)
MAX_DURATION = round(MAX_FRAMES_MODEL / FIXED_FPS, 1)

pipe = LTXConditionPipeline.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16).to("cuda")
dummy_image = Image.new("RGB", (LANDSCAPE_WIDTH, LANDSCAPE_HEIGHT))
video = load_video(export_to_video([dummy_image]))
dummy_cond = LTXVideoCondition(video=video, frame_index=0)
optimize_pipeline_(
    pipe,
    conditions=[dummy_cond],
    prompt="prompt",
    negative_prompt="prompt",
    guidance_scale=1.0,
    height=LANDSCAPE_HEIGHT,
    width=LANDSCAPE_WIDTH,
    num_frames=MAX_FRAMES_MODEL,
    num_inference_steps=2
)

default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
default_negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards, watermark, text, signature"


def resize_image(image: Image.Image) -> Image.Image:
    if image.height > image.width:
        transposed = image.transpose(Image.Transpose.ROTATE_90)
        resized = resize_image_landscape(transposed)
        return resized.transpose(Image.Transpose.ROTATE_270)
    return resize_image_landscape(image)


def resize_image_landscape(image: Image.Image) -> Image.Image:
    target_aspect = LANDSCAPE_WIDTH / LANDSCAPE_HEIGHT
    width, height = image.size
    in_aspect = width / height
    if in_aspect > target_aspect:
        new_width = round(height * target_aspect)
        left = (width - new_width) // 2
        image = image.crop((left, 0, left + new_width, height))
    else:
        new_height = round(width / target_aspect)
        top = (height - new_height) // 2
        image = image.crop((0, top, width, top + new_height))
    return image.resize((LANDSCAPE_WIDTH, LANDSCAPE_HEIGHT), Image.LANCZOS)


def get_duration(
    input_image,
    prompt,
    negative_prompt,
    duration_seconds,
    guidance_scale,
    steps,
    seed,
    randomize_seed,
    progress,
):
    if steps > 4 and duration_seconds > 2:
        return 90
    elif steps > 4 or duration_seconds > 2:
        return 75
    else:
        return 60


@spaces.GPU(duration=get_duration)
def generate_video(
    input_image,
    prompt,
    negative_prompt=default_negative_prompt,
    duration_seconds=MAX_DURATION,
    guidance_scale=1.0,
    steps=8,
    seed=42,
    randomize_seed=False,
    progress=gr.Progress(track_tqdm=True),
):
    """
    Generate a video from an input image using the LTX distilled model.

    This function takes an input image and generates a video animation based on the provided
    prompt and parameters. It uses the LTX 13B Distilled Image-to-Video model for fast generation
    in 4-8 steps.

    Args:
        input_image (PIL.Image): The input image to animate. Will be resized to target dimensions.
        prompt (str): Text prompt describing the desired animation or motion.
        negative_prompt (str, optional): Negative prompt to avoid unwanted elements.
            Defaults to default_negative_prompt (contains unwanted visual artifacts).
        duration_seconds (float, optional): Duration of the generated video in seconds.
            Defaults to 2. Clamped between MIN_FRAMES_MODEL/FIXED_FPS and MAX_FRAMES_MODEL/FIXED_FPS.
        guidance_scale (float, optional): Controls adherence to the prompt. Higher values = more adherence.
            Defaults to 1.0. Range: 0.0-20.0.
        steps (int, optional): Number of inference steps. More steps = higher quality but slower.
            Defaults to 4. Range: 1-30.
        seed (int, optional): Random seed for reproducible results. Defaults to 42.
            Range: 0 to MAX_SEED (2147483647).
        randomize_seed (bool, optional): Whether to use a random seed instead of the provided seed.
            Defaults to False.
        progress (gr.Progress, optional): Gradio progress tracker. Defaults to gr.Progress(track_tqdm=True).

    Returns:
        tuple: A tuple containing:
            - video_path (str): Path to the generated video file (.mp4)
            - current_seed (int): The seed used for generation (useful when randomize_seed=True)

    Raises:
        gr.Error: If input_image is None (no image uploaded).

    Note:
        - The function automatically resizes the input image to the target dimensions
        - Frame count is calculated as duration_seconds * FIXED_FPS (24)
        - Output dimensions are adjusted to be multiples of MOD_VALUE (32)
        - The function uses GPU acceleration via the @spaces.GPU decorator
        - Generation time varies based on steps and duration (see get_duration function)
    """
    if input_image is None:
        raise gr.Error("Please upload an input image.")

    num_frames = np.clip(int(round(duration_seconds * FIXED_FPS)), MIN_FRAMES_MODEL, MAX_FRAMES_MODEL)
    current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
    resized_image = resize_image(input_image)

    video = load_video(export_to_video([resized_image]))
    condition1 = LTXVideoCondition(video=video, frame_index=0)
    output_frames_list = pipe(
        conditions=[condition1],
        prompt=prompt,
        negative_prompt=negative_prompt,
        height=resized_image.height,
        width=resized_image.width,
        num_frames=num_frames,
        guidance_scale=float(guidance_scale),
        num_inference_steps=int(steps),
        generator=torch.Generator(device="cuda").manual_seed(current_seed),
    ).frames[0]

    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
        video_path = tmpfile.name

    export_to_video(output_frames_list, video_path, fps=FIXED_FPS)

    return video_path, current_seed


with gr.Blocks() as demo:
    gr.Markdown("# Fast few-steps LTX 0.9.8 I2V (13B)")
    with gr.Row():
        with gr.Column():
            input_image_component = gr.Image(type="pil", label="Input Image (auto-resized to target H/W)")
            prompt_input = gr.Textbox(label="Prompt", value=default_prompt_i2v)
            duration_seconds_input = gr.Slider(
                minimum=MIN_DURATION,
                maximum=MAX_DURATION,
                step=0.1,
                value=MAX_DURATION,
                label="Duration (seconds)",
                info=f"Clamped to model's {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps.",
            )

            with gr.Accordion("Advanced Settings", open=False):
                negative_prompt_input = gr.Textbox(label="Negative Prompt", value=default_negative_prompt, lines=3)
                seed_input = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42, interactive=True)
                randomize_seed_checkbox = gr.Checkbox(label="Randomize seed", value=True, interactive=True)
                steps_slider = gr.Slider(minimum=1, maximum=10, step=1, value=8, label="Inference Steps")
                guidance_scale_input = gr.Slider(
                    minimum=1.0, maximum=20.0, step=0.5, value=1.0, label="Guidance Scale", visible=False
                )

            generate_button = gr.Button("Generate Video", variant="primary")
        with gr.Column():
            video_output = gr.Video(label="Generated Video", autoplay=True, interactive=False)

    ui_inputs = [
        input_image_component,
        prompt_input,
        negative_prompt_input,
        duration_seconds_input,
        guidance_scale_input,
        steps_slider,
        seed_input,
        randomize_seed_checkbox,
    ]
    generate_button.click(fn=generate_video, inputs=ui_inputs, outputs=[video_output, seed_input])

    gr.Examples(
        examples=[
            ["peng.png", "a penguin playfully dancing in the snow, Antarctica"],
            ["forg.jpg", "the frog jumps around"],
        ],
        inputs=[input_image_component, prompt_input],
        outputs=[video_output, seed_input],
        fn=generate_video,
        cache_examples="lazy",
    )

if __name__ == "__main__":
    demo.queue().launch(mcp_server=True)