app.py

import gradio as gr
import torch
import spaces
from diffusers import LTXConditionPipeline, LTXLatentUpsamplePipeline
from diffusers.utils import export_to_video
from PIL import Image, ImageOps
from gtts import gTTS
from pydub import AudioSegment
import whisper
import ffmpeg
import requests
from io import BytesIO
import os
import gc

# Load LTX models
ltx_model_id = "Lightricks/LTX-Video-0.9.7-distilled"
upscaler_model_id = "Lightricks/ltxv-spatial-upscaler-0.9.7"

pipe = LTXConditionPipeline.from_pretrained(ltx_model_id, torch_dtype=torch.float16)
pipe_upsample = LTXLatentUpsamplePipeline.from_pretrained(
    upscaler_model_id, vae=pipe.vae, torch_dtype=torch.float16
)

pipe.to("cuda")
pipe_upsample.to("cuda")
pipe.vae.enable_tiling()

def prepare_image_condition(image, size=(512, 512), background=(0, 0, 0)):
    image = ImageOps.contain(image, size)
    canvas = Image.new("RGB", size, background)
    offset = ((size[0] - image.width) // 2, (size[1] - image.height) // 2)
    canvas.paste(image, offset)
    return canvas

def round_to_nearest_resolution(height, width, ratio):
    return height - (height % ratio), width - (width % ratio)

@spaces.GPU(duration=180)
def generate_video(prompt, image_url):
    generator = torch.Generator("cuda").manual_seed(42)

    # Aspect-ratio preserving image prep
    image = None
    if image_url:
        raw_image = Image.open(BytesIO(requests.get(image_url).content)).convert("RGB")
        image = prepare_image_condition(raw_image)

    # Dimensions
    base_width, base_height = 512, 512
    downscale = 2 / 3
    w_d, h_d = round_to_nearest_resolution(int(base_width * downscale), int(base_height * downscale), pipe.vae_spatial_compression_ratio)

    # Step 1: Generate latents
    latents = pipe(
        prompt=prompt,
        image=image,
        width=w_d,
        height=h_d,
        num_frames=60,
        num_inference_steps=7,
        output_type="latent",
        guidance_scale=1.0,
        decode_timestep=0.05,
        decode_noise_scale=0.025,
        generator=generator
    ).frames

    torch.cuda.empty_cache()
    gc.collect()

    # Step 2: Upscale
    upscaled = pipe_upsample(latents=latents, output_type="latent").frames

    torch.cuda.empty_cache()
    gc.collect()

    # Step 3: Decode to frames
    frames = pipe(
        prompt=prompt,
        image=image,
        latents=upscaled,
        width=base_width,
        height=base_height,
        num_frames=60,
        num_inference_steps=10,
        output_type="pil",
        guidance_scale=1.0,
        decode_timestep=0.05,
        decode_noise_scale=0.025,
        image_cond_noise_scale=0.025,
        denoise_strength=0.3,
        generator=generator
    ).frames[0]

    # Step 4: Export video
    video_path = "output.mp4"
    export_to_video(frames, video_path, fps=24)

    # Step 5: TTS
    tts = gTTS(text=prompt, lang='en')
    tts.save("voice.mp3")
    AudioSegment.from_mp3("voice.mp3").export("voice.wav", format="wav")

    # Step 6: Subtitles (CPU)
    model = whisper.load_model("base", device="cpu")
    result = model.transcribe("voice.wav", task="transcribe", language="en")
    with open("subtitles.srt", "w", encoding="utf-8") as f:
        f.write(result["srt"])

    # Step 7: Merge video + audio + subtitles
    final_output = "final_with_audio.mp4"
    ffmpeg.input(video_path).output(
        final_output,
        vf="subtitles=subtitles.srt",
        i="voice.mp3",
        c="copy",
        shortest=None,
        loglevel="error"
    ).run()

    return final_output

# Gradio UI
demo = gr.Interface(
    fn=generate_video,
    inputs=[
        gr.Textbox(label="Prompt", placeholder="Describe your scene..."),
        gr.Textbox(label="Optional Image URL (e.g. Pexels)", placeholder="https://...")
    ],
    outputs=gr.Video(label="Generated Video"),
    title="🎬 LTX AI Video Generator",
    description="AI-powered video with voiceover and subtitles. Supports ZeroGPU (PyTorch) runtime."
)

demo.launch()