app.py

import os
import torch
import gradio as gr
import numpy as np
from PIL import Image
import imageio

from diffusers import StableVideoDiffusionPipeline

# =========================
# CONFIG
# =========================
MODEL_ID = "stabilityai/stable-video-diffusion-img2vid"
OUTPUT_DIR = "output"

NUM_FRAMES = 14          # AMAN untuk CPU
FPS = 7
SEED = 42

os.makedirs(OUTPUT_DIR, exist_ok=True)

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32

# =========================
# LOAD MODEL (SAFE)
# =========================
pipe = StableVideoDiffusionPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=DTYPE
)

pipe.to(DEVICE)
pipe.enable_attention_slicing()

if DEVICE == "cuda":
    pipe.enable_model_cpu_offload()

generator = torch.Generator(device=DEVICE).manual_seed(SEED)

# =========================
# SAVE VIDEO (NO CV2)
# =========================
def save_video(frames, path, fps):
    imageio.mimsave(path, frames, fps=fps)

# =========================
# IMAGE → VIDEO FUNCTION
# =========================
def images_to_videos(files):
    if files is None or len(files) == 0:
        return None

    output_videos = []

    for idx, file in enumerate(files):
        image = Image.open(file.name).convert("RGB")
        image = image.resize((768, 432))  # CPU SAFE SIZE

        with torch.autocast(DEVICE if DEVICE == "cuda" else "cpu"):
            result = pipe(
                image=image,
                num_frames=NUM_FRAMES,
                generator=generator
            )

        frames = [
            (frame * 255).astype(np.uint8)
            for frame in result.frames[0]
        ]

        out_path = os.path.join(OUTPUT_DIR, f"video_{idx}.mp4")
        save_video(frames, out_path, FPS)

        output_videos.append(out_path)

    return output_videos

# =========================
# GRADIO UI
# =========================
with gr.Blocks(title="Image to Video (HF Spaces Safe)") as demo:
    gr.Markdown(
        """
        ## 🎥 Image → Video AI (Hugging Face Spaces)
        - Upload **multiple images**
        - Output **MP4**
        - CPU & GPU compatible
        """
    )

    image_input = gr.File(
        file_types=["image"],
        file_count="multiple",
        label="Upload Images"
    )

    video_output = gr.Video(label="Generated Video")

    generate_btn = gr.Button("🚀 Generate")

    def process(files):
        videos = images_to_videos(files)
        return videos[0] if videos else None

    generate_btn.click(
        fn=process,
        inputs=image_input,
        outputs=video_output
    )

demo.launch()