app.py

import os
import gradio as gr
import torch
from diffusers import DiffusionPipeline

# =====================================================
# Pilih model utama (Wan 2.2) - fallback ke damo-vilab
# =====================================================
WAN_MODEL = "BAAI/Wan2.2-videogen"   # ⚠ ganti nama sesuai yg VALID di HF Hub
FALLBACK_MODEL = "damo-vilab/text-to-video-ms-1.7b"

# Ambil token kalau model private/gated
hf_token = os.environ.get("HF_TOKEN")

def load_model(model_id):
    try:
        pipe = DiffusionPipeline.from_pretrained(
            model_id,
            torch_dtype=torch.float16,
            variant="fp16",
            use_auth_token=hf_token
        ).to("cuda")
        print(f"✅ Loaded model: {model_id}")
        return pipe
    except Exception as e:
        print(f"⚠ Error loading {model_id}, fallback ke {FALLBACK_MODEL}. Error: {e}")
        pipe = DiffusionPipeline.from_pretrained(
            FALLBACK_MODEL,
            torch_dtype=torch.float16,
            variant="fp16"
        ).to("cuda")
        return pipe

# Load model saat startup
pipe = load_model(WAN_MODEL)

# =====================================================
# Fungsi generate video
# =====================================================
def generate_video(prompt, num_frames=16, fps=8, seed=42, progress=gr.Progress(track_tqdm=True)):
    generator = torch.manual_seed(seed)
    
    progress(0, desc="🚀 Mulai generate video...")

    output = pipe(
        prompt=prompt,
        num_frames=num_frames,
        generator=generator
    )

    progress(0.7, desc="📸 Menyusun frame jadi video...")

    video_frames = output.frames[0]
    out_path = "output.mp4"

    # Simpan video ke file .mp4
    pipe.save_pretrained_video(video_frames, out_path, fps=fps)

    progress(1, desc="✅ Selesai!")
    return out_path, out_path   # untuk Video preview + Download file

# =====================================================
# Gradio UI
# =====================================================
with gr.Blocks() as demo:
    gr.Markdown("## 🎬 WAN 2.2 Video Generator (Hugging Face Space)")

    with gr.Row():
        prompt_inp = gr.Textbox(
            label="Prompt",
            placeholder="Masukkan deskripsi video...",
            value="Seekor naga robot terbang melintasi kota futuristik"
        )

    with gr.Row():
        frame_slider = gr.Slider(8, 64, step=8, value=16, label="Jumlah Frame")
        fps_slider = gr.Slider(4, 16, step=1, value=8, label="FPS (kecepatan video)")

    btn = gr.Button("🚀 Generate Video")

    with gr.Row():
        video_out = gr.Video(label="Hasil")
        download_link = gr.File(label="Unduh Video", type="file")

    btn.click(
        fn=generate_video,
        inputs=[prompt_inp, frame_slider, fps_slider],
        outputs=[video_out, download_link]
    )

if __name__ == "__main__":
    demo.queue(max_size=5, concurrency_count=1).launch()