import requests
import base64

# Important: the NVIDIA L40S will only support small resolutions, short length and no post-processing.
# If you want those features, you might need to use the NVIDIA A100.
# Use your own Inference Endpoint URL
API_URL = "https://<use your own Inference Endpoint here>.endpoints.huggingface.cloud"

# Use you own API token
API_TOKEN = "hf_<replace by your own Hugging Face token>"

def query(payload):
    response = requests.post(API_URL, headers={
        "Accept": "application/json",
        "Authorization": f"Bearer {API_TOKEN}",
        "Content-Type": "application/json" 
    }, json=payload)
    return response.json()

def save_video(json_response):
    video_data_uri = ""
    try:
        # Extract the video data URI from the response
        video_data_uri = json_response["video"]
    except Exception as e:
        message = str(json_response)
        print(message)
        raise ValueError(message)
    
    # Remove the data URI prefix to get just the base64 data
    # Assumes format like "data:video/mp4;base64,<actual_base64_data>"
    base64_data = video_data_uri.split(",")[1]
    
    # Decode the base64 data
    video_data = base64.b64decode(base64_data)
    
    # Write the binary data to an MP4 file
    with open("video.mp4", "wb") as f:
        f.write(video_data)

# Make the API call
output = query({
    "inputs": {
        "prompt": "Portrait photo, selfie of a beautiful young caucasian woman called Charlotte, wearing a pastel-blue hoodie. She is livestreaming from NYC streets. She looks straight into the camera, looking serious, and she talks. The camera is fixed, static, a medium-shot centered on her face. 4K webcam footage. Intricate details, super resolution, sharp image, award winning."
    },
    "parameters": {
        # ------------------- settings for LTX-Video -----------------------
        # for a vertical video look
        "width": 480,
        "height": 768,

        # LTX-Video requires a frame number divisible by 8, plus one frame
        # note: glitches might appear if you use more than 168 frames
        "num_frames": (8 * 14) + 1,

        # using 30 steps seems to be enough for most cases, otherwise use 50 for best quality
        # I think using a large number of steps (> 30) might create some overexposure and saturation
        "num_inference_steps": 40,

        # values between 3.0 and 4.0 are nice
        "guidance_scale": 3.5,

        # seed: -1,

        # ------------------- settings for Varnish -----------------------
        # This will double the number of frames.
        # You can activate this if you want:
        # - a slow motion effect (in that case use double_num_frames=True and fps=24, 25 or 30)
        # - a HD soap / video game effect (in that case use double_num_frames=True and fps=60)
        "double_num_frames": True,

        # controls the number of frames per second
        # use this in combination with the num_frames and double_num_frames settings to control the duration and "feel" of your video
        "fps": 60, # typical values are: 24, 25, 30, 60

        # upscale the video using Real-ESRGAN.
        # This upscaling algorithm is relatively fast,
        # but might create an uncanny "3D render" or "drawing" effect.
        "super_resolution": True,

        # for cosmetic purposes and get a "cinematic" feel, you can optionally add some film grain.
        # it is not recommended to add film grain if your theme doesn't match (film grain is great for black & white, retro looks)
        # and if you do, adding more than 12% will start to negatively impact file size (video codecs aren't great are compressing film grain)
        # 0% = no grain
        # 10% = a bit of grain
        "grain_amount": 10, # value between 0-100

        # the following parameters are a work in progress
        "enable_audio": False,
        #"audio_prompt": "voices, voice, talking, speaking, speech",
        #"audio_negative_prompt": "",

    }
})

# Save the video
save_video(output)