RAVE-main/scripts/run_experiment.py

import torch
import argparse
import os
import json
import sys
import datetime
import imageio  # Import imageio for MP4 saving
sys.path.append(os.getcwd())
from pipelines.sd_controlnet_rave import RAVE
from pipelines.sd_multicontrolnet_rave import RAVE_MultiControlNet
import utils.constants as const
import utils.video_grid_utils as vgu
import warnings
warnings.filterwarnings("ignore")
import numpy as np

def init_device():
    """Initialize the device (CUDA if available, else CPU)."""
    device_name = 'cuda' if torch.cuda.is_available() else 'cpu'
    device = torch.device(device_name)
    return device

def init_paths(input_ns, video_name, save_folder):
    """Initialize paths for video processing based on video name and save folder."""
    # Set save path directly to the video name (e.g., truck.mp4) under save_folder
    save_dir = save_folder
    os.makedirs(save_dir, exist_ok=True)
    input_ns.save_path = os.path.join(save_dir, video_name)  # Use video_name directly as filename

    # Set video path using the fixed base path and video name
    input_ns.video_path = f'/home/wangjuntong/video_editing_dataset/all_sourse/{video_name}'
    
    # Set Hugging Face ControlNet path based on preprocess_name
    if '-' in input_ns.preprocess_name:
        input_ns.hf_cn_path = [const.PREPROCESSOR_DICT[i] for i in input_ns.preprocess_name.split('-')]
    else:
        input_ns.hf_cn_path = const.PREPROCESSOR_DICT[input_ns.preprocess_name]
    input_ns.hf_path = "runwayml/stable-diffusion-v1-5"
    
    # Set inverse and control paths (though not used for saving)
    input_ns.inverse_path = f'{const.GENERATED_DATA_PATH}/inverses/{video_name}/{input_ns.preprocess_name}_{input_ns.model_id}_{input_ns.grid_size}x{input_ns.grid_size}_{input_ns.pad}'
    input_ns.control_path = f'{const.GENERATED_DATA_PATH}/controls/{video_name}/{input_ns.preprocess_name}_{input_ns.grid_size}x{input_ns.grid_size}_{input_ns.pad}'
    os.makedirs(input_ns.control_path, exist_ok=True)
    os.makedirs(input_ns.inverse_path, exist_ok=True)
    
    return input_ns

def run(input_ns, video_name, positive_prompts, save_folder):
    """Run the video editing process with the given parameters."""
    if 'model_id' not in input_ns.__dict__:
        input_ns.model_id = "None"
    device = init_device()
    input_ns = init_paths(input_ns, video_name, save_folder)

    
    print(f"Save path: {input_ns.save_path}")

    # Prepare video frames as a grid
    input_ns.image_pil_list = vgu.prepare_video_to_grid(input_ns.video_path, input_ns.sample_size, input_ns.grid_size, input_ns.pad)
    input_ns.sample_size = len(input_ns.image_pil_list)
    print(f'Frame count: {len(input_ns.image_pil_list)}')

    # Choose the appropriate ControlNet class
    controlnet_class = RAVE_MultiControlNet if '-' in str(input_ns.controlnet_conditioning_scale) else RAVE
    CN = controlnet_class(device)

    # Initialize models
    CN.init_models(input_ns.hf_cn_path, input_ns.hf_path, input_ns.preprocess_name, input_ns.model_id)
    
    input_dict = vars(input_ns)

    # Run the editing process
    start_time = datetime.datetime.now()
    if '-' in str(input_ns.controlnet_conditioning_scale):
        res_vid, control_vid_1, control_vid_2 = CN(input_dict)
    else: 
        res_vid, control_vid = CN(input_dict)
    end_time = datetime.datetime.now()

    # Convert PIL images to numpy arrays for imageio
    res_vid_np = [np.array(img) for img in res_vid]

    # Save the result video as MP4
    imageio.mimwrite(input_ns.save_path, res_vid_np, format='mp4', fps=30, quality=8) 

if __name__ == '__main__':
    # Parse command-line argument for JSONL file path
    parser = argparse.ArgumentParser(description='Batch video editing with JSONL input.')
    parser.add_argument('--jsonl_path', type=str, required=True, help='Path to the JSONL file containing video info')
    args = parser.parse_args()

    # Fixed parameters
    fixed_params = {
        'preprocess_name': 'depth_zoe',
        'batch_size': 4,
        'batch_size_vae': 1,
        'cond_step_start': 0.0,
        'controlnet_conditioning_scale': 1.0,
        'controlnet_guidance_end': 1.0,
        'controlnet_guidance_start': 0.0,
        'give_control_inversion': True,
        'grid_size': 3,
        'sample_size': -1,
        'pad': 1,
        'guidance_scale': 7.5,
        'inversion_prompt': '',
        'is_ddim_inversion': True,
        'is_shuffle': True,
        'negative_prompts': '',
        'num_inference_steps': 50,
        'num_inversion_step': 50,
        'seed': 0,
        'model_id': 'None'
    }

    # Read and process each line in the JSONL file
    with open(args.jsonl_path, 'r') as f:
        for line in f:
            data = json.loads(line)
            video_name = data['video']  # Use video key directly as filename (e.g., "truck.mp4")
            positive_prompts = data['edit_prompt']
            save_folder = f'/home/wangjuntong/RAVE-main/outputs/lnk_painting/{video_name.rsplit(".", 1)[0]}'  # Folder named after video without extension

            # Create input namespace with fixed and dynamic parameters
            input_ns = argparse.Namespace(**fixed_params)
            input_ns.positive_prompts = positive_prompts
            input_ns.video_name = video_name

            # Run the editing process
            run(input_ns, video_name, positive_prompts, save_folder)