import spaces # Import the spaces module for ZeroGPU import gradio as gr import torch import os import tempfile import shutil import time import ffmpeg import numpy as np from PIL import Image import moviepy.editor as mp from infer import lotus, load_models, pipe_g, pipe_d # Import the global models import logging # Set up logging logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) # Device will be set inside GPU-decorated functions device = 'cuda' # Use 'cuda' as placeholder # Load models once inside a GPU context task_name = 'depth' @spaces.GPU def initialize_models(): load_models(task_name, device) # Call the function to load models initialize_models() # Preprocess the video to adjust frame rate def preprocess_video(video_path, target_fps=24): """Preprocess the video to adjust its frame rate.""" video = mp.VideoFileClip(video_path) # Adjust FPS if target_fps is specified if target_fps > 0: video = video.set_fps(target_fps) return video # Resize image while preserving aspect ratio and adding padding def resize_and_pad(image, target_size): """Resize and pad an image to the target size while preserving aspect ratio.""" # Calculate the new size preserving aspect ratio image.thumbnail(target_size, Image.LANCZOS) # Create a new image with the target size and black background new_image = Image.new("RGB", target_size) new_image.paste( image, ((target_size[0] - image.width) // 2, (target_size[1] - image.height) // 2) ) return new_image @spaces.GPU def process_frame(frame, seed=0, target_size=(512, 512)): """Process a single frame and return depth map.""" try: # Convert frame to PIL Image image = Image.fromarray(frame).convert('RGB') # Resize and pad image input_image = resize_and_pad(image, target_size) # Run inference depth_map = lotus(input_image, 'depth', seed, device) # Crop the output depth map back to original image size width, height = image.size left = (target_size[0] - width) // 2 top = (target_size[1] - height) // 2 right = left + width bottom = top + height depth_map_cropped = depth_map.crop((left, top, right, bottom)) return depth_map_cropped except Exception as e: logger.error(f"Error processing frame: {e}") return None def process_video(video_path, fps=0, seed=0): """Process video frames individually and generate depth maps.""" # Create a persistent temporary directory temp_dir = tempfile.mkdtemp() try: start_time = time.time() # Preprocess the video video = preprocess_video(video_path, target_fps=fps) # Use original video FPS if not specified if fps == 0: fps = video.fps frames = list(video.iter_frames()) total_frames = len(frames) logger.info(f"Processing {total_frames} frames at {fps} FPS...") # Create directory for frame sequence and outputs frames_dir = os.path.join(temp_dir, "frames") os.makedirs(frames_dir, exist_ok=True) # Process frames individually for i, frame in enumerate(frames): # Process each frame with GPU allocation depth_map = process_frame(frame, seed) if depth_map is not None: # Save frame frame_path = os.path.join(frames_dir, f"frame_{i:06d}.png") depth_map.save(frame_path, format='PNG', compress_level=0) # Update live preview every 10% progress if i % max(1, total_frames // 10) == 0: elapsed_time = time.time() - start_time progress = (i / total_frames) * 100 yield depth_map, None, None, f"Processed {i}/{total_frames} frames... ({progress:.2f}%) Elapsed: {elapsed_time:.2f}s" else: logger.error(f"Error processing frame {i}") logger.info("Creating output files...") # Create ZIP of frame sequence zip_filename = f"depth_frames_{int(time.time())}.zip" zip_path = os.path.join(temp_dir, zip_filename) shutil.make_archive(zip_path[:-4], 'zip', frames_dir) # Create MP4 video video_filename = f"depth_video_{int(time.time())}.mp4" output_video_path = os.path.join(temp_dir, video_filename) try: # FFmpeg settings for high-quality MP4 input_pattern = os.path.join(frames_dir, 'frame_%06d.png') ( ffmpeg .input(input_pattern, pattern_type='sequence', framerate=fps) .output(output_video_path, vcodec='libx264', pix_fmt='yuv420p', crf=17, preset='slow') .run(overwrite_output=True, quiet=True) ) logger.info("MP4 video created successfully!") except ffmpeg.Error as e: logger.error(f"Error creating video: {e.stderr.decode() if e.stderr else str(e)}") output_video_path = None total_time = time.time() - start_time logger.info("Processing complete!") # Yield the file paths yield None, zip_path, output_video_path, f"Processing complete! Total time: {total_time:.2f} seconds" except Exception as e: logger.error(f"Error: {e}") yield None, None, None, f"Error processing video: {e}" finally: # Clean up temporary directory if necessary pass def process_wrapper(video, fps=0, seed=0): if video is None: raise gr.Error("Please upload a video.") try: outputs = [] # Initialize models within the Gradio request context initialize_models() # Use video directly, since it's the file path for output in process_video(video, fps, seed): outputs.append(output) yield output return outputs[-1] except Exception as e: raise gr.Error(f"Error processing video: {str(e)}") # Custom CSS for styling (unchanged) custom_css = """ .title-container { text-align: center; padding: 10px 0; } #title { font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif; font-size: 36px; font-weight: bold; color: #000000; padding: 10px; border-radius: 10px; display: inline-block; background: linear-gradient( 135deg, #e0f7fa, #e8f5e9, #fff9c4, #ffebee, #f3e5f5, #e1f5fe, #fff3e0, #e8eaf6 ); background-size: 400% 400%; animation: gradient-animation 15s ease infinite; } @keyframes gradient-animation { 0% { background-position: 0% 50%; } 50% { background-position: 100% 50%; } 100% { background-position: 0% 50%; } } """ # Gradio Interface with gr.Blocks(css=custom_css) as demo: gr.HTML('''
Video Depth Estimation
''') with gr.Row(): with gr.Column(): video_input = gr.Video(label="Upload Video", interactive=True) fps_slider = gr.Slider(minimum=0, maximum=60, step=1, value=0, label="Output FPS (0 for original)") seed_slider = gr.Number(value=0, label="Seed") btn = gr.Button("Process Video") with gr.Column(): preview_image = gr.Image(label="Live Preview") output_frames_zip = gr.File(label="Download Frame Sequence (ZIP)") output_video = gr.File(label="Download Video (MP4)") time_textbox = gr.Textbox(label="Status", interactive=False) btn.click( fn=process_wrapper, inputs=[video_input, fps_slider, seed_slider], outputs=[preview_image, output_frames_zip, output_video, time_textbox] ) demo.queue() if __name__ == "__main__": demo.launch(debug=True)