Haiyang Liu

Upload visualization/visualize.py with huggingface_hub

00eb0db verified 7 months ago

17.8 kB

import os
from pathlib import Path
from typing import List

import numpy as np
from matplotlib import font_manager
from PIL import Image, ImageDraw, ImageFont
from tqdm import tqdm

import imageio

from visualization.HumanML3D263.render import render_frames as render_frames_263
from visualization.MotionStreamer272.render import render_frames as render_frames_272
from visualization.MEI138.render import render_frames as render_frames_138

# Unified tint colors for multi-segment visualization (video frames and text bars)
SEGMENT_TINT_COLORS = [
    [255, 220, 220],  # Light Red (video)
    [220, 255, 220],  # Light Green
    [220, 220, 255],  # Light Blue
]
SEGMENT_TEXT_COLORS = [
    (255, 180, 180),  # Reddish (text caption)
    (180, 255, 180),  # Greenish
    (180, 180, 255),  # Blueish
]

def _apply_tint(images, segments):
    """Apply segment-based tint coloring to a list of images.
    Args:
        images: list of np.ndarray (H, W, 3) uint8.
        segments: np.ndarray of frame boundaries for text segments.
    Returns:
        list of tinted images.
    """
    for i, img in enumerate(images):
        segment_idx = np.searchsorted(segments, i, side="right")
        tint_factor = np.array(SEGMENT_TINT_COLORS[segment_idx % len(SEGMENT_TINT_COLORS)]) / 255.0
        images[i] = (img.astype(np.float32) * tint_factor).clip(0, 255).astype(np.uint8)
    return images


def render_single_video(
    motion: np.ndarray,
    save_path: str,
    dim: int,
    frames: np.ndarray = None,
):
    # Render to image list
    if dim == 263:
        images = render_frames_263(motion)
        fps = 20
    elif dim == 272:
        images = render_frames_272(motion)
        fps = 30
    elif dim == 138:
        images = render_frames_138(motion)
        fps = 30
    else:
        raise ValueError(f"Unsupported motion dimension: {dim}. Supported: [263, 272, 138]")

    # Apply segment tint coloring
    if frames is not None and len(frames) > 1:
        images = _apply_tint(images, frames)

    # Save to video
    writer = imageio.get_writer(save_path, fps=fps)
    for img in images:
        writer.append_data(img)
    writer.close()


def render_video(
    motion_dir: str,
    save_dir: str,
    render_setting,
    frames_dir: str = None,
):
    os.makedirs(save_dir, exist_ok=True)
    motion_path = Path(motion_dir)
    npy_files = list(motion_path.glob("*.npy"))

    motion_count = 0
    error_count = 0

    for npy_file in tqdm(npy_files, desc="Rendering"):
        motion_data = np.load(npy_file)
        output_filename = npy_file.stem + ".mp4"
        output_path = os.path.join(save_dir, output_filename)

        if frames_dir is not None and os.path.exists(frames_dir):
            frames_path = os.path.join(frames_dir, npy_file.name)
            frames = np.load(frames_path)
        else:
            frames = None

        try:
            render_single_video(
                motion=motion_data,
                save_path=output_path,
                dim=render_setting["recover_dim"],
                frames=frames,
            )
        except Exception as e:
            print(f"Error rendering {npy_file}: {e}")
            error_count += 1
            continue
        motion_count += 1

    print(
        f"{motion_count} motion clips rendered. {error_count} errors. Saved to {save_dir}"
    )


def render_text_bar(
    text, width, padding=20, font_size=28, bg_color=(0, 0, 0), fg_color=(255, 255, 255)
):
    """Renders a text bar with automatic wrapping using PIL, returns np.uint8(H,W,3)."""
    try:
        font_path = font_manager.findfont("DejaVu Sans")
    except Exception:
        font_path = font_manager.findfont("Arial")
    font = ImageFont.truetype(font_path, font_size)

    # Split text by separator if present
    segments = []
    if "//////////" in text:
        parts = text.split("//////////")
        for part in parts:
            if part.strip():
                segments.append(part.strip())
    else:
        segments.append(text)

    dummy = ImageDraw.Draw(Image.new("RGB", (10, 10)))
    max_w = width - 2 * padding

    all_lines = []  # List of (text_content, color) tuples

    for i, segment in enumerate(segments):
        color = fg_color
        if len(segments) > 1:
            color = SEGMENT_TEXT_COLORS[i % len(SEGMENT_TEXT_COLORS)]

        # Wrap text for this segment
        cur = ""
        first_word = True
        for w in segment.split():
            test = (cur + " " + w).strip() if cur else w
            if dummy.textlength(test, font=font) <= max_w:
                cur = test
            else:
                all_lines.append((cur, color))
                cur = w
        if cur:
            all_lines.append((cur, color))

    _, top, _, bottom = font.getbbox("A")
    line_h = bottom - top + 4
    bar_h = padding * 2 + line_h * len(all_lines)

    # Ensure height is even for H.264 encoding
    if bar_h % 2 != 0:
        bar_h += 1

    img = Image.new("RGB", (width, bar_h), bg_color)
    draw = ImageDraw.Draw(img)
    y = padding
    for line_text, line_color in all_lines:
        draw.text((padding, y), line_text, font=font, fill=line_color)
        y += line_h
    return np.array(img)


def render_aligned_title_bar(
    total_width,
    widths,
    titles,
    font_size=32,
    bg_color=(255, 255, 255),
    fg_color=(0, 0, 0),
    padding=10,
):
    """Renders a title bar with centered titles aligned to video sections."""
    try:
        font_path = font_manager.findfont("DejaVu Sans")
    except Exception:
        font_path = font_manager.findfont("Arial")
    font = ImageFont.truetype(font_path, font_size)

    # Calculate title bar height
    _, top, _, bottom = font.getbbox("A")
    bar_height = bottom - top + 2 * padding

    # Ensure height is even for H.264 encoding
    if bar_height % 2 != 0:
        bar_height += 1

    # Create image
    img = Image.new("RGB", (total_width, bar_height), bg_color)
    draw = ImageDraw.Draw(img)

    # Calculate positions and draw titles
    x_offset = 0
    for i, (title, width) in enumerate(zip(titles, widths)):
        # Calculate center position for this section
        text_width = draw.textlength(title, font=font)
        x_center = x_offset + width // 2
        x_pos = x_center - text_width // 2
        y_pos = padding

        draw.text((x_pos, y_pos), title, font=font, fill=fg_color)
        x_offset += width

    return np.array(img)


def _get_video_info(video_path: str):
    """Get video width, height, and duration using ffprobe."""
    import subprocess

    cmd = [
        "ffprobe",
        "-v",
        "error",
        "-select_streams",
        "v:0",
        "-show_entries",
        "stream=width,height,duration",
        "-of",
        "csv=p=0",
        video_path,
    ]
    output = subprocess.check_output(cmd, text=True).strip().split(",")
    return int(output[0]), int(output[1]), float(output[2])


def _get_fps(video_path: str):
    """Get video frame rate using ffprobe."""
    import subprocess

    cmd = [
        "ffprobe",
        "-v",
        "error",
        "-select_streams",
        "v:0",
        "-show_entries",
        "stream=r_frame_rate",
        "-of",
        "csv=p=0",
        video_path,
    ]
    fps_str = subprocess.check_output(cmd, text=True).strip()
    num, den = map(int, fps_str.split("/"))
    return num / den


def _build_video_filter(
    input_idx: int,
    video_idx: int,
    target_height: int,
    duration: float,
    max_duration: float,
    target_width: int = -2,
):
    """Build ffmpeg filter for a single video stream."""
    filters = []

    # Scale to target height, ensure even dimensions
    # filters.append(f"[{input_idx}:v]scale=-2:{target_height}[v{video_idx}_scaled]")
    filters.append(
        f"[{input_idx}:v]scale={target_width}:{target_height}[v{video_idx}_scaled]"
    )

    # Handle duration and graying
    if duration < max_duration:
        pad_duration = max_duration - duration
        filters.append(
            f"[v{video_idx}_scaled]tpad=stop_mode=clone:stop_duration={pad_duration}[v{video_idx}_padded]"
        )
        # Apply gray effect after original duration using eq filter with enable
        filters.append(
            f"[v{video_idx}_padded]eq=brightness=-0.5:saturation=0:enable='gte(t,{duration})'[v{video_idx}]"
        )
    else:
        # Video is already long enough, just use scaled version
        filters.append(f"[v{video_idx}_scaled]null[v{video_idx}]")

    return filters


def make_composite_compare_videos(
    result_folder: str,
    save_dir: str,
    text_folder: str = None,
    compare_folders: list = None,
    compare_names: list = None,
):
    """Generates composite videos of (result | compare_folders) with captions and text descriptions.
    Args:
        result_folder: Folder containing result videos (base for comparison)
        save_dir: Directory to save composite videos
        text_folder: Folder containing text descriptions (optional)
        compare_folders: List of folders to compare with result (optional)
        compare_names: List of names for compare folders (optional)
    Uses the longest video duration. Missing videos show black screen.
    Videos that end early show their last frame grayed out.
    Optimized version using ffmpeg directly for much faster processing.
    """
    import subprocess
    import tempfile

    os.makedirs(save_dir, exist_ok=True)
    video_files = list(Path(result_folder).glob("*.mp4"))

    # Handle empty or non-existent compare folders
    if compare_folders is None:
        compare_folders = []
    if compare_names is None:
        compare_names = []

    # Filter out non-existent compare folders
    valid_compare_folders = []
    valid_compare_names = []
    for i, folder in enumerate(compare_folders):
        if folder and os.path.exists(folder):
            valid_compare_folders.append(folder)
            if i < len(compare_names):
                valid_compare_names.append(compare_names[i])
            else:
                valid_compare_names.append(f"Compare {i + 1}")

    compare_folders = valid_compare_folders
    compare_names = valid_compare_names

    for video_file in tqdm(video_files, desc="Creating composite videos"):
        video_id = video_file.stem

        # Prepare video paths - start with result, then add compare folders
        video_paths = [str(video_file)]
        video_names = ["Ours"]

        # Add compare folder videos
        for folder, name in zip(compare_folders, compare_names):
            compare_path = os.path.join(folder, f"{video_id}.mp4")
            video_paths.append(compare_path)
            video_names.append(name)

        video_exists = [os.path.exists(p) for p in video_paths]

        # Load text description
        if text_folder:
            text_file = os.path.join(text_folder, f"{video_id}.txt")
            text_content = (
                Path(text_file).read_text().strip()
                if os.path.exists(text_file)
                else f"Motion: {video_id}"
            )
        else:
            text_content = f"Motion: {video_id}"

        # Find reference video for properties (should always have result video)
        reference_video = str(video_file)
        if not os.path.exists(reference_video):
            print(f"Error: Result video not found for {video_id}, skipping")
            continue

        # Get video properties
        try:
            fps = _get_fps(reference_video)
            ref_width, ref_height, _ = _get_video_info(reference_video)
        except Exception as e:
            print(f"Error probing {video_id}: {e}, skipping")
            continue

        # Collect dimensions and durations for all videos
        widths, heights, durations = [], [], []
        for path, exists, name in zip(video_paths, video_exists, video_names):
            if exists:
                try:
                    w, h, d = _get_video_info(path)
                    widths.append(w)
                    heights.append(h)
                    durations.append(d)
                except Exception as e:
                    print(
                        f"Error probing {name} video for {video_id}: {e}, will use black screen"
                    )
                    widths.append(ref_width)
                    heights.append(ref_height)
                    durations.append(0)
            else:
                print(
                    f"Warning: {name} video missing for {video_id}, will use black screen"
                )
                widths.append(ref_width)
                heights.append(ref_height)
                durations.append(0)

        max_duration = max(durations)
        if max_duration == 0:
            print(f"Warning: All videos for {video_id} have zero duration, skipping")
            continue

        target_height = min(h for h in heights if h > 0)
        # Ensure target height is even
        if target_height % 2 != 0:
            target_height += 1

        # Re-calculate widths based on target_height scaling
        new_widths = []
        for w, h in zip(widths, heights):
            if h > 0:
                # Calculate scaled width maintaining aspect ratio
                aspect_ratio = w / h
                scaled_w = int(target_height * aspect_ratio)
                # Ensure even width
                if scaled_w % 2 != 0:
                    scaled_w += 1
                new_widths.append(scaled_w)
            else:
                new_widths.append(w)

        widths = new_widths
        total_width = sum(widths)

        # print(f"DEBUG: video_id={video_id}")
        # print(f"DEBUG: original widths={widths} (after update), heights={heights}")
        # print(f"DEBUG: target_height={target_height}")
        # print(f"DEBUG: total_width={total_width}")

        # Create and save title/text bars
        title_bar = render_aligned_title_bar(
            total_width, widths, video_names, font_size=32
        )
        text_bar = render_text_bar(text_content, width=total_width, font_size=24)

        with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as f:
            title_path = f.name
            Image.fromarray(title_bar).save(title_path)

        with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as f:
            text_path = f.name
            Image.fromarray(text_bar).save(text_path)

        # Build ffmpeg command
        inputs = []
        filter_parts = []
        input_count = 0

        # Add video inputs and filters
        for i, (path, exists, width, duration) in enumerate(
            zip(video_paths, video_exists, widths, durations)
        ):
            if exists:
                inputs.extend(["-i", path])
                filter_parts.extend(
                    _build_video_filter(
                        input_count,
                        i,
                        target_height,
                        duration,
                        max_duration,
                        target_width=width,
                    )
                )
            else:
                inputs.extend(
                    [
                        "-f",
                        "lavfi",
                        "-i",
                        f"color=c=black:s={width}x{target_height}:d={max_duration}:r={int(fps)}",
                    ]
                )
                filter_parts.append(f"[{input_count}:v]null[v{i}]")
            input_count += 1

        # Add title and text images
        inputs.extend(["-loop", "1", "-i", title_path])
        title_idx = input_count
        input_count += 1

        inputs.extend(["-loop", "1", "-i", text_path])
        text_idx = input_count

        # Compose final layout - horizontally stack all videos
        num_videos = len(video_paths)
        if num_videos == 1:
            filter_parts.append("[v0]null[videos]")
        else:
            video_inputs = "".join([f"[v{i}]" for i in range(num_videos)])
            filter_parts.append(f"{video_inputs}hstack=inputs={num_videos}[videos]")

        filter_parts.append(
            f"[{title_idx}:v][videos][{text_idx}:v]vstack=inputs=3[out]"
        )

        # Execute ffmpeg
        output_path = os.path.join(save_dir, f"{video_id}_composite.mp4")
        cmd = [
            "ffmpeg",
            "-y",
            *inputs,
            "-filter_complex",
            ";".join(filter_parts),
            "-map",
            "[out]",
            "-t",
            str(max_duration),
            "-r",
            str(int(fps)),
            "-c:v",
            "libx264",
            "-preset",
            "ultrafast",
            "-crf",
            "23",
            "-pix_fmt",
            "yuv420p",
            "-profile:v",
            "baseline",
            "-level",
            "3.0",
            "-movflags",
            "+faststart",
            output_path,
        ]

        try:
            result = subprocess.run(cmd, check=False, capture_output=True, text=True)
            if result.returncode != 0:
                print(f"Error processing {video_id}: Return code {result.returncode}")
                print(f"Command: {' '.join(cmd)}")
                print(f"Stderr: {result.stderr}")
            elif os.path.exists(output_path) and os.path.getsize(output_path) == 0:
                print(f"Warning: Generated video {output_path} is empty!")
                print(f"Command: {' '.join(cmd)}")
                print(f"Stderr: {result.stderr}")
        except Exception as e:
            print(f"Unexpected error processing {video_id}: {e}")
        finally:
            for path in [title_path, text_path]:
                try:
                    os.unlink(path)
                except Exception:
                    pass

    print(f"Composite videos saved to {save_dir}")

	import os
	from pathlib import Path
	from typing import List

	import numpy as np
	from matplotlib import font_manager
	from PIL import Image, ImageDraw, ImageFont
	from tqdm import tqdm

	import imageio

	from visualization.HumanML3D263.render import render_frames as render_frames_263
	from visualization.MotionStreamer272.render import render_frames as render_frames_272
	from visualization.MEI138.render import render_frames as render_frames_138

	# Unified tint colors for multi-segment visualization (video frames and text bars)
	SEGMENT_TINT_COLORS = [
	[255, 220, 220], # Light Red (video)
	[220, 255, 220], # Light Green
	[220, 220, 255], # Light Blue
	]
	SEGMENT_TEXT_COLORS = [
	(255, 180, 180), # Reddish (text caption)
	(180, 255, 180), # Greenish
	(180, 180, 255), # Blueish
	]

	def _apply_tint(images, segments):
	"""Apply segment-based tint coloring to a list of images.

	Args:
	images: list of np.ndarray (H, W, 3) uint8.
	segments: np.ndarray of frame boundaries for text segments.

	Returns:
	list of tinted images.
	"""
	for i, img in enumerate(images):
	segment_idx = np.searchsorted(segments, i, side="right")
	tint_factor = np.array(SEGMENT_TINT_COLORS[segment_idx % len(SEGMENT_TINT_COLORS)]) / 255.0
	images[i] = (img.astype(np.float32) * tint_factor).clip(0, 255).astype(np.uint8)
	return images


	def render_single_video(
	motion: np.ndarray,
	save_path: str,
	dim: int,
	frames: np.ndarray = None,
	):
	# Render to image list
	if dim == 263:
	images = render_frames_263(motion)
	fps = 20
	elif dim == 272:
	images = render_frames_272(motion)
	fps = 30
	elif dim == 138:
	images = render_frames_138(motion)
	fps = 30
	else:
	raise ValueError(f"Unsupported motion dimension: {dim}. Supported: [263, 272, 138]")

	# Apply segment tint coloring
	if frames is not None and len(frames) > 1:
	images = _apply_tint(images, frames)

	# Save to video
	writer = imageio.get_writer(save_path, fps=fps)
	for img in images:
	writer.append_data(img)
	writer.close()


	def render_video(
	motion_dir: str,
	save_dir: str,
	render_setting,
	frames_dir: str = None,
	):
	os.makedirs(save_dir, exist_ok=True)
	motion_path = Path(motion_dir)
	npy_files = list(motion_path.glob("*.npy"))

	motion_count = 0
	error_count = 0

	for npy_file in tqdm(npy_files, desc="Rendering"):
	motion_data = np.load(npy_file)
	output_filename = npy_file.stem + ".mp4"
	output_path = os.path.join(save_dir, output_filename)

	if frames_dir is not None and os.path.exists(frames_dir):
	frames_path = os.path.join(frames_dir, npy_file.name)
	frames = np.load(frames_path)
	else:
	frames = None

	try:
	render_single_video(
	motion=motion_data,
	save_path=output_path,
	dim=render_setting["recover_dim"],
	frames=frames,
	)
	except Exception as e:
	print(f"Error rendering {npy_file}: {e}")
	error_count += 1
	continue
	motion_count += 1

	print(
	f"{motion_count} motion clips rendered. {error_count} errors. Saved to {save_dir}"
	)


	def render_text_bar(
	text, width, padding=20, font_size=28, bg_color=(0, 0, 0), fg_color=(255, 255, 255)
	):
	"""Renders a text bar with automatic wrapping using PIL, returns np.uint8(H,W,3)."""
	try:
	font_path = font_manager.findfont("DejaVu Sans")
	except Exception:
	font_path = font_manager.findfont("Arial")
	font = ImageFont.truetype(font_path, font_size)

	# Split text by separator if present
	segments = []
	if "//////////" in text:
	parts = text.split("//////////")
	for part in parts:
	if part.strip():
	segments.append(part.strip())
	else:
	segments.append(text)

	dummy = ImageDraw.Draw(Image.new("RGB", (10, 10)))
	max_w = width - 2 * padding

	all_lines = [] # List of (text_content, color) tuples

	for i, segment in enumerate(segments):
	color = fg_color
	if len(segments) > 1:
	color = SEGMENT_TEXT_COLORS[i % len(SEGMENT_TEXT_COLORS)]

	# Wrap text for this segment
	cur = ""
	first_word = True
	for w in segment.split():
	test = (cur + " " + w).strip() if cur else w
	if dummy.textlength(test, font=font) <= max_w:
	cur = test
	else:
	all_lines.append((cur, color))
	cur = w
	if cur:
	all_lines.append((cur, color))

	_, top, _, bottom = font.getbbox("A")
	line_h = bottom - top + 4
	bar_h = padding * 2 + line_h * len(all_lines)

	# Ensure height is even for H.264 encoding
	if bar_h % 2 != 0:
	bar_h += 1

	img = Image.new("RGB", (width, bar_h), bg_color)
	draw = ImageDraw.Draw(img)
	y = padding
	for line_text, line_color in all_lines:
	draw.text((padding, y), line_text, font=font, fill=line_color)
	y += line_h
	return np.array(img)


	def render_aligned_title_bar(
	total_width,
	widths,
	titles,
	font_size=32,
	bg_color=(255, 255, 255),
	fg_color=(0, 0, 0),
	padding=10,
	):
	"""Renders a title bar with centered titles aligned to video sections."""
	try:
	font_path = font_manager.findfont("DejaVu Sans")
	except Exception:
	font_path = font_manager.findfont("Arial")
	font = ImageFont.truetype(font_path, font_size)

	# Calculate title bar height
	_, top, _, bottom = font.getbbox("A")
	bar_height = bottom - top + 2 * padding

	# Ensure height is even for H.264 encoding
	if bar_height % 2 != 0:
	bar_height += 1

	# Create image
	img = Image.new("RGB", (total_width, bar_height), bg_color)
	draw = ImageDraw.Draw(img)

	# Calculate positions and draw titles
	x_offset = 0
	for i, (title, width) in enumerate(zip(titles, widths)):
	# Calculate center position for this section
	text_width = draw.textlength(title, font=font)
	x_center = x_offset + width // 2
	x_pos = x_center - text_width // 2
	y_pos = padding

	draw.text((x_pos, y_pos), title, font=font, fill=fg_color)
	x_offset += width

	return np.array(img)


	def _get_video_info(video_path: str):
	"""Get video width, height, and duration using ffprobe."""
	import subprocess

	cmd = [
	"ffprobe",
	"-v",
	"error",
	"-select_streams",
	"v:0",
	"-show_entries",
	"stream=width,height,duration",
	"-of",
	"csv=p=0",
	video_path,
	]
	output = subprocess.check_output(cmd, text=True).strip().split(",")
	return int(output[0]), int(output[1]), float(output[2])


	def _get_fps(video_path: str):
	"""Get video frame rate using ffprobe."""
	import subprocess

	cmd = [
	"ffprobe",
	"-v",
	"error",
	"-select_streams",
	"v:0",
	"-show_entries",
	"stream=r_frame_rate",
	"-of",
	"csv=p=0",
	video_path,
	]
	fps_str = subprocess.check_output(cmd, text=True).strip()
	num, den = map(int, fps_str.split("/"))
	return num / den


	def _build_video_filter(
	input_idx: int,
	video_idx: int,
	target_height: int,
	duration: float,
	max_duration: float,
	target_width: int = -2,
	):
	"""Build ffmpeg filter for a single video stream."""
	filters = []

	# Scale to target height, ensure even dimensions
	# filters.append(f"[{input_idx}:v]scale=-2:{target_height}[v{video_idx}_scaled]")
	filters.append(
	f"[{input_idx}:v]scale={target_width}:{target_height}[v{video_idx}_scaled]"
	)

	# Handle duration and graying
	if duration < max_duration:
	pad_duration = max_duration - duration
	filters.append(
	f"[v{video_idx}_scaled]tpad=stop_mode=clone:stop_duration={pad_duration}[v{video_idx}_padded]"
	)
	# Apply gray effect after original duration using eq filter with enable
	filters.append(
	f"[v{video_idx}_padded]eq=brightness=-0.5:saturation=0:enable='gte(t,{duration})'[v{video_idx}]"
	)
	else:
	# Video is already long enough, just use scaled version
	filters.append(f"[v{video_idx}_scaled]null[v{video_idx}]")

	return filters


	def make_composite_compare_videos(
	result_folder: str,
	save_dir: str,
	text_folder: str = None,
	compare_folders: list = None,
	compare_names: list = None,
	):
	"""Generates composite videos of (result \| compare_folders) with captions and text descriptions.

	Args:
	result_folder: Folder containing result videos (base for comparison)
	save_dir: Directory to save composite videos
	text_folder: Folder containing text descriptions (optional)
	compare_folders: List of folders to compare with result (optional)
	compare_names: List of names for compare folders (optional)

	Uses the longest video duration. Missing videos show black screen.
	Videos that end early show their last frame grayed out.
	Optimized version using ffmpeg directly for much faster processing.
	"""
	import subprocess
	import tempfile

	os.makedirs(save_dir, exist_ok=True)
	video_files = list(Path(result_folder).glob("*.mp4"))

	# Handle empty or non-existent compare folders
	if compare_folders is None:
	compare_folders = []
	if compare_names is None:
	compare_names = []

	# Filter out non-existent compare folders
	valid_compare_folders = []
	valid_compare_names = []
	for i, folder in enumerate(compare_folders):
	if folder and os.path.exists(folder):
	valid_compare_folders.append(folder)
	if i < len(compare_names):
	valid_compare_names.append(compare_names[i])
	else:
	valid_compare_names.append(f"Compare {i + 1}")

	compare_folders = valid_compare_folders
	compare_names = valid_compare_names

	for video_file in tqdm(video_files, desc="Creating composite videos"):
	video_id = video_file.stem

	# Prepare video paths - start with result, then add compare folders
	video_paths = [str(video_file)]
	video_names = ["Ours"]

	# Add compare folder videos
	for folder, name in zip(compare_folders, compare_names):
	compare_path = os.path.join(folder, f"{video_id}.mp4")
	video_paths.append(compare_path)
	video_names.append(name)

	video_exists = [os.path.exists(p) for p in video_paths]

	# Load text description
	if text_folder:
	text_file = os.path.join(text_folder, f"{video_id}.txt")
	text_content = (
	Path(text_file).read_text().strip()
	if os.path.exists(text_file)
	else f"Motion: {video_id}"
	)
	else:
	text_content = f"Motion: {video_id}"

	# Find reference video for properties (should always have result video)
	reference_video = str(video_file)
	if not os.path.exists(reference_video):
	print(f"Error: Result video not found for {video_id}, skipping")
	continue

	# Get video properties
	try:
	fps = _get_fps(reference_video)
	ref_width, ref_height, _ = _get_video_info(reference_video)
	except Exception as e:
	print(f"Error probing {video_id}: {e}, skipping")
	continue

	# Collect dimensions and durations for all videos
	widths, heights, durations = [], [], []
	for path, exists, name in zip(video_paths, video_exists, video_names):
	if exists:
	try:
	w, h, d = _get_video_info(path)
	widths.append(w)
	heights.append(h)
	durations.append(d)
	except Exception as e:
	print(
	f"Error probing {name} video for {video_id}: {e}, will use black screen"
	)
	widths.append(ref_width)
	heights.append(ref_height)
	durations.append(0)
	else:
	print(
	f"Warning: {name} video missing for {video_id}, will use black screen"
	)
	widths.append(ref_width)
	heights.append(ref_height)
	durations.append(0)

	max_duration = max(durations)
	if max_duration == 0:
	print(f"Warning: All videos for {video_id} have zero duration, skipping")
	continue

	target_height = min(h for h in heights if h > 0)
	# Ensure target height is even
	if target_height % 2 != 0:
	target_height += 1

	# Re-calculate widths based on target_height scaling
	new_widths = []
	for w, h in zip(widths, heights):
	if h > 0:
	# Calculate scaled width maintaining aspect ratio
	aspect_ratio = w / h
	scaled_w = int(target_height * aspect_ratio)
	# Ensure even width
	if scaled_w % 2 != 0:
	scaled_w += 1
	new_widths.append(scaled_w)
	else:
	new_widths.append(w)

	widths = new_widths
	total_width = sum(widths)

	# print(f"DEBUG: video_id={video_id}")
	# print(f"DEBUG: original widths={widths} (after update), heights={heights}")
	# print(f"DEBUG: target_height={target_height}")
	# print(f"DEBUG: total_width={total_width}")

	# Create and save title/text bars
	title_bar = render_aligned_title_bar(
	total_width, widths, video_names, font_size=32
	)
	text_bar = render_text_bar(text_content, width=total_width, font_size=24)

	with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as f:
	title_path = f.name
	Image.fromarray(title_bar).save(title_path)

	with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as f:
	text_path = f.name
	Image.fromarray(text_bar).save(text_path)

	# Build ffmpeg command
	inputs = []
	filter_parts = []
	input_count = 0

	# Add video inputs and filters
	for i, (path, exists, width, duration) in enumerate(
	zip(video_paths, video_exists, widths, durations)
	):
	if exists:
	inputs.extend(["-i", path])
	filter_parts.extend(
	_build_video_filter(
	input_count,
	i,
	target_height,
	duration,
	max_duration,
	target_width=width,
	)
	)
	else:
	inputs.extend(
	[
	"-f",
	"lavfi",
	"-i",
	f"color=c=black:s={width}x{target_height}:d={max_duration}:r={int(fps)}",
	]
	)
	filter_parts.append(f"[{input_count}:v]null[v{i}]")
	input_count += 1

	# Add title and text images
	inputs.extend(["-loop", "1", "-i", title_path])
	title_idx = input_count
	input_count += 1

	inputs.extend(["-loop", "1", "-i", text_path])
	text_idx = input_count

	# Compose final layout - horizontally stack all videos
	num_videos = len(video_paths)
	if num_videos == 1:
	filter_parts.append("[v0]null[videos]")
	else:
	video_inputs = "".join([f"[v{i}]" for i in range(num_videos)])
	filter_parts.append(f"{video_inputs}hstack=inputs={num_videos}[videos]")

	filter_parts.append(
	f"[{title_idx}:v][videos][{text_idx}:v]vstack=inputs=3[out]"
	)

	# Execute ffmpeg
	output_path = os.path.join(save_dir, f"{video_id}_composite.mp4")
	cmd = [
	"ffmpeg",
	"-y",
	*inputs,
	"-filter_complex",
	";".join(filter_parts),
	"-map",
	"[out]",
	"-t",
	str(max_duration),
	"-r",
	str(int(fps)),
	"-c:v",
	"libx264",
	"-preset",
	"ultrafast",
	"-crf",
	"23",
	"-pix_fmt",
	"yuv420p",
	"-profile:v",
	"baseline",
	"-level",
	"3.0",
	"-movflags",
	"+faststart",
	output_path,
	]

	try:
	result = subprocess.run(cmd, check=False, capture_output=True, text=True)
	if result.returncode != 0:
	print(f"Error processing {video_id}: Return code {result.returncode}")
	print(f"Command: {' '.join(cmd)}")
	print(f"Stderr: {result.stderr}")
	elif os.path.exists(output_path) and os.path.getsize(output_path) == 0:
	print(f"Warning: Generated video {output_path} is empty!")
	print(f"Command: {' '.join(cmd)}")
	print(f"Stderr: {result.stderr}")
	except Exception as e:
	print(f"Unexpected error processing {video_id}: {e}")
	finally:
	for path in [title_path, text_path]:
	try:
	os.unlink(path)
	except Exception:
	pass

	print(f"Composite videos saved to {save_dir}")