WatermarkRemover-AI/remwm.py

import sys
import click
from pathlib import Path
import cv2
import numpy as np
from PIL import Image, ImageDraw

from transformers import AutoProcessor, Florence2ForConditionalGeneration
from lama_inpaint.runtime import LamaInpaint
import torch
from torch.nn import Module
import tqdm
from loguru import logger
from enum import Enum
import os
import tempfile
import shutil
import subprocess

try:
    from cv2.typing import MatLike
except ImportError:
    MatLike = np.ndarray


def load_lama_model(device):
    """Load checksum-verified LaMA weights using the standalone adapter."""
    logger.info("Loading LaMA (verifying cached weights or downloading if missing)")
    return LamaInpaint(device)


class TaskType(str, Enum):
    OPEN_VOCAB_DETECTION = "<OPEN_VOCABULARY_DETECTION>"
    """Detect bounding box for objects and OCR text"""

def identify(task_prompt: TaskType, image: MatLike, text_input: str, model: Florence2ForConditionalGeneration, processor: AutoProcessor, device: str):
    if not isinstance(task_prompt, TaskType):
        raise ValueError(f"task_prompt must be a TaskType, but {task_prompt} is of type {type(task_prompt)}")

    prompt = task_prompt.value if text_input is None else task_prompt.value + text_input
    inputs = processor(text=prompt, images=image, return_tensors="pt")

    # Cast tensors to match the model's dtype and move to the correct device
    inputs = {k: v.to(device).to(model.dtype) if v.is_floating_point() else v.to(device)
              for k, v in inputs.items()}

    generated_ids = model.generate(
        input_ids=inputs["input_ids"],
        pixel_values=inputs["pixel_values"],
        max_new_tokens=1024,
        do_sample=False,
        num_beams=1,
    )
    generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
    return processor.post_process_generation(
        generated_text, task=task_prompt.value, image_size=(image.width, image.height)
    )

def get_watermark_mask(image: MatLike, model: Florence2ForConditionalGeneration, processor: AutoProcessor, device: str, max_bbox_percent: float, detection_prompt: str = "watermark"):
    """
    Detect watermarks and create a mask for inpainting.

    Args:
        image: PIL Image
        model: Florence-2 model
        processor: Florence-2 processor
        device: cuda or cpu
        max_bbox_percent: Maximum bbox size as percentage of image
        detection_prompt: Text prompt for detection (e.g. "watermark", "watermark Sora logo", "Getty Images")
    """
    task_prompt = TaskType.OPEN_VOCAB_DETECTION
    parsed_answer = identify(task_prompt, image, detection_prompt, model, processor, device)

    mask = Image.new("L", image.size, 0)
    draw = ImageDraw.Draw(mask)

    detection_key = "<OPEN_VOCABULARY_DETECTION>"
    if detection_key in parsed_answer and "bboxes" in parsed_answer[detection_key]:
        image_area = image.width * image.height
        for bbox in parsed_answer[detection_key]["bboxes"]:
            x1, y1, x2, y2 = map(int, bbox)
            bbox_area = (x2 - x1) * (y2 - y1)
            if (bbox_area / image_area) * 100 <= max_bbox_percent:
                draw.rectangle([x1, y1, x2, y2], fill=255)
            else:
                logger.warning(f"Skipping large bounding box: {bbox} covering {bbox_area / image_area:.2%} of the image")

    return mask


def detect_only(image: MatLike, model: Florence2ForConditionalGeneration, processor: AutoProcessor, device: str, max_bbox_percent: float, detection_prompt: str = "watermark"):
    """
    Detect watermarks and return bounding boxes WITHOUT creating mask or inpainting.
    Used for preview mode to show what would be detected.

    Returns:
        list of dicts with bbox info: [{"bbox": [x1,y1,x2,y2], "area_percent": float, "accepted": bool}, ...]
    """
    task_prompt = TaskType.OPEN_VOCAB_DETECTION
    parsed_answer = identify(task_prompt, image, detection_prompt, model, processor, device)

    results = []
    detection_key = "<OPEN_VOCABULARY_DETECTION>"

    if detection_key in parsed_answer and "bboxes" in parsed_answer[detection_key]:
        image_area = image.width * image.height
        for bbox in parsed_answer[detection_key]["bboxes"]:
            x1, y1, x2, y2 = map(int, bbox)
            bbox_area = (x2 - x1) * (y2 - y1)
            area_percent = (bbox_area / image_area) * 100
            accepted = area_percent <= max_bbox_percent

            results.append({
                "bbox": [x1, y1, x2, y2],
                "area_percent": round(area_percent, 2),
                "accepted": accepted
            })

    return results

def process_image_with_lama(image: MatLike, mask: MatLike, model_manager: LamaInpaint):
    """Return the shared LaMA adapter result as a uint8 BGR image."""
    result = model_manager(image, mask)

    if result.dtype in [np.float64, np.float32]:
        result = np.clip(result, 0, 255).astype(np.uint8)

    return result

def make_region_transparent(image: Image.Image, mask: Image.Image):
    image = image.convert("RGBA")
    mask = mask.convert("L")
    transparent_image = Image.new("RGBA", image.size)
    for x in range(image.width):
        for y in range(image.height):
            if mask.getpixel((x, y)) > 0:
                transparent_image.putpixel((x, y), (0, 0, 0, 0))
            else:
                transparent_image.putpixel((x, y), image.getpixel((x, y)))
    return transparent_image

def is_video_file(file_path):
    """Check if the file is a video based on its extension"""
    video_extensions = ['.mp4', '.avi', '.mov', '.mkv', '.flv', '.wmv', '.webm']
    return Path(file_path).suffix.lower() in video_extensions

def process_video(input_path, output_path, florence_model, florence_processor, model_manager, device, transparent, max_bbox_percent, force_format, detection_prompt="watermark", progress_offset=0, progress_scale=100):
    """Process a video file by extracting frames, removing watermarks, and reconstructing the video"""
    cap = cv2.VideoCapture(str(input_path))
    if not cap.isOpened():
        logger.error(f"Error opening video file: {input_path}")
        return

    # Get video properties
    fps = cap.get(cv2.CAP_PROP_FPS)
    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    
    # Determine output format
    if force_format:
        output_format = force_format.upper()
    else:
        output_format = "MP4"  # Default to MP4 for videos
    
    # Create output video file
    output_path = Path(output_path)
    if output_path.is_dir():
        output_file = output_path / f"{input_path.stem}_no_watermark.{output_format.lower()}"
    else:
        output_file = output_path.with_suffix(f".{output_format.lower()}")
    
    # Create a temporary file for the video without audio
    temp_dir = tempfile.mkdtemp()
    temp_video_path = Path(temp_dir) / f"temp_no_audio.{output_format.lower()}"
    
    # Set codec based on output format
    if output_format.upper() == "MP4":
        fourcc = cv2.VideoWriter_fourcc(*'mp4v')
    elif output_format.upper() == "AVI":
        fourcc = cv2.VideoWriter_fourcc(*'XVID')
    else:
        fourcc = cv2.VideoWriter_fourcc(*'mp4v')  # Default to MP4
    
    out = cv2.VideoWriter(str(temp_video_path), fourcc, fps, (width, height))
    
    # Process each frame
    with tqdm.tqdm(total=total_frames, desc="Processing video frames") as pbar:
        frame_count = 0
        while cap.isOpened():
            ret, frame = cap.read()
            if not ret:
                break
            
            # Convert frame to PIL Image
            frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
            pil_image = Image.fromarray(frame_rgb)
            
            # Get watermark mask
            mask_image = get_watermark_mask(pil_image, florence_model, florence_processor, device, max_bbox_percent, detection_prompt)
            
            # Process frame
            if transparent:
                # For video, we can't use transparency, so we'll fill with a color or background
                result_image = make_region_transparent(pil_image, mask_image)
                # Convert RGBA to RGB by filling transparent areas with white
                background = Image.new("RGB", result_image.size, (255, 255, 255))
                background.paste(result_image, mask=result_image.split()[3])
                result_image = background
            else:
                lama_result = process_image_with_lama(np.array(pil_image), np.array(mask_image), model_manager)
                result_image = Image.fromarray(cv2.cvtColor(lama_result, cv2.COLOR_BGR2RGB))
            
            # Convert back to OpenCV format and write to output video
            frame_result = cv2.cvtColor(np.array(result_image), cv2.COLOR_RGB2BGR)
            out.write(frame_result)
            
            # Update progress
            frame_count += 1
            pbar.update(1)
            local_progress = frame_count / total_frames
            progress = int(progress_offset + local_progress * progress_scale)
            print(f"Processing frame {frame_count}/{total_frames}, overall_progress:{progress}%")
    
    # Release resources
    cap.release()
    out.release()
    
    # Combine processed video with original audio using FFmpeg
    try:
        logger.info("Merging processed video with original audio...")
        
        # Check if FFmpeg is available
        try:
            subprocess.check_output(["ffmpeg", "-version"], stderr=subprocess.STDOUT)
        except (subprocess.SubprocessError, FileNotFoundError):
            logger.warning("FFmpeg is not available. Video will be produced without audio.")
            shutil.copy(str(temp_video_path), str(output_file))
        else:
            # Use FFmpeg to combine processed video with original audio
            ffmpeg_cmd = [
                "ffmpeg", "-y",
                "-i", str(temp_video_path),  # Processed video without audio
                "-i", str(input_path),       # Original video with audio
                "-c:v", "copy",              # Copy video without re-encoding
                "-c:a", "aac",               # Encode audio as AAC for better compatibility
                "-map", "0:v:0",             # Use video track from first file (processed video)
                "-map", "1:a:0",             # Use audio track from second file (original video)
                "-shortest",                  # End when the shortest track ends
                str(output_file)
            ]

            # Execute FFmpeg
            subprocess.run(ffmpeg_cmd, check=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
            logger.info("Audio/video merge completed successfully!")
    except Exception as e:
        logger.error(f"Error during audio/video merge: {str(e)}")
        # In case of error, use video without audio
        shutil.copy(str(temp_video_path), str(output_file))
    finally:
        # Clean up temporary files
        try:
            os.remove(str(temp_video_path))
            os.rmdir(temp_dir)
        except:
            pass
    
    final_progress = progress_offset + progress_scale
    logger.info(f"input_path:{input_path}, output_path:{output_file}, overall_progress:{final_progress}")
    return output_file


def process_video_two_pass(input_path, output_path, florence_model, florence_processor, model_manager, device, transparent, max_bbox_percent, force_format, detection_prompt="watermark", detection_skip=1, fade_in_sec=0.0, fade_out_sec=0.0, progress_offset=0, progress_scale=100):
    """
    Two-pass video processing with frame skip detection and fade in/out handling.

    Pass 1: Detect watermarks every N frames (sparse detection)
    Pass 2: Apply inpainting to all frames using interpolated masks

    This is more efficient for videos where watermarks don't change rapidly,
    and handles fade in/out watermarks by extending the mask temporally.
    """
    cap = cv2.VideoCapture(str(input_path))
    if not cap.isOpened():
        logger.error(f"Error opening video file: {input_path}")
        return

    # Get video properties
    fps = cap.get(cv2.CAP_PROP_FPS)
    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))

    # Convert seconds to frames
    fade_in_frames = int(fade_in_sec * fps)
    fade_out_frames = int(fade_out_sec * fps)

    logger.info(f"Two-pass processing: {total_frames} frames, skip={detection_skip}, fade_in={fade_in_frames}f, fade_out={fade_out_frames}f")

    # Determine output format
    if force_format:
        output_format = force_format.upper()
    else:
        output_format = "MP4"

    # Create output video file
    output_path = Path(output_path)
    if output_path.is_dir():
        output_file = output_path / f"{input_path.stem}_no_watermark.{output_format.lower()}"
    else:
        output_file = output_path.with_suffix(f".{output_format.lower()}")

    # ========== PASS 1: DETECTION (sparse) ==========
    logger.info("Pass 1: Detecting watermarks...")
    detections = {}  # frame_idx -> [bbox, bbox, ...]
    detection_frames = list(range(0, total_frames, detection_skip))

    with tqdm.tqdm(total=len(detection_frames), desc="Pass 1: Detection") as pbar:
        for frame_idx in detection_frames:
            cap.set(cv2.CAP_PROP_POS_FRAMES, frame_idx)
            ret, frame = cap.read()
            if not ret:
                break

            pil_image = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
            bboxes = detect_only(pil_image, florence_model, florence_processor, device, max_bbox_percent, detection_prompt)

            if bboxes:
                accepted_bboxes = [b["bbox"] for b in bboxes if b["accepted"]]
                if accepted_bboxes:
                    detections[frame_idx] = accepted_bboxes

            pbar.update(1)
            local_progress = (pbar.n / len(detection_frames)) * 0.5  # Pass 1 = 0-50% local
            progress = int(progress_offset + local_progress * progress_scale)
            print(f"Pass 1: frame {frame_idx}/{total_frames}, overall_progress:{progress}%")

    logger.info(f"Pass 1 complete: found watermarks in {len(detections)} detection points")

    # ========== TIMELINE EXPANSION ==========
    # Create frame->bbox mapping with fade in/out expansion
    frame_masks = {}  # frame_idx -> [bbox, ...]

    for det_frame, bboxes in detections.items():
        # Expand backwards (fade in) - watermark might be fading in before detection
        start_frame = max(0, det_frame - fade_in_frames)
        # Expand forwards (fade out) - continue masking after detection
        # Also include frames until next detection point
        end_frame = min(total_frames, det_frame + detection_skip + fade_out_frames)

        for f in range(start_frame, end_frame):
            if f not in frame_masks:
                frame_masks[f] = []
            # Add bboxes, avoiding duplicates
            for bbox in bboxes:
                if bbox not in frame_masks[f]:
                    frame_masks[f].append(bbox)

    logger.info(f"Timeline expanded: {len(frame_masks)} frames will have inpainting applied")

    # ========== PASS 2: INPAINTING ==========
    logger.info("Pass 2: Applying inpainting...")

    # Create temporary file for video without audio
    temp_dir = tempfile.mkdtemp()
    temp_video_path = Path(temp_dir) / f"temp_no_audio.{output_format.lower()}"

    # Set codec
    if output_format.upper() == "MP4":
        fourcc = cv2.VideoWriter_fourcc(*'mp4v')
    elif output_format.upper() == "AVI":
        fourcc = cv2.VideoWriter_fourcc(*'XVID')
    else:
        fourcc = cv2.VideoWriter_fourcc(*'mp4v')

    out = cv2.VideoWriter(str(temp_video_path), fourcc, fps, (width, height))

    # Reset video to beginning
    cap.set(cv2.CAP_PROP_POS_FRAMES, 0)

    with tqdm.tqdm(total=total_frames, desc="Pass 2: Inpainting") as pbar:
        frame_idx = 0
        while cap.isOpened():
            ret, frame = cap.read()
            if not ret:
                break

            if frame_idx in frame_masks:
                # This frame needs inpainting
                pil_image = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))

                # Create mask from bboxes
                mask = Image.new("L", pil_image.size, 0)
                draw = ImageDraw.Draw(mask)
                for bbox in frame_masks[frame_idx]:
                    x1, y1, x2, y2 = bbox
                    draw.rectangle([x1, y1, x2, y2], fill=255)

                # Apply inpainting or transparency
                if transparent:
                    result_image = make_region_transparent(pil_image, mask)
                    background = Image.new("RGB", result_image.size, (255, 255, 255))
                    background.paste(result_image, mask=result_image.split()[3])
                    result_image = background
                else:
                    lama_result = process_image_with_lama(np.array(pil_image), np.array(mask), model_manager)
                    result_image = Image.fromarray(cv2.cvtColor(lama_result, cv2.COLOR_BGR2RGB))

                frame_result = cv2.cvtColor(np.array(result_image), cv2.COLOR_RGB2BGR)
            else:
                # No watermark detected for this frame, copy original
                frame_result = frame

            out.write(frame_result)
            frame_idx += 1
            pbar.update(1)
            local_progress = 0.5 + (frame_idx / total_frames) * 0.5  # Pass 2 = 50-100% local
            progress = int(progress_offset + local_progress * progress_scale)
            print(f"Pass 2: frame {frame_idx}/{total_frames}, overall_progress:{progress}%")

    cap.release()
    out.release()

    # ========== MERGE WITH AUDIO ==========
    try:
        logger.info("Merging processed video with original audio...")
        try:
            subprocess.check_output(["ffmpeg", "-version"], stderr=subprocess.STDOUT)
        except (subprocess.SubprocessError, FileNotFoundError):
            logger.warning("FFmpeg is not available. Video will be produced without audio.")
            shutil.copy(str(temp_video_path), str(output_file))
        else:
            ffmpeg_cmd = [
                "ffmpeg", "-y",
                "-i", str(temp_video_path),
                "-i", str(input_path),
                "-c:v", "copy",
                "-c:a", "aac",
                "-map", "0:v:0",
                "-map", "1:a:0",
                "-shortest",
                str(output_file)
            ]
            subprocess.run(ffmpeg_cmd, check=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
            logger.info("Audio/video merge completed successfully!")
    except Exception as e:
        logger.error(f"Error during audio/video merge: {str(e)}")
        shutil.copy(str(temp_video_path), str(output_file))
    finally:
        try:
            os.remove(str(temp_video_path))
            os.rmdir(temp_dir)
        except:
            pass

    final_progress = progress_offset + progress_scale
    logger.info(f"input_path:{input_path}, output_path:{output_file}, overall_progress:{final_progress}")
    return output_file


def handle_one(image_path: Path, output_path: Path, florence_model, florence_processor, model_manager, device, transparent, max_bbox_percent, force_format, overwrite, detection_prompt="watermark", detection_skip=1, fade_in=0.0, fade_out=0.0, progress_offset=0, progress_scale=100):
    # SAFETY: Never overwrite the input file
    if image_path.resolve() == output_path.resolve():
        logger.error(f"Cannot overwrite input file: {image_path}. Choose a different output path.")
        print(f"ERROR: Cannot overwrite input file! Choose a different output folder.")
        return

    if output_path.exists() and not overwrite:
        logger.info(f"Skipping existing file: {output_path}")
        return

    # Check if it's a video file
    if is_video_file(image_path):
        # Use two-pass if detection_skip > 1 or fade handling is needed
        use_two_pass = detection_skip > 1 or fade_in > 0 or fade_out > 0
        if use_two_pass:
            return process_video_two_pass(image_path, output_path, florence_model, florence_processor, model_manager, device, transparent, max_bbox_percent, force_format, detection_prompt, detection_skip, fade_in, fade_out, progress_offset, progress_scale)
        else:
            return process_video(image_path, output_path, florence_model, florence_processor, model_manager, device, transparent, max_bbox_percent, force_format, detection_prompt, progress_offset, progress_scale)

    # Process image
    image = Image.open(image_path).convert("RGB")
    mask_image = get_watermark_mask(image, florence_model, florence_processor, device, max_bbox_percent, detection_prompt)

    if transparent:
        result_image = make_region_transparent(image, mask_image)
    else:
        lama_result = process_image_with_lama(np.array(image), np.array(mask_image), model_manager)
        result_image = Image.fromarray(cv2.cvtColor(lama_result, cv2.COLOR_BGR2RGB))

    # Determine output format
    if force_format:
        output_format = force_format.upper()
    elif transparent:
        output_format = "PNG"
    else:
        output_format = image_path.suffix[1:].upper()
        if output_format not in ["PNG", "WEBP", "JPG"]:
            output_format = "PNG"
    
    # Map JPG to JPEG for PIL compatibility
    if output_format == "JPG":
        output_format = "JPEG"

    if transparent and output_format == "JPG":
        logger.warning("Transparency detected. Defaulting to PNG for transparency support.")
        output_format = "PNG"

    new_output_path = output_path.with_suffix(f".{output_format.lower()}")
    result_image.save(new_output_path, format=output_format)
    # Report progress for this image (end of range)
    final_progress = progress_offset + progress_scale
    print(f"input_path:{image_path}, output_path:{new_output_path}, overall_progress:{final_progress}%")
    return new_output_path

@click.command()
@click.argument("input_path", type=click.Path(exists=True))
@click.argument("output_path", type=click.Path(), required=False, default=None)
@click.option("--preview", is_flag=True, help="Preview mode: detect watermarks and output JSON with base64 image (no processing).")
@click.option("--overwrite", is_flag=True, help="Overwrite existing files in bulk mode.")
@click.option("--transparent", is_flag=True, help="Make watermark regions transparent instead of removing.")
@click.option("--max-bbox-percent", default=10.0, help="Maximum percentage of the image that a bounding box can cover.")
@click.option("--force-format", type=click.Choice(["PNG", "WEBP", "JPG", "MP4", "AVI"], case_sensitive=False), default=None, help="Force output format. Defaults to input format.")
@click.option("--detection-prompt", default="watermark", help="Text prompt for watermark detection (e.g. 'watermark', 'watermark Sora logo', 'Getty Images').")
@click.option("--detection-skip", default=1, type=int, help="Detect watermarks every N frames for videos (1-10). Higher = faster but may miss brief watermarks.")
@click.option("--fade-in", default=0.0, type=float, help="Extend mask backwards by N seconds to handle fade-in watermarks.")
@click.option("--fade-out", default=0.0, type=float, help="Extend mask forwards by N seconds to handle fade-out watermarks.")
def main(input_path: str, output_path: str, preview: bool, overwrite: bool, transparent: bool, max_bbox_percent: float, force_format: str, detection_prompt: str, detection_skip: int, fade_in: float, fade_out: float):
    # Input validation
    if detection_skip < 1 or detection_skip > 10:
        logger.warning(f"detection_skip must be 1-10, got {detection_skip}. Using 1.")
        detection_skip = max(1, min(10, detection_skip))
    if fade_in < 0:
        fade_in = 0
    if fade_out < 0:
        fade_out = 0

    input_path = Path(input_path)

    # ========== PREVIEW MODE ==========
    if preview:
        import json
        import base64
        from io import BytesIO
        import random

        device = "cuda" if torch.cuda.is_available() else "cpu"

        # Force no dtype for CUDA (intentional default)
        # Apply float32 for CPU (compatibility)
        model_dtype = torch.float32 if device == "cpu" else None

        florence_model = Florence2ForConditionalGeneration.from_pretrained(
            "florence-community/Florence-2-large",
            torch_dtype=model_dtype).to(device).eval()
        florence_processor = AutoProcessor.from_pretrained("florence-community/Florence-2-large")

        # Get sample image from input
        if input_path.is_dir():
            # Get a random image from directory
            images = list(input_path.glob("*.[jp][pn]g")) + list(input_path.glob("*.webp"))
            videos = list(input_path.glob("*.mp4")) + list(input_path.glob("*.avi")) + list(input_path.glob("*.mov"))
            files = images + videos
            if not files:
                print(json.dumps({"error": "No supported files found in directory"}))
                return
            sample_path = random.choice(files)
        else:
            sample_path = input_path

        # Load image (extract frame if video)
        if is_video_file(sample_path):
            cap = cv2.VideoCapture(str(sample_path))
            total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
            # Get frame from middle of video
            cap.set(cv2.CAP_PROP_POS_FRAMES, total_frames // 2)
            ret, frame = cap.read()
            cap.release()
            if not ret:
                print(json.dumps({"error": f"Could not read frame from video: {sample_path}"}))
                return
            pil_image = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
            source_type = "video"
            source_frame = total_frames // 2
        else:
            pil_image = Image.open(sample_path).convert("RGB")
            source_type = "image"
            source_frame = None

        # Run detection
        detections = detect_only(pil_image, florence_model, florence_processor, device, max_bbox_percent, detection_prompt)

        # Draw bounding boxes on image
        draw = ImageDraw.Draw(pil_image)
        for det in detections:
            x1, y1, x2, y2 = det["bbox"]
            color = (0, 255, 0) if det["accepted"] else (255, 0, 0)  # Green if accepted, red if rejected
            draw.rectangle([x1, y1, x2, y2], outline=color, width=3)
            # Draw label
            label = f"{det['area_percent']:.1f}%"
            draw.text((x1, y1 - 15), label, fill=color)

        # Convert to base64
        buffer = BytesIO()
        pil_image.save(buffer, format="PNG")
        img_base64 = base64.b64encode(buffer.getvalue()).decode("utf-8")

        # Output JSON result
        result = {
            "image": img_base64,  # Just base64, GUI adds prefix
            "detections": detections,
            "source": str(sample_path),
            "source_type": source_type,
            "source_frame": source_frame,
            "prompt_used": detection_prompt,
            "max_bbox_percent": max_bbox_percent
        }
        print(json.dumps(result))
        return

    # ========== NORMAL PROCESSING MODE ==========
    output_path = Path(output_path)

    device = "cuda" if torch.cuda.is_available() else "cpu"
    print(f"Using device: {device}")

    # Force no dtype for CUDA (intentional default)
    # Apply float32 for CPU (compatibility)
    model_dtype = torch.float32 if device == "cpu" else None

    florence_model = Florence2ForConditionalGeneration.from_pretrained(
        "florence-community/Florence-2-large",
        torch_dtype=model_dtype).to(device).eval()
    florence_processor = AutoProcessor.from_pretrained("florence-community/Florence-2-large")
    logger.info("Florence-2 Model loaded")

    if not transparent:
        model_manager = load_lama_model(device)
        logger.info("LaMa model loaded")
    else:
        model_manager = None

    if input_path.is_dir():
        if not output_path.exists():
            output_path.mkdir(parents=True)

        # Include video files in the search
        images = list(input_path.glob("*.[jp][pn]g")) + list(input_path.glob("*.webp"))
        videos = list(input_path.glob("*.mp4")) + list(input_path.glob("*.avi")) + list(input_path.glob("*.mov")) + list(input_path.glob("*.mkv"))
        files = images + videos
        total_files = len(files)

        for idx, file_path in enumerate(tqdm.tqdm(files, desc="Processing files")):
            output_file = output_path / file_path.name
            # Calculate progress range for this file
            progress_offset = int(idx / total_files * 100)
            progress_scale = int(100 / total_files)
            handle_one(file_path, output_file, florence_model, florence_processor, model_manager, device, transparent, max_bbox_percent, force_format, overwrite, detection_prompt, detection_skip, fade_in, fade_out, progress_offset, progress_scale)
    else:
        # Single file mode - if output is a directory, construct file path
        if output_path.is_dir():
            output_file = output_path / input_path.name
        else:
            output_file = output_path

        # Ensure video output has proper extension
        if is_video_file(input_path) and output_file.suffix.lower() not in ['.mp4', '.avi', '.mov', '.mkv']:
            if force_format and force_format.upper() in ["MP4", "AVI"]:
                output_file = output_file.with_suffix(f".{force_format.lower()}")
            else:
                output_file = output_file.with_suffix(".mp4")  # Default to mp4

        handle_one(input_path, output_file, florence_model, florence_processor, model_manager, device, transparent, max_bbox_percent, force_format, overwrite, detection_prompt, detection_skip, fade_in, fade_out)
        print(f"input_path:{input_path}, output_path:{output_file}, overall_progress:100")

if __name__ == "__main__":
    main()