completed codebase

.gitignore

@@ -160,3 +160,5 @@ cython_debug/
 #  and can be added to the global gitignore or merged into this file.  For a more nuclear
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 #.idea/
+
+.gradio/

florence_sam/app.py → app.py

@@ -2,26 +2,25 @@ from main import infer
 import moviepy.editor as mp
 import gradio as gr
 import os
+import tempfile
 
 def pre_processor(video_path, scale_factor, prompt, crop_duration):
     video = mp.VideoFileClip(video_path)
-
     cropped_video = video.subclip(0, min(crop_duration, video.duration))
 
-    temp_output = "cropped_video.mp4"
-    cropped_video.write_videofile(temp_output, codec="libx264")
+    with tempfile.NamedTemporaryFile(suffix='.mp4', delete=False) as temp_file:
+        temp_output = temp_file.name
+        cropped_video.write_videofile(temp_output, codec="libx264")
 
-    output = infer(temp_output, scale_factor, prompt)
+        output = infer(temp_output, scale_factor, prompt)
 
-    # Clean up temporary files
-    if os.path.exists(temp_output):
-        os.remove(temp_output)
+        # Clean up temporary file
+        os.unlink(temp_output)
 
     return output
 
-
 demo = gr.Interface(
-    title="Text based video inpainting",
+    title="Text Based Video Inpainting 🔥 (SAM2+Florance2+ProPainter)",
     fn=pre_processor, 
     inputs=[
         gr.Video(label="Upload Video"), 
@@ -32,4 +31,4 @@ demo = gr.Interface(
     outputs="video"
 )
 
-demo.launch()
+demo.launch(server_port=7555)

florence_sam/florancesam_pipeline.py → florancesam_pipeline.py

@@ -52,15 +52,15 @@ class VideoProcessor:
         self.scale_factor = scale_factor
 
         # Process video based on the prompt
-        output_video_path, session_path = self._process_prompt(video_path, prompt)
+        output_video_path, session_path, input_frames_dir, output_directory_path = self._process_prompt(video_path, prompt)
 
         # Create frames from the output video
-        fps = self._create_frames(output_video_path, os.path.join(session_path, "output_frames"))
+        fps = self._create_frames(output_video_path, output_directory_path)
         
         # Delete the output video
         os.remove(output_video_path)
 
-        return session_path, fps
+        return session_path, fps, input_frames_dir, output_directory_path
 
     def _create_frames(self, video_path, output_dir):
         create_directory(output_dir)
@@ -119,10 +119,14 @@ class VideoProcessor:
 
         # Generate unique name for video processing
         name = generate_unique_name()
-        session_path = os.path.join("tmp", name)
-        create_directory(session_path)
-        frame_directory_path = os.path.join(session_path, "input_frames")
-        create_directory(frame_directory_path)
+        # session_path = os.path.join("tmp", name)
+        # create_directory(session_path)
+        # frame_directory_path = os.path.join(session_path, "input_frames")
+        # create_directory(frame_directory_path)
+        import tempfile
+        session_path = tempfile.mkdtemp(prefix="video_processing_")
+        frame_directory_path = tempfile.mkdtemp(prefix="input_frames_", dir=session_path)
+        output_directory_path = tempfile.mkdtemp(prefix="output_frames_", dir=session_path)
 
         frames_sink = sv.ImageSink(
             target_dir_path=frame_directory_path,
@@ -157,7 +161,8 @@ class VideoProcessor:
             )
 
         # Create output video path
-        output_video_path = os.path.join("tmp", f"{name}.mp4")
+        # output_video_path = os.path.join("tmp", f"{name}.mp4")
+        output_video_path = os.path.join(session_path, f"{name}.mp4")
         frames_generator = sv.get_video_frames_generator(video_path)
         masks_generator = self.sam_video_model.propagate_in_video(inference_state)
 
@@ -185,7 +190,7 @@ class VideoProcessor:
                 annotated_frame = (annotated_frame > 0).astype(np.uint8) * 255
                 sink.write_frame(annotated_frame)
 
-        return output_video_path, session_path
+        return output_video_path, session_path, frame_directory_path, output_directory_path
 
 #Example usage
 # output_video = video_processor.process_video(

florence_sam/README.md

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:031c6f979f28b8b0f3b42af389fd1f741046ca89de9bc05a33912db7e30e8741
-size 258

florence_sam/requirements.txt

@@ -1,12 +0,0 @@
-tqdm
-einops
-spaces
-timm
-transformers
-samv2
-gradio
-supervision
-opencv-python
-pytest
-imageio
-moviepy

florence_sam/main.py → main.py

@@ -3,6 +3,7 @@ import shutil
 import torch
 from propainter_pipeline import process_video
 from florancesam_pipeline import VideoProcessor
+import tempfile
 
 video_processor = VideoProcessor()
 video_processor._enable_mixed_precision()
@@ -12,7 +13,7 @@ def infer(video_path, scale_factor, prompt):
     video_processor._enable_mixed_precision()
     # --------------------------------------------------------------------------------
     print("Processing video with FlorenceSam...")
-    session_path, fps = video_processor.process_video(
+    session_path, fps, input_frames_dir, output_frames_dir = video_processor.process_video(
         video_path=video_path,
         scale_factor=scale_factor,
         prompt=prompt
@@ -26,13 +27,14 @@ def infer(video_path, scale_factor, prompt):
     torch.cuda.empty_cache()
 
     print("Processing video with ProPainter...")
-    process_video(video=os.path.join(session_path, "input_frames"), mask=os.path.join(session_path, "output_frames"), save_fps=int(fps), fp16=True)
+    result_path = tempfile.mkdtemp()
+    inpainted_video = process_video(video=input_frames_dir, mask=output_frames_dir, save_fps=int(fps), fp16=True, output=result_path)
 
-    # remove intermediate files
-    shutil.rmtree("tmp")
+    shutil.rmtree(input_frames_dir)
+    shutil.rmtree(output_frames_dir)
     torch.cuda.empty_cache()
-
-    return "results/input_frames/inpaint_out.mp4"
+    
+    return inpainted_video
 
 if __name__ == "__main__":
     infer("/home/ubuntu/ahmedghani/clip-07-camera-2.mp4", 0.5, "players, basketball, rim, players shadow")