Perf: ONNX multi-thread, 320px det_size for video, fast-mode frame skip UI

app.py

@@ -66,7 +66,7 @@ def body_swap_image(source_img, target_img, blend_strength):
         return None, f"Error: {e}"
 
 
-def face_swap_video(source_img, target_video, enhance, progress=gr.Progress()):
+def face_swap_video(source_img, target_video, enhance, fast_mode, progress=gr.Progress()):
     if source_img is None or target_video is None:
         return None, "Please upload a source face image and a target video."
 
@@ -77,12 +77,13 @@ def face_swap_video(source_img, target_video, enhance, progress=gr.Progress()):
         target_video,
         mode="face",
         enhance=enhance,
+        fast_mode=fast_mode,
         progress=progress,
     )
     return output_path, msg
 
 
-def body_swap_video(source_img, target_video, blend_strength, progress=gr.Progress()):
+def body_swap_video(source_img, target_video, blend_strength, fast_mode, progress=gr.Progress()):
     if source_img is None or target_video is None:
         return None, "Please upload a source body image and a target video."
 
@@ -93,6 +94,7 @@ def body_swap_video(source_img, target_video, blend_strength, progress=gr.Progre
         target_video,
         mode="body",
         blend_strength=blend_strength,
+        fast_mode=fast_mode,
         progress=progress,
     )
     return output_path, msg
@@ -174,6 +176,10 @@ with gr.Blocks(title="Face & Body Swapper", theme=gr.themes.Soft()) as demo:
                         label="Enhance faces (sharpening + contrast)",
                         value=True,
                     )
+                    fv_fast = gr.Checkbox(
+                        label="⚡ Fast Mode — skip every other frame (~2× speed, slight motion blur)",
+                        value=False,
+                    )
                     fv_btn = gr.Button("Swap Faces in Video", variant="primary")
                 with gr.Column(scale=1):
                     fv_output = gr.Video(label="Result Video")
@@ -181,7 +187,7 @@ with gr.Blocks(title="Face & Body Swapper", theme=gr.themes.Soft()) as demo:
 
             fv_btn.click(
                 face_swap_video,
-                inputs=[fv_source, fv_target, fv_enhance],
+                inputs=[fv_source, fv_target, fv_enhance, fv_fast],
                 outputs=[fv_output, fv_status],
             )
 
@@ -199,6 +205,10 @@ with gr.Blocks(title="Face & Body Swapper", theme=gr.themes.Soft()) as demo:
                         minimum=0.1, maximum=1.0, value=0.85, step=0.05,
                         label="Blend Strength",
                     )
+                    bv_fast = gr.Checkbox(
+                        label="⚡ Fast Mode — skip every other frame (~2× speed)",
+                        value=False,
+                    )
                     bv_btn = gr.Button("Swap Body in Video", variant="primary")
                 with gr.Column(scale=1):
                     bv_output = gr.Video(label="Result Video")
@@ -206,7 +216,7 @@ with gr.Blocks(title="Face & Body Swapper", theme=gr.themes.Soft()) as demo:
 
             bv_btn.click(
                 body_swap_video,
-                inputs=[bv_source, bv_target, bv_blend],
+                inputs=[bv_source, bv_target, bv_blend, bv_fast],
                 outputs=[bv_output, bv_status],
             )
 

processors/face_swap.py

@@ -102,16 +102,25 @@ class FaceSwapper:
 
         import insightface
         from insightface.app import FaceAnalysis
+        import onnxruntime as ort
+        import multiprocessing
 
-        # Face analysis (buffalo_l auto-downloads on first run)
-        # Initialize face analysis (CPU-only for free HF Spaces tier)
+        n_threads = multiprocessing.cpu_count()
+
+        # Use all available CPU cores for ONNX inference
+        sess_opts = ort.SessionOptions()
+        sess_opts.intra_op_num_threads = n_threads
+        sess_opts.inter_op_num_threads = n_threads
+        sess_opts.execution_mode = ort.ExecutionMode.ORT_PARALLEL
+
+        # Face analysis — 640 for images, 320 for video (set via swap_frame)
         self._app = FaceAnalysis(
             name="buffalo_l",
             providers=["CPUExecutionProvider"],
         )
         self._app.prepare(ctx_id=-1, det_size=(640, 640))
 
-        # inswapper model
+        # inswapper model with multi-thread session options
         _download_inswapper()
         self._swapper = insightface.model_zoo.get_model(
             str(INSWAPPER_PATH),
@@ -270,7 +279,10 @@ class FaceSwapper:
             )
 
         if cached_target_faces is None:
+            # Use smaller det_size for video to speed up detection
+            self._app.det_model.input_size = (320, 320)
             target_faces = self._app.get(target_bgr)
+            self._app.det_model.input_size = (640, 640)  # restore for images
         else:
             target_faces = cached_target_faces
 

processors/video_processor.py

@@ -40,6 +40,7 @@ class VideoProcessor:
         mode: str = "face",          # "face" | "body"
         enhance: bool = False,
         blend_strength: float = 0.85,
+        fast_mode: bool = False,     # skip every other frame (~2x speed)
         progress=None,
     ) -> tuple[str | None, str]:
         """
@@ -82,6 +83,7 @@ class VideoProcessor:
         processed        = 0
         errors           = 0
         cached_tgt_faces = None   # reused across DET_INTERVAL frames
+        last_result      = None   # for fast_mode frame duplication
 
         while True:
             ret, frame = cap.read()
@@ -94,6 +96,12 @@ class VideoProcessor:
                     f"Processing frame {frame_idx + 1} / {total_frames}",
                 )
 
+            # Fast mode: skip odd frames — duplicate the previous processed result
+            if fast_mode and frame_idx % 2 == 1 and last_result is not None:
+                writer.write(last_result)
+                frame_idx += 1
+                continue
+
             # Only re-detect target faces every DET_INTERVAL frames
             use_cache = (mode == "face") and (frame_idx % DET_INTERVAL != 0) and (cached_tgt_faces is not None)
 
@@ -109,9 +117,11 @@ class VideoProcessor:
 
             if result_frame is not None:
                 writer.write(result_frame)
+                last_result = result_frame
                 processed += 1
             else:
                 writer.write(frame)  # keep original on failure
+                last_result = frame
                 errors += 1
 
             frame_idx += 1