Update app.py

app.py

@@ -867,6 +867,36 @@ def _fit_keyframe(image_path, current_canvas):
     return gr.update(value=image_path), gr.update(value=label)
 
 
+# Client-side only (`fn=None`, no server round-trip): reads the real `<video>` element's current playback
+# position, not a property of the file — so "grab this frame" means whatever's on screen when the button is
+# pressed, paused or scrubbed to, not automatically the clip's last frame.
+_FRAME_GRAB_JS = """
+function() {
+    const video = document.querySelector('#h3-generated-video video');
+    return video ? video.currentTime : 0;
+}
+"""
+
+
+def _extract_frame(video_path, timestamp):
+    """The frame at `timestamp` seconds into `video_path`, as a numpy RGB array — Gradio converts it to a PIL
+    image for whichever `gr.Image` this is wired to. Runs on CPU; no GPU time, no interaction with `_generate`."""
+    if not video_path:
+        return None
+    import cv2
+
+    cap = cv2.VideoCapture(video_path)
+    if not cap.isOpened():
+        return None
+    fps = cap.get(cv2.CAP_PROP_FPS) or FPS
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    target_frame = min(int(float(timestamp) * fps), max(0, total_frames - 1))
+    cap.set(cv2.CAP_PROP_POS_FRAMES, target_frame)
+    ok, frame = cap.read()
+    cap.release()
+    return cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) if ok else None
+
+
 load_models()
 
 INTRO = """# PlagueKind · MiniMax-H3
@@ -885,6 +915,16 @@ CSS = """
 .main.fillable {max-width: 1250px !important}
 .dark .gradio-container { color: var(--body-text-color); }
 .status p {font-size: 0.8rem; opacity: 0.65; text-align: center;}
+.h3-hidden-timestamp {
+    opacity: 0;
+    height: 0px;
+    width: 0px;
+    margin: 0px;
+    padding: 0px;
+    overflow: hidden;
+    position: absolute;
+    pointer-events: none;
+}
 """
 
 with gr.Blocks(title="PlagueKind · MiniMax-H3") as demo:
@@ -979,7 +1019,12 @@ with gr.Blocks(title="PlagueKind · MiniMax-H3") as demo:
                 )
 
         with gr.Column():
-            video = gr.Video(label="Video + soundtrack")
+            video = gr.Video(label="Video + soundtrack", elem_id="h3-generated-video")
+            with gr.Row():
+                grab_first_btn = gr.Button("📸 Use current frame as First frame", size="sm", variant="secondary")
+                grab_last_btn = gr.Button("📸 Use current frame as Last frame", size="sm", variant="secondary")
+            first_frame_timestamp = gr.Number(value=0, visible=True, elem_classes="h3-hidden-timestamp")
+            last_frame_timestamp = gr.Number(value=0, visible=True, elem_classes="h3-hidden-timestamp")
             report = gr.Markdown()
             with gr.Accordion("Distilled / Turbo LoRAs", open=False):
                 lora_1_strength = gr.Slider(
@@ -1060,6 +1105,18 @@ with gr.Blocks(title="PlagueKind · MiniMax-H3") as demo:
     first_frame.upload(_fit_keyframe, [first_frame, canvas], [first_frame, canvas])
     last_frame.upload(_fit_keyframe, [last_frame, canvas], [last_frame, canvas])
 
+    # Grabbing the currently-displayed frame: the button's own click runs only the JS above (`fn=None`, no
+    # server round-trip) to read the real `<video>` element's playback position into a hidden number box; that
+    # box's `.change()` is what actually decodes and writes the frame, server-side.
+    grab_first_btn.click(fn=None, inputs=None, outputs=[first_frame_timestamp], js=_FRAME_GRAB_JS)
+    first_frame_timestamp.change(
+        _extract_frame, [video, first_frame_timestamp], first_frame, show_progress="hidden"
+    )
+    grab_last_btn.click(fn=None, inputs=None, outputs=[last_frame_timestamp], js=_FRAME_GRAB_JS)
+    last_frame_timestamp.change(
+        _extract_frame, [video, last_frame_timestamp], last_frame, show_progress="hidden"
+    )
+    
     controls = [
         prompt,
         canvas,