Update app.py
Browse files
app.py
CHANGED
|
@@ -609,6 +609,41 @@ def _refine_video_alpha(result_rgba: Image.Image, guide_rgb: np.ndarray) -> Imag
|
|
| 609 |
return result_rgba
|
| 610 |
|
| 611 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 612 |
def _reset_session(session_id: str) -> bool:
|
| 613 |
if not session_id:
|
| 614 |
return False
|
|
@@ -1459,6 +1494,7 @@ def _process_frames_sync(raw_frames, resolved_model, transparent,
|
|
| 1459 |
fallback_chain = ['silueta'] # Keep video RAM predictable; no heavyweight fallback.
|
| 1460 |
active_session = None # Disable frame-to-frame alpha ghosting on moving subjects.
|
| 1461 |
results = []
|
|
|
|
| 1462 |
for idx, content in enumerate(raw_frames):
|
| 1463 |
try:
|
| 1464 |
if not content:
|
|
@@ -1489,12 +1525,22 @@ def _process_frames_sync(raw_frames, resolved_model, transparent,
|
|
| 1489 |
results.append({"data":"", "format":"png", "error":"all_models_failed"})
|
| 1490 |
continue
|
| 1491 |
if result.mode != "RGBA": result = result.convert("RGBA")
|
|
|
|
|
|
|
| 1492 |
# v32.9.8: guided filter + soft contrast
|
| 1493 |
try:
|
| 1494 |
guide_rgb = np.array(img.convert("RGB"))
|
| 1495 |
result = _refine_video_alpha(result, guide_rgb)
|
| 1496 |
except Exception as e:
|
| 1497 |
logger.warning(f"[frames-batch] alpha refine failed: {e}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1498 |
blended = False
|
| 1499 |
if active_session is not None:
|
| 1500 |
arr = np.array(result)
|
|
@@ -1517,8 +1563,7 @@ def _process_frames_sync(raw_frames, resolved_model, transparent,
|
|
| 1517 |
results.append({"data":base64.b64encode(buf.getvalue()).decode("ascii"),
|
| 1518 |
"format":fmt,"size":buf.tell(),
|
| 1519 |
"model":used_model,"smoothed":blended,
|
| 1520 |
-
"matting":used_matting})
|
| 1521 |
-
del arr
|
| 1522 |
except Exception as e:
|
| 1523 |
logger.error(f"[frames-batch] frame {idx} failed: {e}")
|
| 1524 |
results.append({"data":"","format":"png","error":str(e)[:120]})
|
|
|
|
| 609 |
return result_rgba
|
| 610 |
|
| 611 |
|
| 612 |
+
def _recover_video_mask_dropout(current: np.ndarray, raw: np.ndarray,
|
| 613 |
+
previous: Optional[np.ndarray]) -> tuple:
|
| 614 |
+
"""Conservative, within-batch dropout recovery; never copies opaque old pixels.
|
| 615 |
+
|
| 616 |
+
A previous mask is used only as a spatial hint for pixels which the *current*
|
| 617 |
+
Silueta inference still assigns some foreground probability. This prevents
|
| 618 |
+
replacing a moving subject with an old silhouette. No state survives batches.
|
| 619 |
+
"""
|
| 620 |
+
if previous is None or previous.shape != current.shape:
|
| 621 |
+
return current, False
|
| 622 |
+
h, w = current.shape
|
| 623 |
+
prev_area = int(np.count_nonzero(previous >= 128))
|
| 624 |
+
cur_area = int(np.count_nonzero(current >= 128))
|
| 625 |
+
if prev_area < h*w*0.025 or cur_area < h*w*0.008:
|
| 626 |
+
return current, False
|
| 627 |
+
ratio = cur_area / max(1, prev_area)
|
| 628 |
+
if not (0.25 <= ratio < 0.63):
|
| 629 |
+
return current, False
|
| 630 |
+
prev_core = (previous >= 175).astype(np.uint8)
|
| 631 |
+
cur_core = (current >= 175).astype(np.uint8)
|
| 632 |
+
# Reject large subject motion; old outlines would cause ghosting.
|
| 633 |
+
intersection = np.count_nonzero((prev_core > 0) & (cur_core > 0))
|
| 634 |
+
if intersection / max(1, np.count_nonzero(cur_core)) < 0.65:
|
| 635 |
+
return current, False
|
| 636 |
+
# Recover only pixels already supported by current raw model confidence.
|
| 637 |
+
near_previous = cv2.dilate(prev_core, np.ones((7, 7), np.uint8)) > 0
|
| 638 |
+
candidate = near_previous & (raw >= 55) & (current < 128)
|
| 639 |
+
if not np.any(candidate):
|
| 640 |
+
return current, False
|
| 641 |
+
out = current.copy()
|
| 642 |
+
# A modest boost, never turn faint background evidence fully opaque.
|
| 643 |
+
out[candidate] = np.maximum(out[candidate], np.minimum(raw[candidate], 175))
|
| 644 |
+
return out, True
|
| 645 |
+
|
| 646 |
+
|
| 647 |
def _reset_session(session_id: str) -> bool:
|
| 648 |
if not session_id:
|
| 649 |
return False
|
|
|
|
| 1494 |
fallback_chain = ['silueta'] # Keep video RAM predictable; no heavyweight fallback.
|
| 1495 |
active_session = None # Disable frame-to-frame alpha ghosting on moving subjects.
|
| 1496 |
results = []
|
| 1497 |
+
previous_alpha = None
|
| 1498 |
for idx, content in enumerate(raw_frames):
|
| 1499 |
try:
|
| 1500 |
if not content:
|
|
|
|
| 1525 |
results.append({"data":"", "format":"png", "error":"all_models_failed"})
|
| 1526 |
continue
|
| 1527 |
if result.mode != "RGBA": result = result.convert("RGBA")
|
| 1528 |
+
# Keep raw Silueta confidence for conservative dropout recovery.
|
| 1529 |
+
raw_alpha = np.array(result.convert("RGBA"))[:, :, 3].copy()
|
| 1530 |
# v32.9.8: guided filter + soft contrast
|
| 1531 |
try:
|
| 1532 |
guide_rgb = np.array(img.convert("RGB"))
|
| 1533 |
result = _refine_video_alpha(result, guide_rgb)
|
| 1534 |
except Exception as e:
|
| 1535 |
logger.warning(f"[frames-batch] alpha refine failed: {e}")
|
| 1536 |
+
arr = np.array(result.convert("RGBA"))
|
| 1537 |
+
recovered, did_recover = _recover_video_mask_dropout(
|
| 1538 |
+
arr[:, :, 3], raw_alpha, previous_alpha)
|
| 1539 |
+
arr[:, :, 3] = recovered
|
| 1540 |
+
result = Image.fromarray(arr, "RGBA")
|
| 1541 |
+
# Store only reliable frames as references, never a recovered frame.
|
| 1542 |
+
if not did_recover:
|
| 1543 |
+
previous_alpha = recovered.copy()
|
| 1544 |
blended = False
|
| 1545 |
if active_session is not None:
|
| 1546 |
arr = np.array(result)
|
|
|
|
| 1563 |
results.append({"data":base64.b64encode(buf.getvalue()).decode("ascii"),
|
| 1564 |
"format":fmt,"size":buf.tell(),
|
| 1565 |
"model":used_model,"smoothed":blended,
|
| 1566 |
+
"matting":used_matting,"dropout_recovered":did_recover})
|
|
|
|
| 1567 |
except Exception as e:
|
| 1568 |
logger.error(f"[frames-batch] frame {idx} failed: {e}")
|
| 1569 |
results.append({"data":"","format":"png","error":str(e)[:120]})
|