From 16db1d2a010d4593c55423e6426bcc45b8a9041e Mon Sep 17 00:00:00 2001
From: leether <leether@outlook.com>
Date: Sun, 7 Jun 2026 15:55:58 +0800
Subject: [PATCH] fix(concat): correct xfade offset calculation for variable
 transition durations
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Bug: offset = cum_duration - trans_duration * i
  This assumed all transitions have the same duration. When a later
  transition duration increased (e.g., 0.4 -> 0.5), trans_duration*i
  could grow faster than cum_duration, causing negative offset regression.

  Example from rsi_cn timeline:
    i=38 offset=520.58  (td=0.4)
    i=39 offset=519.66  (td=0.5) ← regression!

Fix: offset = cum_duration - sum(all_transition_durations[:i])
  Each transition's actual duration is now accumulated individually.

Verified: all 48 offsets are monotonically increasing after fix.

Closes: concat-engine-offset-regression
---
 core/concat_engine.py | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/core/concat_engine.py b/core/concat_engine.py
index 51732db..fcb46fb 100644
--- a/core/concat_engine.py
+++ b/core/concat_engine.py
@@ -385,9 +385,13 @@ def _build_filter_complex(
                 trans_duration = trans.get("duration", self.config.default_transition_duration)
 
                 # 计算 offset
-                # offset = sum(clips[0:i].duration) - trans_duration * i
+                # offset = sum(clips[0:i].duration) - sum(all_transition_durations[0:i])
                 cum_duration = sum(c["duration"] for c in clips[:i])
-                offset = cum_duration - trans_duration * i
+                cum_trans_duration = sum(
+                    clips[j].get("transition", {}).get("duration", 0.0)
+                    for j in range(i)
+                )
+                offset = cum_duration - cum_trans_duration
 
                 # xfade 视频转场
                 xfade_types = {