nex_animation/backend/tests/test_compose_transition_tim...

188 lines
9.5 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters!

This file contains ambiguous Unicode characters that may be confused with others in your current locale. If your use case is intentional and legitimate, you can safely ignore this warning. Use the Escape button to highlight these characters.

"""转场与直切混排时的链路时间基:xfade 两侧必须一致(任务 #36 的返工)。
症状:同一条时间线,全直切能出片、全转场能出片,**只要「前面有一处直切、后面还有转场」
就必炸**,报错原文(服务端任务 #36 compose):
[Parsed_xfade_2] First input link main timebase (1/1000000) do not match
the corresponding second input link xfade timebase (1/90000)
Failed to configure output pad on Parsed_xfade_2 / Error reinitializing filters!
根因:阶段 1 归一化用 `-video_track_timescale 90000` 写 mp4,解出来的链路时间基是 1/90000;
而 `concat` 滤镜的**输出**链路时间基是 ffmpeg 默认的 AV_TIME_BASE(1/1000000)。串接链是
「累积链 + 下一路裸输入」,所以一旦 concat 之后又出现 xfade,两侧就变成了
1/1000000 vs 1/90000。xfade 要求两侧时间基完全相同,于是滤镜图重配置失败,
整条导出退化成「返回码 234 + 空文件」,用户看到的是一句"命令执行失败(返回码 1)"。
(这也是为什么它看起来像偶发:炸点取决于转场怎么分布,不取决于素材。)
口径:每一路进链之前先 `settb=AVTB`(音频 `asettb=AVTB`),把串接链上所有链路钉在同一个
时间基上,xfade ↔ concat 怎么交替都成立。
只有 xfade 严格这件事是量出来的,别误读成"所有混接滤镜都卡时间基"。同一批 90000 素材实测:
`[0:a][1:a]concat=v=0:a=1[a1];[a1][2:a]acrossfade` 不给 asettb 也 EXIT=0;
48k 归一后的样本与 24k 裸样本直接 amix 也 EXIT=0。音频侧的 asettb 是让整张图只剩一个
时间基口径的保险(原生音轨链与视频链同构,视频那一半才是必须的),命门在视频侧。
"""
from __future__ import annotations
import re
import subprocess
import pytest
from app.models import AudioClip, Timeline, TimelineSegment
from app.services import compose
from app.services import ffmpeg as ff
_CAPS = ("xfade", "concat", "timebase") # timebase = settb/asettb,ffmpeg 4.4+
needs_ffmpeg = pytest.mark.skipif(
not ff.available() or not all(ff.capabilities()[c] for c in _CAPS),
reason="需要支持 xfade/concat/settb 的 ffmpeg(缺哪个看 /api/system/health)")
SEG_MS, TD_MS, FPS, W, H, N = 2000, 500, 25, 320, 240, 4
# 转场分布:三元组依次是「段0→段1 / 段1→段2 / 段2→段3」是否给 500ms 转场。
# 最后那列 broken_before 是**去掉 settb 之后实测**的结果(见 test 里的反证),
# 只有「直切在前」和「交替」两种分布会炸 —— 与用户那句「有时能生成成功」严丝合缝。
SHAPES = {
"直切在前": ((False, True, True), True),
"转场在前": ((True, False, False), False),
"交替": ((True, False, True), True),
"全直切": ((False, False, False), False),
"全转场": ((True, True, True), False),
}
def _timeline(keep_audio: bool, shape: str) -> Timeline:
tds, _broken = SHAPES[shape]
tl = Timeline(project_id=1, fps=FPS, width=W, height=H,
settings_json={"keep_shot_audio": keep_audio})
for idx in range(N):
seg = TimelineSegment(timeline_id=1, order_index=idx + 1, duration_ms=SEG_MS,
source_in_ms=0, source_out_ms=SEG_MS, start_ms=idx * SEG_MS,
shot_id=idx + 1)
if idx:
# transition_duration_ms 挂在「前一段」上:段 idx-1 → 段 idx 的转场
tl.segments[idx - 1].transition_duration_ms = TD_MS if tds[idx - 1] else 0
tl.segments.append(seg)
tl.audio_clips.append(AudioClip(timeline_id=1, kind="voiceover", label="配音",
source_path="x.wav", start_ms=200, duration_ms=3000))
return tl
def _graph(tl: Timeline, *, keep_audio: bool):
"""输入号 0..N-1 是镜头片段,N 是配音缺失时的静音轨(混音总线那边的事)。"""
return compose.build_graph(tl, list(tl.segments), [], list(tl.audio_clips), {},
total_ms=N * SEG_MS, width=W, height=H, fps=FPS, ass_path=None,
target_lufs=-16, audio_inputs=[(N, None, "silence")])
# ---------------------------------------------------------------- 图形状:谁都不许是裸输入
STEP = re.compile(r"(xfade=|concat=n=|acrossfade=)")
def _inputs_of(part: str) -> list[str]:
"""取一步滤镜(xfade / concat / acrossfade)前面的那些输入标签。"""
head = part[:STEP.search(part).start()]
return re.findall(r"\[([^\]]+)\]", head)
@pytest.mark.parametrize("shape", list(SHAPES))
@pytest.mark.parametrize("keep", [False, True])
def test_chain_inputs_are_timebase_normalized(shape: str, keep: bool):
"""`[i:v]` / `[i:a]` 只许出现在 settb 那一行;链上每一步的输入必须是归一化过的标签。"""
tds, _broken = SHAPES[shape]
graph, _video = _graph(_timeline(keep, shape), keep_audio=keep)
parts = graph.split(";")
# 每一路(keep_shot_audio 时连音轨一起)都先过一次 settb=AVTB
assert len([p for p in parts if "settb=AVTB[" in p]) == N * (2 if keep else 1), shape
for part in parts:
# 归一化片段那几路(输入号 0..N-1)只许在 settb 行里被引用;
# N 之后是配音/BGM/静音轨,它们进的是混音总线,不归这条规矩管
head = re.match(r"^\[(\d+):(v|a)\]", part)
if head and int(head.group(1)) < N:
assert "settb=AVTB[" in part, part
# 串接步数:每个接缝一步;带转场且保留原生音轨时,acrossfade 与 xfade 各占一步
steps = [p for p in parts if STEP.search(p)]
assert len(steps) == (N - 1) + (keep * sum(1 for t in tds if t)), (shape, steps)
for part in steps:
for label in _inputs_of(part):
assert re.fullmatch(r"(vsrc|asrc|v|av)\d+", label), (label, part)
# ---------------------------------------------------------------- 真实 ffmpeg:五种分布都要出片
@pytest.fixture(scope="module")
def seg_files(tmp_path_factory) -> list[str]:
"""4 段 2s 素材,和阶段 1 一样用 -video_track_timescale 90000 写盘(时间基 1/90000)。"""
folder = tmp_path_factory.mktemp("tb-segs")
out = []
for i in range(N):
path = folder / f"seg-{i}.mp4"
ff.run([ff.FFMPEG, "-v", "error", "-y",
"-f", "lavfi", "-i", f"testsrc=size={W}x{H}:rate={FPS}:duration={SEG_MS / 1000}",
"-f", "lavfi", "-i", f"sine=frequency={300 + i * 90}:duration={SEG_MS / 1000}",
"-map", "0:v", "-map", "1:a", "-t", f"{SEG_MS / 1000:.3f}",
"-c:v", "libx264", "-preset", "ultrafast", "-pix_fmt", "yuv420p",
"-c:a", "aac", "-ar", "48000", "-ac", "2",
"-video_track_timescale", "90000", str(path)], timeout=300)
out.append(str(path))
return out
def _render(graph: str, vlabel: str, seg_files: list[str], out) -> subprocess.CompletedProcess:
cmd = [ff.FFMPEG, "-v", "error", "-y"]
for seg in seg_files:
cmd += ["-i", seg]
cmd += ["-f", "lavfi", "-t", "8", "-i", "anullsrc=r=48000:cl=stereo",
"-filter_complex", graph, "-map", vlabel, "-map", "[aout]",
"-c:v", "libx264", "-preset", "ultrafast", "-pix_fmt", "yuv420p",
"-c:a", "aac", "-ar", "48000", "-ac", "2", str(out)]
return ff.run(cmd, timeout=600, check=False)
def _unfix(graph: str) -> str:
"""还原成改前的图:删掉 settb 那一层,链上重新引用裸输入 [i:v]/[i:a]。"""
parts = [p for p in graph.split(";") if "settb=AVTB[" not in p]
out = []
for part in parts:
for idx in range(N):
part = part.replace(f"[vsrc{idx}]", f"[{idx}:v]").replace(f"[asrc{idx}]", f"[{idx}:a]")
out.append(part)
return ";".join(out)
def _stream_ms(path, selector: str) -> float:
proc = subprocess.run(["ffprobe", "-v", "error", "-select_streams", selector,
"-show_entries", "stream=duration", "-of", "csv=p=0", str(path)],
capture_output=True, text=True)
return float(proc.stdout.strip() or 0) * 1000.0
@needs_ffmpeg
@pytest.mark.parametrize("shape", list(SHAPES))
@pytest.mark.parametrize("keep", [False, True])
def test_every_transition_shape_renders(shape: str, keep: bool, seg_files, tmp_path):
"""改后五种转场分布都要出片;改前只有「直切在前 / 交替」两种会炸 —— 两件事一起钉住。"""
tds, broken_before = SHAPES[shape]
tl = _timeline(keep, shape)
graph, vlabel = _graph(tl, keep_audio=keep)
out = tmp_path / "film.mp4"
proc = _render(graph, vlabel, seg_files, out)
assert proc.returncode == 0, proc.stderr
assert "do not match" not in proc.stderr
probe = ff.probe(out)
assert (probe.width, probe.height) == (W, H)
# 量视频流本身:音频总线按 total_ms 补齐(apad + atrim),容器时长从来不代表画面长度
video_ms = _stream_ms(out, "v:0")
# 每个转场把时间线缩短 500ms(xfade 的重叠语义),直切不缩短
expect_ms = N * SEG_MS - TD_MS * sum(1 for t in tds if t)
assert abs(video_ms - expect_ms) < 350, (video_ms, expect_ms, probe.duration, proc.stderr)
old = tmp_path / "film-before-fix.mp4"
before = _render(_unfix(graph), vlabel, seg_files, old)
if broken_before:
assert before.returncode != 0 and "do not match" in before.stderr, \
"改前应当复现任务 #36 的报错;现在却不炸,说明 ffmpeg 或归一化参数变了,口径要重估"
else:
assert before.returncode == 0, before.stderr