nex_animation/backend/tests/test_loudnorm_two_pass.py

103 lines
4.5 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.

"""成片响度归一(EBU R128):两遍 loudnorm 必须命中目标响度。
背景(验收 8.4 实测踩到):单遍 loudnorm 是动态归一,十几秒的短片实测目标 -16
LUFS 只落到 -18.5 LUFS,成片偏轻、多个项目之间响度不一致。合成前先跑一遍只测量,
再把实测值喂给第二遍(linear=true),整片响度才稳定命中目标。
"""
from __future__ import annotations
import re
import subprocess
import pytest
from app.services import compose
from app.services import ffmpeg as ff
needs_ffmpeg = pytest.mark.skipif(not ff.available() or not ff.capabilities()["loudnorm"],
reason="需要支持 loudnorm 的 ffmpeg")
PULSED = ("aevalsrc=0.9*sin(2*PI*440*t)*between(mod(t\\,1.2)\\,0\\,0.15)"
":d=6:s=48000") # 短促爆音 + 大段留白:单遍动态归一最容易在这类素材上跑偏
def _base_cmd() -> list[str]:
return [ff.FFMPEG, "-v", "error", "-y",
"-f", "lavfi", "-i", "color=c=black:s=64x64:r=25:d=6",
"-f", "lavfi", "-i", PULSED]
def _graph(tail: str = "") -> str:
return f"[1:a]loudnorm=I=-16.0:TP=-1.5:LRA=11{tail},atrim=0:6.000,asetpts=PTS-STARTPTS[aout]"
def test_without_loglevel_strips_only_log_flags():
args = ["-v", "error", "-y", "-i", "a.mp4", "-loglevel", "warning", "-t", "3"]
assert compose._without_loglevel(args) == ["-y", "-i", "a.mp4", "-t", "3"]
def test_loudnorm_measure_parses_report_and_builds_tail(monkeypatch):
report = ('{ "input_i" : "-9.42", "input_tp" : "-3.10", "input_lra" : "4.20",\n'
' "input_thresh" : "-20.55", "target_offset" : "-0.31" }')
seen: dict = {}
class Proc:
returncode = 0
stderr = report
stdout = ""
def fake_run(cmd, **kw):
seen["cmd"] = cmd
return Proc()
monkeypatch.setattr(ff, "run", fake_run)
measured = compose.loudnorm_measure(_base_cmd(), _graph(":print_format=json"), "[0:v]", 3000)
assert measured["input_i"] == pytest.approx(-9.42)
# 测量必须自己抬升日志级别:主命令是 -v error,会把 loudnorm 的测量报告整段吞掉
assert seen["cmd"][:3] == [ff.FFMPEG, "-loglevel", "info"]
assert "-v" not in seen["cmd"] and "error" not in seen["cmd"]
tail = compose.loudnorm_tail(measured)
assert "measured_I=-9.42" in tail and "linear=true" in tail
def test_loudnorm_measure_raises_on_silence(monkeypatch):
class Proc:
returncode = 0
stderr = ('{ "input_i" : "-inf", "input_tp" : "-inf", "input_lra" : "0.00",'
' "input_thresh" : "-inf", "target_offset" : "0.00" }')
stdout = ""
monkeypatch.setattr(ff, "run", lambda *a, **k: Proc())
with pytest.raises(compose.ComposeError):
compose.loudnorm_measure(_base_cmd(), _graph(":print_format=json"), "[0:v]", 3000)
def test_build_graph_embeds_measured_loudnorm_params():
"""两遍归一的落点:第二遍的 filtergraph 必须带上实测值 + linear=true。"""
tail = ":measured_I=-9.42:measured_TP=-3.10:measured_LRA=4.20:" \
":measured_thresh=-20.55:offset=-0.31:linear=true"
assert tail not in _graph()
assert tail in _graph(tail)
@needs_ffmpeg
def test_two_pass_loudnorm_hits_target(tmp_path):
"""同一段「爆音 + 留白」素材:两遍归一命中目标,单遍动态归一明显跑偏。"""
cmd = _base_cmd()
measured = compose.loudnorm_measure(cmd, _graph(":print_format=json"), "0:v", 6000)
assert measured["input_i"] < -6.0 and measured["input_tp"] > -3.0
def render(tail: str, name: str) -> float:
out = tmp_path / name
subprocess.run(cmd + ["-filter_complex", _graph(tail), "-map", "0:v", "-map", "[aout]",
"-c:v", "libx264", "-preset", "ultrafast", "-pix_fmt", "yuv420p",
str(out)], check=True, capture_output=True)
text = subprocess.run([ff.FFMPEG, "-hide_banner", "-nostats", "-i", str(out),
"-filter_complex", "ebur128=framelog=quiet", "-f", "null", "-"],
capture_output=True, text=True).stderr
return float(re.findall(r"I:\s*(-?\d+\.?\d*)\s*LUFS", text)[-1])
one_pass = render("", "one.mp4")
two_pass = render(compose.loudnorm_tail(measured), "two.mp4")
assert abs(two_pass - (-16.0)) <= 0.5, f"两遍归一偏离目标 -16 LUFS:实测 {two_pass} LUFS"
assert abs(one_pass - (-16.0)) > abs(two_pass - (-16.0)) + 0.3, \
f"单遍动态归一应明显更偏:单遍 {one_pass} vs 两遍 {two_pass}"