Skip to content

Generate clean 2D game sprites & animation atlases — component-row pipeline: state rows, alpha cleanup, frame extraction, runtime atlases. Codex/Claude skill.

fix: meet short sound loudness targets with peak limiting

Soohong Kim1bed2d2ed5fd
diff --git a/sprite_gen/audio/processing.py b/sprite_gen/audio/processing.py
index ec18955..44a85e0 100644
--- a/sprite_gen/audio/processing.py
+++ b/sprite_gen/audio/processing.py
@@ -78,7 +78,25 @@ def normalize(samples, target, loop):
result = _ffmpeg([*_pcm_args(), '-af', filt, '-ar', str(RATE), '-f', 'f32le', 'pipe:1'],
source.astype('<f4').tobytes())
output = np.frombuffer(result.stdout, dtype='<f4').reshape(-1, 2).copy()
- return output[len(samples):2*len(samples)] if loop else output[:len(samples)]
+ output = output[len(samples):2*len(samples)] if loop else output[:len(samples)]
+ corrections = 0
+ # A short transient can make loudnorm stop at the peak ceiling below the LUFS target.
+ # Measure the actual playback, then apply gain with an oversampled lookahead limiter.
+ # These are local DSP passes over the paid source, never provider retries.
+ for _ in range(3):
+ context = np.tile(output, (3, 1)) if loop else output
+ stats = measure(context, target)
+ if abs(stats['input_i'] - target) <= .2 and stats['input_tp'] <= -1.8:
+ break
+ gain = target - stats['input_i']
+ filt = (f'aresample=192000,volume={gain}dB,'
+ 'alimiter=limit=0.794328:level=false:latency=true:attack=1:release=20')
+ corrected = _ffmpeg([*_pcm_args(), '-af', filt, '-ar', str(RATE), '-f', 'f32le', 'pipe:1'],
+ context.astype('<f4').tobytes())
+ pcm = np.frombuffer(corrected.stdout, dtype='<f4').reshape(-1, 2).copy()
+ output = pcm[len(output):2*len(output)] if loop else pcm[:len(output)]
+ corrections += 1
+ return output, corrections
def export(source: Path, out_dir: Path, *, target_lufs, loop, crossfade):
@@ -86,7 +104,7 @@ def export(source: Path, out_dir: Path, *, target_lufs, loop, crossfade):
source_seconds = len(samples) / RATE
if loop:
samples = make_loop(samples, crossfade)
- samples = normalize(samples, target_lufs, loop)
+ samples, corrections = normalize(samples, target_lufs, loop)
duration = len(samples) / RATE
files = {}
for fmt, codec in (('ogg', ['-c:a', 'libopus', '-b:a', '160k']), ('m4a', ['-c:a', 'aac', '-b:a', '192k', '-movflags', '+faststart'])):
@@ -107,4 +125,5 @@ def export(source: Path, out_dir: Path, *, target_lufs, loop, crossfade):
files[fmt]['seam_delta'] = float(np.max(np.abs(playback[-1] - playback[0])))
return {'files': files, 'duration_seconds': duration, 'source_duration_seconds': source_seconds,
'sample_rate': RATE, 'loop_start_seconds': 0, 'loop_end_seconds': duration,
- 'crossfade_seconds': crossfade if loop else 0}
+ 'crossfade_seconds': crossfade if loop else 0,
+ 'normalization': {'method': 'loudnorm-two-pass', 'peak_limit_dbtp': -2, 'gain_limiter_passes': corrections}}
diff --git a/sprite_gen/compose/game_pack.py b/sprite_gen/compose/game_pack.py
index 6c56923..d6d17bf 100644
--- a/sprite_gen/compose/game_pack.py
+++ b/sprite_gen/compose/game_pack.py
@@ -227,7 +227,8 @@ def pack(request: Path, out_dir: Path, *, page_size=MAX_PAGE_SIZE, max_bytes=MAX
raise ValueError(f"duplicate audio id: {asset_id}")
if not isinstance(entry.get("source"), str) or not entry["source"]:
raise ValueError("each audio entry requires a source path")
- record, files, paths = load_sound(request.parent / entry["source"], max_bytes)
+ remaining = max_bytes - sum(map(len, audio_payloads.values()))
+ record, files, paths = load_sound(request.parent / entry["source"], remaining)
for fmt, data in files.items():
name = f"audio-{asset_id}.{fmt}"
record["files"][fmt]["path"] = name