fix(worker): render_plan 保留视频原声并尊重每clip音量 #1474

Merged

auto-approve-bot merged 12 commits from fix/render-plan-audio-integrity into develop 2026-08-24 01:16:00 +08:00

Conversation 5 Commits 12 Files Changed 4

+443 -118

12 Commits

Author	SHA1	Message	Date
xiaoxia	af72a26fab	fix: unify audio filter order (denoise -> speed -> reverse -> volume) ... - _render_pass_through: move volume filter to last position (after speed/reverse), matching concat_main_audio order for consistent output - render_audio: remove redundant -t when atrim already controls duration	2026-08-24 01:04:02 +08:00
xiaoxia	cdf955ea3a	fix: use atrim filter for precise+efficient audio trimming in simple path ... Replace -ss output seeking (which decodes all data before trim point) with atrim filter in -af chain. Filter order: atrim -> asetpts -> volume, ensuring volume is only applied to the retained segment. When no trimming needed, use direct extraction for maximum efficiency.	2026-08-24 00:55:27 +08:00
xiaoxia	4f6ac9c8aa	fix: add speed_cfg.clamp() + comment explaining multi-stage atempo in pass-through ... SpeedEngine.build_audio_filter already handles out-of-range speeds via multi-stage atempo chaining (e.g. 0.25x -> atempo=0.5,atempo=0.5). Add explicit clamp() and comment to make this clear.	2026-08-24 00:49:03 +08:00
xiaoxia	b81e91e79a	fix: move -ss back after -i for sample-accurate audio trimming ... Input seek (-ss before -i) uses keyframe seeking which may skip to the wrong keyframe for audio, causing A/V desync in concatenated output. For a rendering service, precision > speed. Use output seeking.	2026-08-24 00:42:46 +08:00
xiaoxia	115455bacc	fix: volume=0 clips retained with silence filter to preserve A/V sync ... - render_audio.py: stop filtering out volume=0 clips from concat; they produce silence via volume=0 filter, keeping timeline aligned - unified_render_service.py: probe audio regardless of volume; volume=0 applied as filter, not by dropping stream - test: update test_silent_clip to verify volume=0 clip retained with volume=0 filter	2026-08-24 00:36:17 +08:00
xiaoxia	0081ce7641	Merge remote-tracking branch 'origin/develop' into fix/render-plan-audio-integrity	2026-08-24 00:30:34 +08:00
CI Bot	d7d68c0ceb	style: auto-format with black + isort + prettier [skip ci-format-check]	2026-08-23 16:24:38 +00:00
xiaoxia	41d503299a	fix: -ss as input option before -i, probe exception re-raises ... - render_audio.py: move -ss before -i so FFmpeg uses seek (fast) instead of decode-and-discard (slow) - unified_render_service.py: probe_has_audio fatal exception now logs error and re-raises (fail-fast, no silent audio loss) - test: update probe exception test to expect RuntimeError	2026-08-24 00:21:32 +08:00
xiaoxia	a628b4ca8e	fix: audio filter order (denoise before volume) + clarify reverse path comment ... - _render_pass_through: move volume filter after denoise to avoid amplifying noise before denoising - concat_main_audio: add comment explaining speed/reverse go through filter_complex else branch, not simple path	2026-08-24 00:15:24 +08:00
xiaoxia	2f76509a5b	fix: address AI code review - probe defaults to no audio, safe af filter ... - render_audio.py: simple mode builds af_parts list then joins into single -af (defensive against future multi-clip) - unified_render_service.py: probe_has_audio exception now defaults to False with warning log (was unsafe True) - test: add test_pass_through_probe_exception_defaults_to_no_audio	2026-08-24 00:10:07 +08:00
CI Bot	e3420610a3	style: auto-format with black + isort + prettier [skip ci-format-check]	2026-08-23 15:59:36 +00:00
saas-backend-agent	c2a510331e	fix(worker): render_plan 保留视频原声并尊重每clip音量 ... 根因：render_audio.mix_audio 中 main_clips = [] 无条件丢弃主图层源视频 原声（PR #1249 "避免录入杂音"），导致走新路径 render_plan 生成的成片 没有原声，与预览不一致（用户反馈"音频没有"）。 修复： - mix_audio 保留 main/broll 原声，按 clip.config.volume 控制音量， 无音频流/静音(volume=0)的 clip 安全过滤 - concat_main_audio 单clip与多clip路径均应用 per-clip volume 滤镜 - 直通渲染(_render_pass_through)改为先 probe_has_audio 再决定是否编码 音频，并应用 per-clip volume；不再无条件假设 main 带音频 - stream copy 在音量非默认时回退重编码 - 新增 _clip_volume 辅助函数(render_audio + UnifiedRenderService) - 更新 test_unified_render_service 旧"丢弃原声"断言为新行为 - 新增 test_original_audio_retained.py 9 个回归测试 验证：161 个相关单测全部通过	2026-08-23 23:54:44 +08:00