FireRedAudio 批量保存/下载 · T8star-Aix

Individual save nodes are fine when you have one file. The moment you have a full AudioBatch - a dozen, a hundred takes - you want SaveAudioBatch. It copies or transcodes every successful take in the batch into a timestamped export folder under ComfyUI/output/, writes a portable export-manifest.json with per-file SHA-256, optionally zips the whole thing, and registers a native listen/download list in the results area (16 players by default). It's the "give me all the good takes, in a folder, with receipts" node.

How it works

It iterates audio_batch's successful items, names each file from its index/speaker/line ID (e.g. 0001-roleA-line-003.wav), and either copies the original or transcodes to your chosen audio_format (wav/flac/mp3/ogg). Each exported file gets an export_sha256 recorded in the manifest so you can later prove what shipped. Failed exports are collected in the report - continue_on_error (default on) keeps going past a bad file instead of dying mid-batch. When create_zip is on, everything in the export folder gets compressed to a sibling .zip.

Outputs:

• 原 AudioBatch（透传） - the batch passes through, so you can chain this into a final save or another export at a different format.
• 导出 Manifest 路径 - the portable manifest (project name, source manifest path, format, per-item paths + hashes).
• ZIP 路径 - the archive, when created.
• 导出报告 - counts, failures, previewed count, output directory.

The inputs that matter

• audio_format - the export codec. Copying WAVs as WAV is lossless and instant; FLAC is the compact-lossless choice; MP3/OGG when size wins.
• project_name and subfolder - the export lands at output/<subfolder>/<project>-<timestamp>/. A fresh timestamped folder every run means exports never clobber each other.
• preview_limit - how many native players appear in the results area (default 16, max 100). Lower it for a huge batch so ComfyUI doesn't try to render a hundred players.

Where it fits

End of the pipeline, after QA and review: BatchDubbing → SpeechQA → LineReview → this node. The 20_creator_qa_repair_delivery example is exactly that: chunked batch dubbing, QA-fail repair, take selection, then a batch export + ZIP. It also accepts batches restored via the resume node, so an interrupted session that reuses its manifest can still export cleanly - the export writes its own separate folder and manifest, never touching the source batch or its WAVs.

Installing and the gotcha

Pack-wide install (ComfyUI Manager search comfyui-fireredaudio-T8, or clone + python scripts\setup_runtime.py); no model needed. The one thing to remember: only successful takes are exported. Failed or held items in the batch are silently skipped (reported in the report), so check the QA/review status before you ship - an "all exported" zip with a missing line is the classic quiet failure. Also note the results-area list caps at preview_limit, but the export on disk always contains everything.