FireRedAudio 分轨制作交付包 · T8star-Aix

At some point the dub is done and someone else needs the files: an editor, a mixer, a client, your future self. ProductionPackage is the pack's answer to that handoff. It takes your finished AudioBatch and produces a proper delivery folder - equal-length dialogue stems per role and per scene, a dialogue master, the real SRT/VTT generated from actual timeline placements, any external Master/BGM/room tone you connect, and a manifest that records every version, parameter, and SHA-256 hash. Then it zips the lot. This is the node you reach for when "just save the WAVs" stops being enough.

How it works

It renders the batch's successful lines into a single dialogue master (sequence mode when lines have no timing, timeline mode when they carry start_seconds/end_seconds), then slices that same render into stems grouped by speaker (role stems) and scene (scene stems), each padded with silence to the same end point as the master. That equal-length padding is what makes the stems drop straight into a DAW and line up - no nudging, no "why is this stem short."

What it won't do is pretend to be a mixer. If you connect a bgm_audio or room_tone_audio, those files are included in the package as-is with mixed_into_master_by_this_node: false in the manifest. It won't secretly blend BGM under your dialogue. The SRT/VTT are built from the render's real placements (not the script's guesses), so subtitle timing matches the audio that's actually in the zip.

The inputs that matter

• audio_batch - required; must contain at least one successful take or the node refuses.
• sample_rate (default 48000) and crossfade_ms (default 40) - the production format. A delivery_preset input overrides both, so if you're using the pack's audiobook/podcast/video_dialogue presets you don't need to set these by hand.
• include_role_stems / include_scene_stems - both on by default; untick what you don't need to keep the folder small.
• create_zip - on by default. The ZIP also carries code/model revisions, worker revision, generation parameters, and the full SHA-256 inventory.

Outputs: 原 AudioBatch（透传）, 交付 Master (the AUDIO), plus the 制作 Manifest 路径, 交付 ZIP 路径, and a 制作包报告 string. Everything lands under ComfyUI/output/<subfolder>/<project_name>-<timestamp>/.

Where it fits

End of the production chain, after QA and review: BatchDubbing → SpeechQA → LineReview → this. The README's example 22_production_package shows it pulling from a project-exchange AudioBatch, so you can even hand off a project you finished in the author's desktop companion app. It's an output node (is_output_node: true), so it's the last stop - nothing downstream.

Installing

Pack-wide: ComfyUI Manager search comfyui-fireredaudio-T8, or clone + python scripts\setup_runtime.py. No model needed for the packaging itself, but the AudioBatch it consumes comes from model-bearing nodes, so the full install applies. The one trap: if your batch has failed lines, they're excluded from the package - check the QA report first, or you'll ship a dub with a hole in it.