FireRedAudio 可恢复批量配音 · T8star-Aix

This is the node the whole pack is built around. T8_FireRedAudio_BatchDubbing takes a parsed script (ScriptPlan) and a voice bank, and turns every line into a WAV, chunked, checkpointed, and resumable - the difference between "dub a chapter overnight" and "dub 20 lines and babysit it." If you came here from a reddit thread about FireRedAudio in ComfyUI, this is almost certainly the node you were told to find.

What it does

You connect a model handle, a script_plan (from ScriptParser/TextNormalizer), and a voice_bank (your characters' voice profiles). It walks the script in order and sends each line to the isolated worker's TTS. The clever part is the batch_size (default 8): the worker does latent-first, decode-later, meaning it generates the latent for the whole batch before switching to the audio decoder once. That avoids loading the heavy decoder back and forth per line, which is the difference between 8 lines per GPU trip and 1. The tooltip on batch_size even tells you the sweet spot: 4–8 on a 24 GB card.

Everything lands in ComfyUI/output/<subfolder>/<project_name>/. After each line succeeds, the manifest is atomically updated. That single detail is why this pack can claim interruption recovery: kill ComfyUI mid-batch and the completed entries are already durable.

Resume, cache, and the fingerprint

Re-run the same node and it doesn't re-generate everything. Each item carries a fingerprint combining the line text, voice profile, generation settings, and the model's identity. On resume, an item is only skipped if the fingerprint matches and the WAV still exists. Change a single line of script, swap the reference audio, or bump the seed - that line regenerates, everything else is skipped. The README's v0.15 numbers are the pitch: 100 lines / 8 voices in 13 batches, 100/100 done, then a re-run that hit the manifest cache in 0.136 seconds. Same-machine numbers, but the shape is real.

resume (default on) and continue_on_error (default on) are the two switches that make this usable unattended. Turn continue_on_error off only when you genuinely want the whole batch to stop on the first failure.

Inputs and outputs

Realistically you touch four things: model, script_plan, voice_bank, and batch_size. project_name and subfolder just name the output directory. Outputs are the full audio_batch, the manifest_path (feed it to AudioBatchResume later), and the batch report with per-item status.

Install

Standard pack install:

cd ComfyUI\custom_nodes
git clone https://github.com/T8mars/comfyui-fireredaudio-T8.git
cd comfyui-fireredaudio-T8
python scripts\setup_runtime.py

Then models: python scripts\download_models.py --profile full --variant int8-wo-safe-v1 into ComfyUI/models/TTS/FireRedAudio/. The full profile includes the RedAE decoder that generation needs. setup_runtime.py builds the isolated Python 3.10 / Transformers 5.8 worker - the pack deliberately keeps your ComfyUI Python clean, which is the sane answer to the Transformers-version wars that plague audio nodes.

Gotchas

A 100-line script is a 13+ minute run with a cold-start model load on top, so don't panic at the first run's wall time - the README explicitly warns its 120-second cold-start number is load, not speed. If a voice_bank profile is missing for a speaker, the node errors before generating anything, which is annoying and correct. And the fingerprint cache is strict: if the model identity changed (different variant, different acceleration mode), everything regenerates. That's by design - it refuses to silently reuse audio generated by a different model. The one thing you should never do is edit the WAVs in place and re-run; that's what the repair chain is for.