FireRedAudio 恢复批次/审核会话 · T8star-Aix

You ran a 100-line dubbing batch. It got through 83 lines, your GPU driver hiccuped, ComfyUI died, and you're staring at a terminal full of output paths. The bad old way: re-run the batch and re-generate all 100 lines. The way this pack designed it: point this node at the manifest, get your 83 finished takes back as a proper AudioBatch, and let BatchDubbing resume from where it stopped. This is the node that makes long production runs survivable.

What it is

T8_FireRedAudio_AudioBatchResume reads a manifest - from batch dubbing, a repair run, duration fitting, or a review session - and reconstructs the AudioBatch object those nodes work with. Everything else in the pack consumes T8_FIREREDAUDIO_AUDIO_BATCH, and normally that object comes out of the node that produced it. This node re-creates it from disk, after the fact, which is what turns a half-finished session into a resumable one.

The inputs are the point. manifest_path is the only one you actually need: paste the path from the batch report, or a relative path under ComfyUI's output folder. missing_policy defaults to mark_missing, which is the gentle behavior - if a WAV vanished, the restored batch flags the item instead of throwing. Set it to error if you'd rather the node refuse to continue with gaps. verify_hashes (default off) re-checks recorded SHA-256s; turn it on when you suspect files were edited behind your back. allow_external_manifest is the advanced safety valve: by default the node only reads manifests inside ComfyUI's output, because letting arbitrary files in is how you get confused. Flip it only when your project lives on an external drive.

Why the trust model matters

This pack is aggressively non-destructive, and resume is where that philosophy pays off. Restoring a batch runs no model and modifies nothing - it reads a manifest, checks files, reports. That means you can load a session at the start of a new ComfyUI run, inspect the dashboard of approved/review/retry/missing items, and decide what actually needs work before you spend a single GPU second.

The UI dashboard (the fireredaudio_resume_dashboard in the node's output) summarizes statuses and tells you the next step. That's the feature the README is proudest of: "已通过/待审核/待返修/缺失和导出就绪" - approved, pending review, pending repair, missing, and ready-to-export. You don't have to guess where you left off.

Install

It's part of comfyui-fireredaudio-T8, so one install covers everything:

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

No model download needed for this node itself - it's pure bookkeeping. You only need the model when you feed the restored batch back into something that generates (like BatchRetry).

Gotchas

The default sandbox is real: a manifest outside ComfyUI/output will be refused unless you explicitly enable allow_external_manifest. That's annoying the first time, and it's correct the other ninety-nine times - it stops you from pointing the node at random JSON. Also: restoring doesn't validate that the generation parameters still match, only that the files exist (and optionally hash-match). If you changed your script and want to re-dub, resume won't catch that; re-running BatchDubbing with a changed script is the right move, and its fingerprinting will correctly regenerate only what changed.

In short: use it at the start of every session that touches a long batch. It's one input, zero risk, and it's the difference between "resume 83/100" and "redo everything."