AuK editor assembly X (internal)

What this node is

One input: gap_seconds. Two outputs: audio and segment_report. That's the whole node - and it's the one that actually hands your edited recording to the rest of the graph.

AuK editor assembly X (internal) is the assembly step of the segment editor in AuK Chained Clone X. When you've browsed the lines, regenerated the two that came out wrong, added a missing one, and hit Finalize, this is what runs: it takes the retained takes, joins them with the current gap_seconds, and emits the result. The pack's README lists it, along with the review step, among the internal queue targets that support the segment editor and aren't intended for direct workflow use. Don't add it to a graph.

Why it's separate from the generator

Because the two have different jobs, and conflating them is what makes review loops feel broken.

Regenerating a line means running the model again on one script line - expensive, seeded, and only meaningful per line. Finalizing means taking what you already have and producing a file - cheap, deterministic, and graph-level.

If Finalize regenerated speech, then every press of the button would risk changing lines you were happy with, and your carefully curated takes would be pointless. The chained clone node's documentation is explicit that Finalize does not regenerate speech or run upstream model/text-generation nodes. It assembles, and only assembles. Its retained takes may even have deliberately different voices or settings - that's allowed, and it's the point of a review loop.

The gap_seconds input is the only thing you get to decide at assembly time, and it sits between finished lines exactly as it does during a normal batch run: silence between lines, no crossfades, no normalization, no trimming.

The behaviour that follows from this

Understanding the split explains several things you'd otherwise read as bugs:

Finalize is instant, and that's not a caching illusion. Nothing is being computed except the join.

Changing a setting doesn't change your existing takes. Settings affect future generations. Retained audio keeps the voice, seed and duration it was recorded with. If you change the model and hit Finalize, you get the old recording joined - genuinely, not as a bug.

It needs every slot. Assembly won't emit an incomplete recording as a success. A line whose generation failed stays as an ungenerated slot to retry, and Finalize refuses until it has audio.

Refresh can reconnect; a restart cannot. The takes live in a temp area with a 24-hour inactivity window and are cleared by a server restart or by switching the editor off. The regeneration state isn't durable, so finish a review session rather than leaving it overnight.

segment_report is the second output and it's worth wiring to a text display while you're learning the node's behaviour - it describes the assembled segments.

Install

Nothing extra is needed for this node - it's part of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/haroonaslam/WorkflowX-Configurator
cd WorkflowX-Configurator
pip install "transformers>=5.3,<6" safetensors PyYAML tqdm

Restart ComfyUI and hard-refresh the browser. The generator half of the workflow needs the AuK weights from drbaph/AuK-comfyui in models/diffusion_models/, models/text_encoders/ and models/vae/ before any of it produces sound.

Where it goes wrong

Adding it manually. If you drag this onto a canvas hoping to join two audio clips, you'll get a node with a gap_seconds widget and nothing to feed it. There's no audio input. Use a core audio-concatenation node instead; this one reads the editor's internal takes, not sockets.

Assuming it re-renders. "I changed the script and Finalized and nothing changed" is expected: Finalize assembles retained takes, and the script only affects what gets generated next. Run a fresh full batch if you want the new text.

Losing takes to a restart. Clear editor, switching the editor off, a server restart, or 24 hours idle all discard them. There's no previous-take history and no automatic recovery - save what you like downstream by Finalizing and writing the output somewhere before you shut ComfyUI down.