AuK Generate / Edit X

Why you'd reach for it

This is where AuK actually makes sound. Everything upstream - the model, the encoder, the VAE, the instruction - exists to feed this node, and its single AUDIO output is a mono 24 kHz waveform you hand to core Preview Audio or Save Audio.

The thing to internalise early: AuK does not use KSampler. It has a dedicated model socket, its own multimodal conditioning and its own sampling loop, so a native KSampler wired to it will do nothing useful. This node is the sampler.

How it works

model comes from AuK Model Loader X (Base or Flash), conditioning from AuK Instruction Encode X, and vae from AuK VAE Loader X. Then the four settings that decide the result:

seconds is output length, and its behaviour depends on your task. 0 matches the source audio's duration - that's the right starting point for every editing task where you're swapping or changing words at roughly the same length. Text-only generation needs a positive value, and roughly 2–3 spoken words per second is the working estimate. If the last words of your clip get clipped, this number is almost always why.

seed drives the diffusion noise and reference-latent sampling, and each batch item uses seed + index.

steps, guidance and sway are Base-only settings. The tooltips are blunt about it: steps is Euler steps for Base, 32 is the upstream default and 16–50 is the sensible band; guidance is AuK's own guidance strength, not KSampler CFG, default 2; sway defaults to -1 with 0 disabling it. Flash ignores all three and always runs its fixed four-step schedule with no guidance. Nothing warns you - the widgets just sit there doing nothing while you're on a Flash checkpoint.

The sway field, briefly

sway is an inference-time sampling trick, not a knob invented for this pack. The name comes from Sway Sampling, introduced with F5-TTS - an SDE-timestep weighting that improves output quality and speed without retraining, and it has since turned up in audio models generally. Here, -1 is upstream's default and 0 turns it off. The honest advice from the node's own tooltip is to leave it alone unless you're experimenting; it's the kind of field where a small change produces a difference you'll struggle to attribute.

Tuning for hard edits

The task guide has one genuinely counterintuitive recommendation and it's worth repeating: for difficult edits, setting the output about 0.2–0.5 seconds shorter than the source improves instruction adherence. A six-second source, try 5.5. That means disconnecting AuK Prompt Enhance X's seconds output and typing the duration manually, since the enhancer's job is to compute a task-aware duration for you.

The other thing to set expectations about: AuK's own documentation opens with a warning that prompt adherence is not strong - cloning, TTS and some edits work, others are unstable, and that's the model rather than the integration. Expect to iterate, and listen to every result.

Install

ComfyUI Manager → search WorkflowX Configurator, or:

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

You'll need three files in place before this node does anything: an AuK checkpoint in models/diffusion_models/, qwen_omni_*.safetensors in models/text_encoders/, and auk_vae.safetensors in models/vae/ - all from drbaph/AuK-comfyui, weights only, no sidecars. Restart ComfyUI after adding them.

Where it goes wrong

Sampling settings that do nothing. If you're on Flash and fiddling with steps, that's the wasted afternoon. Check which checkpoint the model loader has before you tune.

A batch of silence or a truncated line. Usually seconds - too short for inserted words, or 0 on a text-only generation, which can't match a source that doesn't exist.

Long clips cost memory and time. There's no combined source-plus-target duration cap in this integration (it follows upstream's Python tooling rather than the standalone wrapper's 30-second budget), so nothing stops you from queueing something that makes the card unhappy. Audio length is the main driver of peak VRAM.