AuK Model Loader X

Why you'd reach for it

AuK is a local speech model that does cloning, editing, conversion and separation from a text instruction - think "replace these words with those words in this recording" rather than "generate a voice". It's a diffusion-style audio model, not a TTS vocoder, which is why it has its own loader nodes rather than plugging into a text-encoder prompt.

AuK Model Loader X picks the diffusion checkpoint. There are two lines of it, and the choice between them is the first decision you make:

• Base - full quality, 32 steps with guidance. The one for results you'll keep.
• Flash - the distilled model. Four fixed steps, no guidance, ignores the sampling controls entirely. The one for iterating.

model_name also carries the weight format, because the file suffix says so: bf16 is the baseline, int8 or w4a4 are smaller. Keep the original filenames from the model repo - the loader identifies the variant from them.

attention is optional and defaults to auto, which hands the decision to ComfyUI's attention dispatcher and matches upstream inference. Choosing an explicit Flash or Sage backend is opt-in: they need CUDA with fp16/bf16, and they fall back to PyTorch SDPA when the kernel is unavailable, when the dtype is unsupported, or when an edit mask is present. So if you're specifically benchmarking attention backends, note that a masked edit step silently isn't a Flash/Sage step.

One output: WORKFLOWX_AUK_MODEL, which goes to AuK Instruction Encode and AuK Generate / Edit. Both connections matter - Base and Flash carry their own learned Qwen layer-fusion weights, so the model is required at the encoding stage, not just for sampling.

The precision dropdown will confuse you

precision controls the compute dtype, and its tooltip says exactly what to expect: auto means bf16 on GPU and fp32 on CPU. It does not change how the weights are stored - that comes from the checkpoint file. So you can load an int8 file and select bf16 compute, and that's a legitimate combination, not a mistake. The naming is unfortunate and it catches people out.

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

Those four are the pack's optional AuK dependency set - the base pack only needs PyWORLD, numpy and scipy, but the AuK nodes need a current Transformers (5.3.x or a compatible 5.x). Nothing here pins or replaces your PyTorch build.

Then the weights. Download the converted checkpoints from drbaph/AuK-comfyui on HuggingFace into the normal ComfyUI folders:

ComfyUI/models/diffusion_models/   auk_base_bf16.safetensors   (3.06 GB)
ComfyUI/models/text_encoders/      qwen_omni_bf16.safetensors  (8.07 GB)
ComfyUI/models/vae/                auk_vae.safetensors         (0.64 GB)

Download only the weights - Base and Flash configs are bundled in the pack under auk/assets/, and the Qwen tokenizer and config files live in the pack too, loaded locally from there. Put no sidecars in the model folders. Restart ComfyUI after dropping new files in.

Where it goes wrong

Quantized files need a recent enough ComfyUI. The int8 and w4a4 formats depend on ComfyUI / Comfy Kitchen builds exposing the right quantized ops (int8_tensorwise with ConvRot, convrot_w4a4, asym_w4a8_int8). Those exist in current builds; they should not be assumed for an old one. BF16 is the baseline and always works, which is why it's the right first download even though it's the biggest.

The missing-model dropdown on Windows. ComfyUI lists subfolder models with backslashes, so a value in a loaded workflow can show as missing. Re-select the checkpoint from the dropdown and move on - the file is fine.

VRAM is audio-length dependent. Peak use scales with clip length and the attention backend, not with a fixed floor. One reported measurement with a quantized model and encoder pair was around 8 GB peak; treat that as an observation, not a requirement. Longer clips cost more memory and more time.