AuK Model Loader

AuK is a speech generation and editing model - text-to-speech, voice cloning, and a long list of "edit this recording" tasks (pitch, speed, volume, lyric swaps, denoise, speaker separation). If that's the job you came for, this is the node where the workflow starts. It loads the AuK checkpoint and hands you an AUK_MODEL socket.

The first thing to understand: that socket is not a MODEL. You cannot feed it to a KSampler and you cannot drive it with CLIP Text Encode. AuK has its own sampling loop and its own multimodal conditioning, so the whole sampler side of the graph is replaced by the pack. If you're expecting a drop-in for a normal diffusion graph, you'll be confused for about ten minutes and then it clicks.

Base or Flash

Both variants do the same tasks. Base is the full model: 32 Euler steps, guidance 2, sway -1 by default, and you get to fiddle with all three. Flash is the distilled one - four fixed steps, guidance disabled, and the pack ignores your steps, guidance and sway values entirely. Same reason Lightning and Turbo checkpoints behave differently from their parents: distillation collapses the trajectory into fewer, bigger jumps, which buys a lot of speed and costs some quality. If you're iterating on wording, run Flash. When you want the take you keep, run Base.

The inputs that matter

• model_name - the dropdown is your ComfyUI/models/diffusion_models/ list. Drop the .safetensors file there. The filename suffix tells you the weight format: bf16 is the baseline, int8 and w4a8 are smaller. In sizes: 6.12 GB fp32, 3.06 GB bf16, 1.55 GB int8, 0.88 GB w4a8. That's disk size, not VRAM.
• precision - compute dtype only, not storage. auto means bf16 on GPU and fp32 on CPU. The actual weight format is read out of the checkpoint's own metadata, so picking bf16 here does not magically upgrade an int8 file.
• attention (optional) - auto uses ComfyUI's attention dispatcher. You can force flash_attention or sageattention, but they need CUDA fp16/bf16 and fall back to PyTorch SDPA when the kernel is unavailable or when an edit mask is present. So an edit task may quietly run SDPA no matter what you selected. sdpa is the honest baseline if you're comparing output.

One output: AUK_MODEL, wired into both AuK Instruction Encode (.model) and AuK Generate / Edit (.model). Both need it - the model carries the learned Qwen layer-fusion weights used at encode time, which is why the encoder node asks for it too.

Install

ComfyUI Manager → search the pack title ComfyUI-AuK, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Saganaki22/ComfyUI-AuK
cd ComfyUI-AuK
python -m pip install -r requirements.txt

That pulls transformers>=5.3,<6, safetensors, PyYAML and tqdm. It does not touch your torch build. Then grab a checkpoint from drbaph/AuK-comfyui on Hugging Face - those are the converted files the pack expects, and only they carry the variant metadata the loader reads. Restart ComfyUI afterwards.

Where people get burned

Quantized checkpoints need quantized-op support in ComfyUI itself. The int8/w4a8 files use formats (int8_tensorwise with ConvRot, convrot_w4a4, asym_w4a8_int8) that only exist in recent ComfyUI / Comfy Kitchen builds. If yours doesn't have them you get a blunt error naming the format and telling you to update or use BF16 - that's the node being helpful, not broken. troubleshooting.md's quant ladder is the background: INT8-ConvRot is the format that gave 20/30-series cards a real fit lever, and w4a8/w4a4 are the four-bit-weight relatives of it.

Keep the original filenames. The loader identifies Base vs Flash from checkpoint metadata, and for unconverted files it falls back to the names auk_base.safetensors / auk_flash.safetensors. Rename one to something cute and you get "Cannot identify this AuK checkpoint."

Windows subfolder paths. ComfyUI stores subfolder model names with backslashes, so a workflow loaded from someone else can show the checkpoint as missing even though the file is right there. Re-select it in the dropdown.

Set expectations on adherence. The author's own README opens with a warning: this model's prompt adherence is uneven. TTS, cloning, volume, enhancement and singing extraction are reliable; insert/replace speech, lyric edits and whisper conversion land in the "less stable" pile. Budget two or three seeds for those, and don't conclude your setup is broken.

The author here is Saganaki22, who has a habit of being first with local ports of new audio models (Zonos2 TTS, Higgs v3 TTS) and quantizing them for consumer cards. That's the flavor of this pack: fast, capable port, with the llama-3-shaped caveats spelled out in the README rather than hidden.