AuK · T8star-Aix Native ComfyUI Nodes

Run AuK speech generation and editing directly inside ComfyUI

中文说明 · Model repository · Complete Windows package / overseas

This is a standalone ComfyUI V3 custom-node package. It loads AuK, AuK-Flash, and Qwen2.5-Omni-3B directly in the ComfyUI process. It does not require AuK Local, a server at 127.0.0.1:7860, or a service token.

This repository publishes only the standalone ComfyUI node package. There is one separate AuK Local one-click package. Each can be installed and run independently; they only share model sources and documentation links.

Nodes

• AuK Model Loader defaults to the higher-quality AuK Base and can switch to the speed-oriented AuK-Flash. ComfyUI manages staged loading and offloading of the VAE, Qwen encoder, and DiT.
• AuK Generate / Edit exposes 17 Chinese task entries that cover all 16 upstream low-level tasks. Target-speaker extraction is a content-based entry for speaker separation. The node returns standard ComfyUI AUDIO, the final instruction, run metadata JSON, and the resolved target duration.
• AuK Audio Trim / Duration cuts by start/end seconds while preserving sample rate and channels. It returns cropped AUDIO, its exact FLOAT duration, and a trim description. An end of 0 means the end of the source.

Tasks include instruction TTS, zero-shot voice cloning, speech and lyric editing, pitch/speed/volume/emotion/timbre editing, de-accenting, nonverbal editing, whisper conversion, enhancement, quality repair, speaker separation, vocal extraction, and target-speaker extraction.

Install

ComfyUI Manager

Search for AuK · T8star-Aix in ComfyUI Manager, install it, and restart ComfyUI. Confirm that Manager offers version 2.0.7 or later; if Registry processing still shows an older release, use the Git installation until the new version becomes active.

Git

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-Auk-T8
cd Comfyui-Auk-T8
python -m pip install -r requirements.txt

Install dependencies with the same Python interpreter that runs ComfyUI. The requirements do not install or replace PyTorch or TorchAudio.

Models

Download the model files from t8star/Auk-Comfy and keep this layout:

ComfyUI/models/auk/
├── AuK-Flash/
│   ├── auk_flash.safetensors
│   ├── vae.safetensors
│   └── config.yaml
├── AuK/
│   ├── auk_base.safetensors
│   ├── vae.safetensors
│   └── config.yaml
└── Qwen2.5-Omni-3B/
    ├── config.json
    ├── model-00001-of-00003.safetensors
    ├── model-00002-of-00003.safetensors
    ├── model-00003-of-00003.safetensors
    └── the remaining repository files

You can also run the downloader with ComfyUI's Python from the node directory. Both model variants require Qwen:

python download_models.py --variant flash
python download_models.py --variant base
python download_models.py --variant all

The downloader is pinned to the tested Hugging Face snapshot recorded in MODEL_MANIFEST.json and verifies SHA-256 by default; use --skip-sha256 only when you intentionally want a faster size-only check. The loader also searches every path registered as auk in extra_model_paths.yaml; the AuK checkpoint and Qwen folder may be stored in different registered roots.

Flash plus Qwen requires about 18.7 GB. Both AuK variants plus Qwen require about 25.5 GB.

Run

1. Drag the JSON for the required task from example_workflows into ComfyUI. The folder contains an executable workflow for every one of the 17 entries.
2. AuK Model Loader defaults to AuK Base. Prefer Base for emotion, accent, timbre, nonverbal, whisper, and repair tasks; Flash is intended for fast previews.
3. Select a task in AuK Generate / Edit. The node displays its official input requirement, field meanings, example, and warning. For source/reference tasks, connect Load Audio → AuK Audio Trim / Duration → AuK Generate / Edit. Trim long originals before generation. Set both crop values to 0 to pass the full source.
4. Queue the workflow. AuK-Flash always uses NFE=4 and CFG=0; Base uses the advanced sampling controls.

Automatic duration adaptation: click ↻ 自动适配时长 in the node guide, or select 自动适配（按任务规则） in the visible duration mode widget (the default). Editing uses the actual cropped input and task rules, ignoring a stale target value or a connected Float. No separate Float node is needed. Instruction TTS and voice cloning estimate duration from text. This estimates the spoken length from the target text and prevents a short sentence from continuing into AuK's internal no-reference marker when a much longer duration is requested. Select 手动指定 for exact TTS timing; editing rules remain automatic. The previous automatic TTS option is kept for existing workflows. Text estimates over 30 seconds now report an error requesting shorter text or separate segments instead of silently truncating. The Seed widget uses ComfyUI's standard randomize after generation mode by default; switch its control mode to fixed to reproduce a result. The metadata output records the actual seed, requested duration, resolved duration, and duration mode.

Pitch, volume, timbre, de-accent, and whisper match the effective speech duration measured by official VAD. Enhancement, quality repair, and separation preserve the full input length. Speed uses effective speech duration / multiplier. Emotion uses the official factors: 1.22× for sad, 1.16× for fearful, and 1.06× for the other supported emotions. Speech and lyric edits estimate the result from the text added or removed. Nonverbal edits add or remove the official event duration. These automatic rules ignore a stale target-duration widget, so a 48-second source trimmed to four seconds is validated as the actual four-second node input.

Before inference, the node uses official Silero VAD to measure the unpadded speech interval and adds 0.1 seconds only around the model input. Whisper conversion uses the official -44.47 LUFS target; lyric and music-separation outputs use the official -14 LUFS downward limiter and 0.95 peak ceiling. All tasks reject nonfinite audio. If a model peak exceeds the standard audio range, gain is reduced to a 0.99 peak ceiling to prevent FLAC export clipping; ordinary levels and quiet whispers are preserved. Metadata records this protection and its gain in dB. Lyric editing requires clean isolated a cappella solo vocals. Emotion, de-accenting, and whisper conversion require ordinary spoken speech. Singing is not a valid test source for those tasks, and already-standard speech is not a valid de-accenting test. Model loading and generation report native ComfyUI progress through configuration, Qwen, VAE, AuK, text/reference encoding, sampling, and decoding stages. Speed supports 0.5, 0.75, 1.25, 1.5, or 2.0; pitch uses +1/+2/+3 or -1/-2/-3 semitones, and volume uses +5/+10/+15 or -5/-10/-15 dB.

Source/reference audio and generated output are each limited to 30 seconds independently. They are not added together, so a 30-second input may produce a 30-second output. Slowing speech or changing emotion may resolve to more than 30 seconds; trim the source first in that case. VAD removes some leading/trailing silence in speech tasks. Metadata separately records the cropped raw input, prepared input, and resolved output duration. CPU mode is available for compatibility testing but is very slow; NVIDIA CUDA with bf16 is recommended.

Version 2.0.7 adds a clickable automatic-duration button and a standalone audio-trim node, connected in every audio workflow. Input validation uses exact sample counts for the 30-second limit, rejects invalid sampling parameters, and exposes the resolved target duration. Restart ComfyUI and reload the browser after upgrading. Add the crop node to an old graph or import a new workflow.

Version 2.0.6 fixes validation against stale pre-trim durations and applies independent 30-second limits to input and output. It completes the official duration and preprocessing rules, adds quality repair, and defaults to AuK Base. The package includes drag-and-drop workflows for all 17 task entries.

Version 2.0.6 uses official unpadded Silero speech duration and LUFS handling, validates nonverbal/quality/speaker-order requests before inference, and displays the matching official guide inside the node.

Standalone local package

Complete AuK Local Windows package (models and Python included): Hugging Face overseas download — AuK-Local.rar, 20.37 GB (18.97 GiB). Fully extract it and double-click AuK-Local.exe in the AuK-Local folder. GitHub AuK Local Release ZIPs contain program updates only.

The AuK Local one-click package remains a separate light-themed web workstation with its own Python runtime, model management, task history, and launch scripts. It is no longer a runtime prerequisite for these ComfyUI nodes.

Compatibility

• ComfyUI >=0.23.0 with the V3 custom-node API and staged model-management interfaces.
• Python >=3.10.
• Verified with PyTorch/TorchAudio 2.7.x + CUDA 12.8 and a 24 GB NVIDIA GPU.
• Model construction uses substantial host memory; 48 GB or more system RAM is recommended.
• Output is 24 kHz float audio.

Links

• Bilibili
• YouTube
• API
• Online AI apps
• Standalone local package
• AuK-Comfy models
• Hugging Face profile
• Original AuK project

License

The node code is released under the MIT License. Model files retain the licenses included by their upstream repositories.