AuK-Flash-ComfyUI

Native instruction-guided speech generation, reference voice conditioning and editing; experimental quality acceptance.

This is a converted model package for Jack Spooner’s ComfyUI fork, derived from tencent/AuK-Flash. It is not a new fine-tune.

Required code

The integration modifies ComfyUI core; this is not a drop-in custom_nodes package.

• Model integration commit
• Implementation · Offline converter
• Development branch · Pinned compatible reference snapshot 9d9053c5e
• Installation, node usage and acceptance details

For an existing ComfyUI installation, use the linked core changes and their prerequisites against a compatible base. The integration commit alone is not asserted to be an independently cherry-pickable patch. The pinned reference snapshot includes the required shared loaders, attention and memory changes. Follow the fork’s installation README and dependencies when reproducing that snapshot; update the application only while its queue is idle. Upstream ComfyUI is Comfy-Org/ComfyUI; its README describes the underlying node engine and model management.

Files and installation

Copy the model category folders into ComfyUI/models/, preserving their names and subfolders. Supporting pronunciation/configuration assets in this repository are part of the package.

File	Size
diffusion_models/auk_flash.safetensors	6.122 GB
text_encoders/auk_qwen25_omni_3b.safetensors	7.459 GB
vae/auk_audio_vae.safetensors	0.587 GB

Manifest/provenance files record the conversion and source revisions. SHA256SUMS and FILES.json describe this release. Filenames and model folders matter. No original source snapshot needs to be converted again.

Usage

Copy diffusion_models, text_encoders and vae into the corresponding ComfyUI model folders. Load the AuK diffusion model, CLIP type auk, and AuK VAE. Use the supplied direct-synthesis editor/API example first. Prompt: Based on the following description: "A warm conversational voice", generate speech content "Hello there.". Select your own reference.wav for the reference example. The optional MOSS helper workflow requires a separately downloaded MOSS model.

Example graphs are in workflows.

Verification and limits

Original-precision storage, BF16 transformer/conditioning and FP32 VAE execution were tested in bounded CPU/GPU generation and lifecycle comparisons. Listening quality, edit effectiveness and longer resource tests remain incomplete. This is an experimental converted release, not a general quality claim. Other GPU backends and quantized exports are unverified.

These statements describe the linked retained tests, not a new cross-platform acceptance suite. File size is not a minimum RAM/VRAM requirement. No inference-time downloads or external inference worker are required by these native model paths.

Conversion and provenance

Splits Tencent AuK transformer/VAE and combines AuK fusion weights with retained Qwen2.5-Omni-3B conditioning weights. Folds VAE weight normalization; omits unused Qwen vision/Talker/vocoder/logits and training-only components. Embeds tokenizer/config assets and retains the source bundle.

Source model revision: 575b92f0895f75180bf2cbd35f2e176c5732b8ed. See PROVENANCE.json, the conversion manifest and NOTICE. No training data, private inputs or generated user recordings are included.

License and attribution

Mixed terms: Tencent AuK transformer/VAE are MIT. The bundled Qwen2.5-Omni-3B-derived conditioning encoder is under Qwen RESEARCH LICENSE AGREEMENT, restricted to research/evaluation; commercial use requires permission from Alibaba Cloud. Built with Qwen. A copy of that agreement, the required Qwen Notice attribution and prominent conversion notices are included. Do not interpret Tencent’s MIT tag as licensing the Qwen component.

See LICENSE, NOTICE, and the retained upstream model card/source notices. Model authors and publishers retain credit; this is a third-party ComfyUI repackaging and does not imply endorsement. The weights’ terms and the separately linked ComfyUI code license are distinct.

The example workflow files are copied/adapted from the linked ComfyUI fork; its GPL-3.0 license is retained in LICENSES/COMFYUI-GPL-3.0.txt. This does not relicense the model weights.