v1.1.0

What's Changed

Features

• Scale TTS seconds to the reference voice's measured pace (#5): AuK Prompt Enhance now derives the seconds output for voice cloning and other TTS requests from the reference clip's actual speaking pace (weighted transcript length over clip duration) when both the clip and its Whisper transcript are connected — fast talkers get less time, slow talkers more — preventing trailing silence and clipped or rushed speech. Falls back to the generic per-character estimate when no reference clip or transcript is available.

Fixes

• Support junctioned model subfolders (#6): model loaders no longer reject checkpoints reached through Windows directory junctions or symlinked subfolders inside the registered model categories (ValueError: Model path leaves its registered model directory). Path-traversal protection is retained.

Full Changelog: v1.0.1...v1.1.0

v1.0.1

ComfyUI-AuK v1.0.1

This update expands the local Qwen2.5-Omni-3B Prompt Enhance path and refreshes the bundled workflows and documentation.

What changed

• Prompt Enhance audio input: accepts the source/reference audio and returns prepared_audio for Instruction Encode
• ASR-aware editing: connect the Whisper transcript to context so replacement, insertion, removal and lyric-edit durations can account for the source content
• Task-aware duration: handles source length, speech edits, speed changes, emotion, nonverbal sounds and TTS text length
• Whisper preparation: applies the upstream whisper RMS targets for whisper conversion tasks
• Local operation: canonical task commands still run through the bundled Qwen2.5-Omni-3B language head without an external LLM API
• Updated workflows: Prompt Enhance, Whisper, Instruction Encode and Generate / Edit are wired together with Markdown guidance
• No combined duration cap: removed the obsolete source-plus-output 30-second workflow warning
• Documentation parity: synchronized the English and Chinese READMEs and updated the single task guide

Editing observations

FP32 and BF16 comparisons against the Hugging Face demo remained inconsistent on the tasks marked unstable in the README. In testing, difficult edits often worked better when output duration was set 0.2–0.5 seconds shorter than the source; for example, 5.5 seconds for a 6-second input. Some unstable tasks also succeeded more often with the Chinese cookbook instruction than with its English version.

Validation

• 43 regression tests passed
• 9 contract and workflow tests passed; 2 optional parity/kernel checks skipped

README · Task guide · 中文说明

v1.0.0

ComfyUI-AuK v1.0.0

Local AuK speech generation, voice cloning and editing for ComfyUI — 8 nodes, no cloud calls, no mid-run downloads.

What is in this release

• 8 nodes: Model Loader, Encoder Loader, VAE Loader, Instruction Encode, Generate / Edit, Instruction Builder, Whisper Transcribe, Prompt Enhance
• Base and Flash: same task set. Flash is the distilled model (4 fixed steps, no guidance) for speed; Base runs 32 steps with guidance for quality
• Quantization: BF16, INT8, W4A8 (converted with Comfy Kitchen) plus FP32 originals — pick by VRAM budget, mix and match with encoders
• Local Prompt Enhance: maps loose requests onto canonical task instructions using the encoder's own language head — no API keys
• Whisper Transcribe with opt-in download_if_missing (fetches openai/whisper checkpoints into models/whisper/ once, offline afterwards)
• Instruction Builder with canonical templates for all 26 tasks

Model downloads

All 12 checkpoints: drbaph/AuK-comfyui
Diffusion (Base/Flash × FP32/BF16/INT8/W4A8) → models/diffusion_models/, Qwen encoder (BF16 8.07 GB / INT8 4.67 GB / W4A8 3.18 GB) → models/text_encoders/, VAE (0.64 GB) → models/vae/. Tokenizer/config sidecars ship with the node pack.

Requirements

• ComfyUI with Transformers 5.3.x
• INT8/W4A8 checkpoints need Comfy Kitchen builds exposing int8_tensorwise (ConvRot) or asym_w4a8_int8; BF16/FP32 have no extra requirements

Known limitations (model, not the node)

Prompt adherence varies by task — voice cloning, TTS, volume, enhancement and separation are the reliable end; replace/insert speech, lyric edits and whisper conversion often need several attempts. Full tested ranking: Prompt adherence. Lyric editing additionally requires acapella input and 1–2 word changes with phonetically similar replacements.

Docs

README · Task guide · 中文说明