Skip to content

Models & software

A record of the software, AI models, and third-party components that make up NK-Studio and its runtime engine. This list is part of the deliverable and will be kept up to date as the render pipeline is finalised.

Studio application

Component Description License
Unity (URP) Real-time 3D engine powering the studio interface and avatar preview Unity Software License
FMOD Studio Audio engine for voice playback, preview, and recording capture Firelight Technologies License
TextMesh Pro UI text rendering, included with Unity Unity Companion License

Runtime engine

Component Description License
Python 3.11 Runtime language for the local API and all background services PSF License
FastAPI HTTP API framework for the local runtime engine MIT
Uvicorn ASGI server running the runtime API BSD
SoundFile / libsndfile Audio file I/O for WAV handling BSD / LGPL-2.1
FFmpeg Video encoding — raw frames + audio to MP4 LGPL-2.1+

Voice generation (TTS)

Component Description License
Qwen3-TTS-12Hz-1.7B-Base Base TTS model (Alibaba / Qwen team); provides the zero-shot voice cloning foundation Apache 2.0
Custom trained talker model Character-specific voice model trained on top of Qwen3-TTS. Combines zero-shot reference cloning with a trained adapter for a more stable and consistent voice identity than zero-shot alone Proprietary
PyTorch (CUDA) GPU inference framework BSD

Lip sync

Component Description License
Montreal Forced Aligner (MFA) Forced phoneme alignment — timestamps phonemes against a transcript MIT
G2P (grapheme-to-phoneme) Generates pronunciations for words outside the phoneme dictionary; bundled with MFA MIT

AI render

Component Description License
ComfyUI Node-based workflow engine executing the AI render pipeline locally GPL-3.0
PyTorch (CUDA) Inference runtime for the AI render models BSD

Render pipeline

The AI render runs two sequential workflows on the local ComfyUI server. The studio injects the recorded take as input and retrieves the output video on completion.

Workflow 1 — Stylization. Transforms the low-resolution 3D capture into the final visual look of the avatar.

Stage Role
Input loader Loads the recorded take capture
Video generation model — to be finalised Core stylization pass; renders the avatar in its final visual style
LoRA / style adapter — to be finalised Character and look-specific fine-tuning applied on top of the base model
Output saver Writes the stylized video for use as input to Workflow 2

Workflow 2 — Lip sync refinement. Runs a dedicated lip-sync pass on the stylized output to produce accurate mouth movement in the final image.

Stage Role
Input loader Loads the stylized video from Workflow 1
Audio loader Loads the voice WAV captured with the take
Lip sync model — to be finalised Drives mouth movement in the stylized video against the voice audio
Post-processing — to be finalised Blending, compositing, or cleanup as needed
Output saver Writes the final rendered video

Note

Specific render models, their versions, and licensing will be completed here as the render pipeline is finalised.

Audio enhancement

Component Description License
Resemble Enhance Two-stage audio enhancement: denoising and bandwidth restoration (Resemble AI) Apache 2.0
PyTorch (CUDA) GPU runtime for the enhancement model BSD

Output and delivery

Component Description License
Topaz Video AI Post-render upscaling from pipeline output to 4K delivery quality Topaz Labs Commercial License