BonFrame LatentSync ONNX Runtime Pack
This repository contains the ONNX inference assets used by BonFrame's local
Windows and Linux LatentSync runtime. The application downloads one complete
pack from Model Zoo; inference then runs in Node.js with onnxruntime-node and
does not require Python or PyTorch.
Layout
1.5/: 256px distilled UNet graph and runtime metadata.1.6/: 512px distilled UNet graph and runtime metadata.common/: FP16 VAE encoder/decoder, Whisper audio encoder, mel filters, face detector, 106-point landmark model, and the official mouth mask.
The distilled checkpoints use the fixed four-timestep schedule recorded in
each version's runtime.json. They are not interchangeable with the original
teacher checkpoints.
The two large external-data blobs are distributed from BonFrame's Cloudflare R2 bucket. Model Zoo preserves the student filenames referenced inside each ONNX graph while downloading those blobs from stable canonical R2 object keys.
Upstream and licensing
- LatentSync source/model architecture: ByteDance, Apache-2.0.
- Whisper tiny audio encoder: OpenAI Whisper model components.
- VAE: LatentSync's configured Stable Diffusion VAE component.
- Face detection and landmarks: InsightFace-compatible ONNX assets used by LatentSync/FasterLivePortrait.
This pack is intended for use under the licenses and acceptable-use terms of the corresponding upstream projects. Users are responsible for obtaining consent for the people and voices represented in generated media.