| component | role | license | how it is used |
|---|---|---|---|
| ComfyUI-TBDub (this repository) | integration | Apache License 2.0 | - |
TaoLiveAIGC/TBDub 19e18e4 |
upstream code (derived from X-Dub and DiffSynth-Studio, see its NOTICE.md) | Apache License 2.0 | installed separately by the administrator; imported unmodified, patched in memory (runtime/tbdub_patches.py) |
TaoLiveAIGC/TBDub b1ac4369 |
tbdub_student.safetensors, null_prompt_emb.pt, config.json |
Apache License 2.0 | downloaded by the administrator |
KlingTeam/X-Dub c84e9612 |
Wan2.2_VAE.safetensors |
Apache License 2.0 | downloaded by the administrator |
facebook/hubert-large-ll60k ff022d09 |
audio encoder | Apache License 2.0 | downloaded by the administrator |
MediaPipe 0.10.21 and Face Landmarker face_landmarker.task (float16/1) |
face detection and landmarks (CPU) | Apache License 2.0 | pip package + downloaded model |
| PyTorch, transformers, diffusers, safetensors, OpenCV, NumPy, SciPy, scikit-image, JAX, ... | runtime dependencies | their own licenses (BSD/Apache/MIT families) | installed by the administrator |
| ffmpeg / ffprobe | decoding, encoding, muxing | LGPL/GPL depending on the build | installed by the administrator; called as separate processes |
This repository contains no model weights, no upstream code and no ffmpeg binary.
runtime/tbdub_patches.py contains code written for this integration that follows upstream code closely in two
places: StreamingSafeOpen reads the safetensors file format (Hugging Face, Apache License 2.0), and the patched
causal-convolution forward of P4 repeats the padding logic of CausalConv3d from the Wan VAE in DiffSynth-Studio as
shipped in TBDub (Apache License 2.0). Process control, configuration and tests are adapted from the same author's
ComfyUI-LeapTalk / ComfyUI-MonarchRT / ComfyUI-CausalForcing (Apache License 2.0).
The demo videos and speech in the v0.1.0 release are AI-generated and come from the public ComfyUI-LeapTalk v0.1.0 release: the portrait was generated with stabilityai/stable-diffusion-xl-base-1.0 (CreativeML Open RAIL++-M) from a text description of a fictional person, the speech with hexgrad/Kokoro-82M (Apache License 2.0), and the talking video with LeapTalk. For the full-frame demo that video was placed on a flat 1280x720 background (a synthetic canvas). No real person and no private media are involved.