ComfyUI_FunCineForge

ComfyUI_FunCineForge

Fun-CineForge: A Unified Dataset Pipeline and Model for Zero-Shot Movie Dubbing in Diverse Cinematic Scenes

Update

• 修复双人对话口音不对齐问题，调整双人对话的推理逻辑，以demo的代码为基础，切分片段，而非infer里的单jsonl 推理全部，按理demo是对的，因为参考音频可以区分男女；

• 修复音频口型不对齐的问题，注意输入视频必须是25帧

1.Installation

In the ./ComfyUI/custom_nodes directory, run the following:

git clone https://github.com/smthemex/ComfyUI_FunCineForge

2.requirements

• 注意安装pyannote.audio 库引发numpy 版本提升到2.0以上，记得装回低版本

pip install -r requirements.txt

3.checkpoints

• ALL FunCineForge FunCineForge
• speech_fsmn_vad_zh-cn-16k-common-pytorchlinks
• speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorchlinks
• speech_campplus_sv_zh-cn_16k-commonlinks
• punc_ct-transformer_zh-cn-common-vocab272727-pytorchlinks
• segmentation-3.0
• 去我云盘拉取全部模型的zip吧,解压到funcineforge目录下（结构如下图）

├── ComfyUI/models/   # 13G
|     ├── funcineforge/llm   #注意 没有模型库那样的子目录，直接是config和模型文件，其他几个一样
|          ├── config.yaml
|          ├── mp_rank_00_model_states.pt
... 

4 Example

5 Citation

@misc{liu2026funcineforgeunifieddatasettoolkit,
    title={FunCineForge: A Unified Dataset Toolkit and Model for Zero-Shot Movie Dubbing in Diverse Cinematic Scenes}, 
    author={Jiaxuan Liu and Yang Xiang and Han Zhao and Xiangang Li and Zhenhua Ling},
    year={2026},
    eprint={2601.14777},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
}