Projects

• 
  Style-Talking 06/2026 · Zero-shot speaking-style clone for video-driven avatars

  A zero-shot audio-driven digital human system that builds speaking-style prompts from a reference video, predicts LivePortrait expression motion from wav2vec audio features, and renders talking videos while preserving the original non-face regions.

  Project Code
• 
  Voice Studio 05/2026 · Local voice creation workspace for Apple Silicon Macs

  A local-first macOS voice creation app for premium voice generation, voice cloning, voice design, dialogue generation, and DeepSeek-assisted dialogue rewrite. It targets novels, audio drama, narration, character dialogue, and multi-role audio export workflows.

  Voice Studio 0.01 dialogue audio sample about 9 min 52 sec

  Multi-role dialogue generated in Voice Studio, covering role voice binding, per-line synthesis, and final full-audio merge.

  Show synthesis text excerpt

  旁白: 航行又持续了两个多小时，如果在三维，太空艇已经航行了二十万千米左右。突然间，硬币大小的“魔戒”顶天立地地出现在前方。
  卓文: 紧急转向！
  旁白: 卓文用目光操纵太空艇紧急转向，使撞向环箍的太空艇从“魔戒”的圆环中穿过。从艇中看去，像是通过了太空中一道巨大的拱门。太空艇全力减速，然后返回，悬停在距“魔戒”的圆心不远处。
  关一帆: 我看不出后六个数的规律。
  韦斯特: 那就假设没有规律。前六个数是我们发送的，最可能的含义就是表示我们，后六个数没有规律且不断出现不同的组合，可能代表“一切”，我们的一切。
  魔戒: 我是墓地。
  关一帆: 谁的墓地？
  魔戒: 这个墓地的建造者的墓地。
  

  Full demo page

  Code Download
• 
  IConFace 05/2026 · Unified reference-aware face restoration

  A unified reference-aware and no-reference face restoration framework. It uses identity-structure asymmetric conditioning so one checkpoint can exploit same-identity references when available and fall back to degraded-only restoration when references are absent.

  Project arXiv Code
• 
  flux-restoration 04/2026 · FLUX-based restoration adapter

  An inference-oriented release package for unified blind and reference-based face restoration on FLUX.2-klein-base-4B. It includes example outputs, pretrained model instructions, and a practical restoration workflow for reuse.

  Code
• 
  ComfyUI-Flux-FaceIR 04/2026 · Face restoration workflow for ComfyUI

  A ComfyUI extension for aligned face restoration and full-image restoration. It supports face detection, alignment, parameter export, restoration, and paste-back so the model can be used inside visual node workflows.

  Code
• 
  flux-inference 04/2026 · FLUX inference acceleration stack

  A FLUX.1-dev inference optimization stack covering PyTorch baseline execution, model export, ONNX Runtime, and TensorRT backends. The project focuses on practical image generation acceleration experiments.

  Code
• 
  3d-viewers 05/2026 · Browser viewers for 3D vision and digital-human models

  Lightweight browser-based viewers for inspecting 3D vision and digital-human assets. It covers FLAME, parametric models, Gaussian avatars, topology, joints, bindings, and motion before fitting or training.

  Code
• 
  EfficientTime 05/2026 · Local-first AI daily planning assistant

  A local-first macOS productivity app for turning rough task notes into editable daily schedules. It keeps the current task visible and supports local reminders plus AI-assisted planning drafts.

  Code