README.md

metadata

title: Bernini-R-S2V Talking Head
emoji: 🎤
colorFrom: red
colorTo: purple
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
short_description: Portrait + speech audio into a lip-synced talking video
python_version: '3.12'
startup_duration_timeout: 1h
models:
  - rzgar/Bernini-R-S2V
  - tolgacangoz/Wan2.2-S2V-14B-Diffusers

🎤 Bernini-R-S2V — Talking Head

Animate a portrait image + speech audio into a lip-synced talking video.

Bernini-R-S2V is a speech-driven video extension of Bernini-R, built on Wan 2.2 S2V. It adds single-speaker audio conditioning so a still portrait can be driven to speak or sing in sync with an audio track.

How it works

• Input: a clear front-facing portrait + a mono speech clip (best results at ~3–10 seconds of clear speech).
• Output: an MP4 of the portrait animated to speak, with the audio muxed in.
• Runs on ZeroGPU with the community WanSpeechToVideoPipeline (diffusers PR #12258).

Note on weights

The Bernini-R-S2V release ships ComfyUI single-file checkpoints (dual high/low noise Wan 2.2 MoE) plus a wav2vec2 audio encoder, with no diffusers-format export. The community diffusers Wan 2.2 S2V pipeline uses a single-transformer S2V architecture, so to deliver the same self-hosted image + audio → talking video experience this Space runs the base Wan 2.2 S2V-14B weights (tolgacangoz/Wan2.2-S2V-14B-Diffusers) that Bernini-R-S2V is finetuned from.