---
title: Bernini-R-S2V Talking Head
emoji: 🎙️
colorFrom: red
colorTo: purple
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
short_description: Speech-driven talking-head video (Bernini-R S2V, Wan2.2)
python_version: "3.12"
startup_duration_timeout: 1h
---

# Bernini-R-S2V — Talking Head

Speech-driven video generation demo for
[**rzgar/Bernini-R-S2V**](https://huggingface.co/rzgar/Bernini-R-S2V) — a Wan2.2
Speech-to-Video (S2V) model grafted onto ByteDance's Bernini-R renderer with
lip-sync from audio.

Upload a portrait image and a speech clip; the model animates the portrait into
a lip-synced talking-head video.

The inference path reproduces the author's ComfyUI reference workflow
(`ComfyUI-WanBerniniS2V` custom node + Wan2.2 S2V nodes), driven directly in
Python. It uses the FP8-scaled two-expert (high/low noise) checkpoints plus the
LightX2V 4-step distillation LoRAs for fast, ZeroGPU-friendly inference.

- Text encoder: `umt5_xxl_fp8_e4m3fn_scaled`
- VAE: Wan 2.1 VAE
- Audio encoder: `wav2vec2_large_english_fp16`
- Sampler: `dpmpp_2m_sde` / `sgm_uniform`, 4 steps, CFG 1.0