rzgar

/

Bernini-R-S2V

like 23

thank u for the complement and you're welcome,
achieving very high reference consistency with I2V is difficult, but we can avoid making it worse.

• Try not to use LoRAs. If you have to, on the low-noise use low strength (0.5 or 0.65 max), most Wan2.2 LoRAs effecting the input image at anything higher than 0.6.
• Use the model’s native resolutions like 1280x720 or 480x832 instead of square or random sizes.
• Higher precision on the low-noise (fp16 or fp32) while using fp8/fp16 on high-noise also helps maintain details. Most importantly, use a sharp, high-quality input image and avoid heavily compressed JPGs.

If the speaker's voice identity is not that important, you could use the LTX-2.3 like the screenshot. import the Wan video and add speech to it, in my testing denoise the video at 0.25 is the sweet spot for lip sync and preserving the facial identity

WAN input video

LTX2.3 boasts impressive audio capabilities and appears more convenient than S2V;
it eliminates the need to prepare local audio files in advance and can generate multi-person conversations based on prompts.
Combining Bernini for video generation with LTX2.3 for audio or voiceover is an excellent use case.

@rzgar bernini + LTX2.3 Lip-sync Phoneme Mapping and Workflow

https://github.com/princepainter/ComfyUI-PainterNodes
https://github.com/princepainter/ComfyUI-PainterNodes/blob/main/workflows/

Official models are required; this node does not support KJ model loading.

bernini+ltx2.3

@rzgar bernini + LTX2.3 Lip-sync Phoneme Mapping and Workflow

https://github.com/princepainter/ComfyUI-PainterNodes
https://github.com/princepainter/ComfyUI-PainterNodes/blob/main/workflows/

Thanks for the links, tested it and works great!