fq

data in modelscope for more convenient to download?

3

[modelscope](https://www.modelscope.cn/home)

as the v0.2.0 example. the obj file?

deploy sadtalker with trition by cuda optimization can see here. real time.

19

[SadTalkerTriton](https://github.com/foocker/SadTalkerTriton) 25 fps , 84 frame, cost 2.4s。3090ti

Live Speech Portraits: Real-Time Photorealistic Talking-Head Animation. i mean, the method seems be very familiar. and inference fps of 3090ti is?

for more stable generator. and for speed infer, can reference (lora, lcm) + [trt](https://github.com/NVIDIA/TensorRT/blob/release/10.5/demo/Diffusion/flux_pipeline.py)