关于我的测试结果，不是很理想？

[zhanghongyong123456]

我的驱动视频： 预处理：python scripts/extract_kps_sequence_and_audio.py \ --video_path "./test_samples/short_case/10/gt.mp4" \ --kps_sequence_save_path "./test_samples/short_case/10/kps.pth" \ --audio_save_path "./test_samples/short_case/10/aud.mp3"

https://github.com/tencent-ailab/V-Express/assets/48466610/60db96fb-1841-4de3-acf2-430407236a4a

我的参考图像：（截图 512x512）

我的结果： 运行脚本： python inference.py \ --reference_image_path "./test_samples/short_case/tys/ref.jpg" \ --audio_path "./test_samples/short_case/tys/aud.mp3" \ --kps_path "./test_samples/short_case/tys/kps.pth" \ --output_path "./output/short_case/talk_tys_fix_face.mp4" \ --retarget_strategy "fix_face"

https://github.com/tencent-ailab/V-Express/assets/48466610/b4ce5263-9eb1-4478-b126-36859346912f

我不确定我哪里有问题，希望指点一下

[FurkanGozukara]

i planned to make a gradio app for this but this result looks very bad

[tiankuan93]

1. Our model is trained using English audio, and our audio feature extractor is also trained in English, so our model will perform more consistently on English audio for now. Other languages may yield some reasonable results, but it will require some experimentation with the parameters.
2. For the mode of _"fixface", we provide parameters to adjust the effect of the audio. We also commit the default parameters in the new commit.

• We get the same results if we use the reference_attention_weight=1.0 and audio_attention_weight=1.0 parameters.
• We get results bellow if we use the reference_attention_weight=0.95 and audio_attention_weight=3.0 parameters.

https://github.com/tencent-ailab/V-Express/assets/19601425/d88f9a9b-cc06-4476-b997-82fcb88e57d4

• If we crop the reference image more properly and use English audio, we get the following results.

https://github.com/tencent-ailab/V-Express/assets/19601425/79e4ddfe-e9a1-4ba1-b4ce-294023a0f1ab

https://github.com/tencent-ailab/V-Express/assets/19601425/4d4a04e8-d412-40a8-a370-1fc603addf6d

[FurkanGozukara]

@tiankuan93 thanks huge info do you plan to make a gradio demo app? or we have to make ourselves

[tiankuan93]

@tiankuan93 thanks huge info do you plan to make a gradio demo app? or we have to make ourselves

@FurkanGozukara We have no plans to implement a gradio demo app for now, thank you for your interest.

[FurkanGozukara]

@tiankuan93 thanks huge info do you plan to make a gradio demo app? or we have to make ourselves

@FurkanGozukara We have no plans to implement a gradio demo app for now, thank you for your interest.

Thanks then hopefully I will do myself and publish

I hope you don't change much with consistency Lora so I can implement that too

[tiankuan93]

@tiankuan93 thanks huge info do you plan to make a gradio demo app? or we have to make ourselves

@FurkanGozukara We have no plans to implement a gradio demo app for now, thank you for your interest.

Thanks then hopefully I will do myself and publish

I hope you don't change much with consistency Lora so I can implement that too

Consistency Lora only reduces the number of steps for inference and doesn't change much else.

[cantonalex]

@zhanghongyong123456 what retarget strategy did you use for your first video? how did you get the output video uncropped to the reference face

EDIT: oh thats your source video. bummer..

[boboji21]

牛逼A