BadToBest / EchoMimic

Lifelike Audio-Driven Portrait Animations through Editable Landmark Conditioning
https://badtobest.github.io/echomimic.html
Apache License 2.0
2.26k stars 263 forks source link

CFG Test (Gradio Configuration) #49

Closed oisilener1982 closed 1 month ago

oisilener1982 commented 1 month ago

1.1 CFG This is lowest setting that works in my test with no error

https://github.com/user-attachments/assets/d4e432e0-9e72-4bdb-98cb-47908f70304c

1.5

https://github.com/user-attachments/assets/caf72c92-c982-4a42-a12c-4cd49fa11fb3

2

https://github.com/user-attachments/assets/3fc78d18-da99-4582-9792-85e1da1e052f

2.5

https://github.com/user-attachments/assets/5bf281d5-757e-4e65-96de-989ebd366084

3

https://github.com/user-attachments/assets/d165ec5d-6916-46ec-8402-d934ab2d3a34

4

https://github.com/user-attachments/assets/ce7bd485-1297-4dae-9fda-52ad7818b224

5

https://github.com/user-attachments/assets/cf19cc51-c55e-4a84-8ea6-74b90870f3b9

7

https://github.com/user-attachments/assets/7521402c-9788-4071-b34b-ff3506c13425

10

https://github.com/user-attachments/assets/6500608c-c157-4c54-be11-4f39124aeeed

oisilener1982 commented 1 month ago

I personally like CFG 4. You can also play with this setting, it does not affect the generation time The lowest settings of 1.1 is not acceptable and the highest CFG 10 is also not good

THe reference image Ivana4

yuange250 commented 1 month ago

good job 👍

JoeFannie commented 1 month ago

We have found that the current code only supports cfg > 1. We are working on fix this bug. Thank you for ablation study.