search

ali-vilab / VGen

Official repo for VGen: a holistic video generation ecosystem for video generation building on diffusion models
https://i2vgen-xl.github.io
2.75k stars 243 forks source link

Question about the result in DreamVideo #123

Open Amnesiae opened 3 weeks ago

Amnesiae commented 3 weeks ago

Why are my results after training inference so poor? The example I use is “a * is playing guitar” https://github.com/ali-vilab/VGen/assets/67503039/83b3358e-3385-40d6-b6c4-7d1739bf5ae8

weilllllls commented 3 weeks ago

The results appear to show that the subject ID is overly preserved. You might try reducing the identity_adapter_index and increasing the motion_adapter_index during inference.

Amnesiae commented 3 weeks ago

Thank you for your answer！I did not modify any code, so do I need to adjust the parameters in the code myself to achieve the effect in the paper?

weilllllls commented 3 weeks ago

Yes, you will need to adjust the hyperparameters to suit your specific cases and we also provide some examples.