search

ICTMCG / Make-Your-Anchor

[CVPR 2024] Make-Your-Anchor: A Diffusion-based 2D Avatar Generation Framework.
327 stars 20 forks source link

Different input return totally same result #6

Closed YahooKID closed 6 months ago

YahooKID commented 7 months ago

Thanks for open source,

Here is a little problem, I can't make sure whether it's a bug or features of model. I pull the code and setup the environment. And the demo has been repeated successfully. However, when I try to change the appearance image, the body infer model has return me the totally same result, when the clip embedding is different obviously. I wanna to make sure whether it's a bug or a feature, or somewhere I did was wrong.

bone-11 commented 7 months ago

Thank you for your attention.

As a person-specific approach, each of our model blinds with one specific appearance during fine-tuning stage. If you want to generate other appearances, further fine-tuning is necessary. This training setup is partly due to limitations in both data and computational resources of ours. To help training other appearance, we plan to release the pre-trained weight from pre-training stage, and the fine-tuning code. The guidance and code for preprocess training data will be updated further.

kike-0304 commented 7 months ago

Thank you for your attention.

As a person-specific approach, each of our model blinds with one specific appearance during fine-tuning stage. If you want to generate other appearances, further fine-tuning is necessary. This training setup is partly due to limitations in both data and computational resources of ours. To help training other appearance, we plan to release the pre-trained weight from pre-training stage, and the fine-tuning code. The guidance and code for preprocess training data will be updated further.

great