How was the artistic background of T2I model preserved in IP-Face-Plus?

tencent-ailab / IP-Adapter

The image prompt adapter is designed to enable a pretrained text-to-image diffusion model to generate images with image prompt.

Apache License 2.0

4.5k stars 296 forks source link

How was the artistic background of T2I model preserved in IP-Face-Plus? #288

Open zechenghe opened 5 months ago

zechenghe commented 5 months ago

Thanks for the great work! I am curious how the text-to-image model is preserved to generate artistic background? The Liaon face dataset contains all real image such that the model will be generating real background after training? Any guidance how to incorporate the synthetic data for training? Thanks!

xiaohu2015 commented 5 months ago

a main reason is that we freeze the UNet