search

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

ip-adapter faceID portrait cannot generate reliable 1024 * 1024 images #279

Closed hnhparitosh closed 5 months ago

hnhparitosh commented 5 months ago

I recently tried the ip-adapter faceID portrait model just like the example shared on the hugging face. The 512 512 or 768 768 images are great but when the resolution is 1024 * 1024, it generates very bad anatomy like two identical heads growing out of a single body, etc. Please look at the image given below. portraut_out

How can we avoid generating such outputs in 1024-resolution and can someone please share more information about this particular model? Thank you.

xiaohu2015 commented 5 months ago

you should use high resolution fix trick to generate high resolution images.

hnhparitosh commented 5 months ago

Is there any way to use the high-resolution fix programmatically as I am not using UIs like automatic111 or comfyUI?

xiaohu2015 commented 5 months ago

a simple way is firstly generate 512x512 images, then use a SR model to upscale to 1024x1024, and using image to image pipeline