Image-conditioned model rarely condition on image

[dqj5182]

I have run Paint3D in my own data, and I think image-conditioned model rarely get conditional input from image (not as pipeline perspective, but in the perspective of the final output).

May I ask how we can put more emphasis on image condition rather than text condition?

[samiede]

Decrease the denoising strength

[Zzlongjuanfeng]

Two options may enhance the influence of the image condition:

1. Increase the weight of the image condition.
2. Replace the IP-adapter model with 'ip-adapter-plus_sd15'.

[dqj5182]

@Zzlongjuanfeng Thanks for your feedback! Sorry but where can I increase the weight of image condition?

[Zzlongjuanfeng]

@Zzlongjuanfeng Thanks for your feedback! Sorry but where can I increase the weight of image condition?

During class initialization, you can set the weight of the IP adapter using 'pipeline.set_ip_adapter_scale(xx)'.

[dqj5182]

Thanks a lot! I think my issue has been solved, so I will close the issue.