After the update, the generated results became worse

[norsizu]

After the update, the similarity of the generated characters has dropped significantly. Even when I use the exact same parameters as before, the generated results are much worse than the previous workflow results. Original image Generated results before update Updated generation results

[cubiq]

I need to know the exact update and the workflow you are using.

A previous version was applying the embeds twice, so it was actually an error but the resemblance was better of course

[charliedzm]

I need to know the exact update and the workflow you are using.

A previous version was applying the embeds twice, so it was actually an error but the resemblance was better of course

Actually,the previous version is better resemblance. How can I get previous effect in this updated version?

[norsizu]

It really seems to be because the previous version was applying the embeds twice, I refer to the previous workflow, added ControlNet application once, and now the generated effect is very close to the previous version. InstantID_zoe.json

[cubiq]

use the "advanced" node and increase the ip_weight/cn_strength. the result will be the same

[fadygit]

use the "advanced" node and increase the ip_weight/cn_strength. the result will be the same

I got what " norsizu " is talking about and I agree with him that the previous one was better result than this new node . If you noticed at the above two pictures and especially at the 2nd one , you will notice that there is a " X " mark at the middle of the picture and I figured out that new updated node (Apply InstantID) has now the input "control_net" than the previous one , so when I connected the controlnet model "InstantID-diffusion-pytorch" to the node ("Apply InstantID" ) and the controlnet depth model to ( "Apply Controlnet Advanced") , it gave me the same X mark.

please check below this screen shot:

[cubiq]

from the picture your wf looks wrong, check the examples

[lyksdu]

from the picture your wf looks wrong, check the examples

I got the same black X in my picture. and I just use simple wf, just include instantID-custom-nodes.

can you see and check if there are wrong

[lyksdu]

can you see and check if there are wrong

I have checked and the X is the same as the result of face analysis

[lyksdu]

seems get the reason.

because the X is so similar with the face result, give more concern about code of face analysis.

In “ ApplyInstantID” class, there are two lines about face analysis, https://github.com/cubiq/ComfyUI_InstantID/blob/8b7932a66bf43b5e2499e2ba77df5c373af006da/InstantID.py#L283 and in "extractFeatures", when extract_kps=True, it will return facekps(of course), AND DRAW an X result on origin pictures. This makes changes https://github.com/cubiq/ComfyUI_InstantID/blob/8b7932a66bf43b5e2499e2ba77df5c373af006da/InstantID.py#L170

and the solution currently remove the draw_kps function call, and just use the pic. https://github.com/cubiq/ComfyUI_InstantID/blob/8b7932a66bf43b5e2499e2ba77df5c373af006da/InstantID.py#L170 this may cause the pic look blue. there must be more errors.

[cubiq]

this has been covered multiple times, check the closed issues. Upgrade the extension and ipadapter if you have it.

check to have downloaded the models correctly, re-download if needed. Everything works as it should