shiimizu
commented
1 week ago Yes, that's because the preceding token gets repeated n * 2 amount of times for v2, where n is the amount of images. It's just n for v1.
So if you have 3 images and the word before img is man, there will be a total of 6 man tokens in the prompt (man * 3 * 2).
With the recent change, I added the * 2 cause I wanted it to be the same as the original code, but seems like it was okay before, since I got the same results if I just repeated my images twice using the RepeatImageBatch node.
Perhaps not a bug but I just want to confirm;
Anything I enter after the trigger word is ignored. If I start with “photo of girl img in a red dress” then I get a closeup photo of just the face.
If I enter “photo of a girl in a red dress img” then I get a girl in a red dress, but the resemblance isn’t as strong as it would be if img was at the beginning.
I should add that I am use openai comfy, it shows no updates, and it did the same thing when I loaded the v2 example. Before the most recent update though I do know that the v2 example default prompt would render per the prompt.