Prompt for specific style ?

[104madlife]

Thanks for ur great work!

I'm trying to do the ID customization task(faceID) for anime character, but the color style of the output image seems to be odd and low quality. Could u please share what kind of prompt you used for specifying the style of the image? Especially those for the shown examples of ur posted results

[lehduong]

I have attached below some examples with different output resolutions. CFG is set to 4 and num_inference_steps is set to 65. The trick is copying part of the source images' captions to the target prompt (as in here).

You can increase the resolution by multiplying height and width by a factor such as 1.25, 1.5, or 2 (make sure height and width are both divisible by 16). In most cases, increasing the resolution will improve the image quality but might reduce the ID preservation capability (as the model was not trained on those resolutions).

For some odd styles, using more input images might help.

1 input image:

3 input image: