Understanding re-aging results

Hello, thanks for this awesome work! I have a question trying to understand the following re-aging results, given an input face here:

I got the following different results with the same inference code:
python age_editing.py --image_path xxx --age_init 39 --gender male --save_aged_dir xxx --specialized_path xxx --target_ages 0 10 20 30 40 50 60 70 80 90 100

Do you have any idea why the results are not deterministic in the above case?

Also in case 2:

Do you have any idea why there are artifacts?

Hi, I'm not an author of this paper, but have some insight.

First, I conjecture artifacts are raised by non-enough optimization in null-text optimization process (before Prompt2Prompt editing process) or huge gap between source prompt and inputted image. One potential solution is to increase the steps of null-text optimization to obtain more faithful null-text embeddings. To do that, you can modify num_inner_step in

FADING/null_inversion.py

Line 152 in 14fefbf

def null_optimization(self, latents, num_inner_steps, epsilon):

like from num_inner_step=10 to num_inner_step=20. It takes more time, but might generate more faithful results.

In addition, I observed that input age also affects the resulting images like below figure:

To mitigate this issue, you can use off-the-shelf age estimator (e.g., DEX, InsightFace) than age constant via command line.

I hope it works well 😀