Closed kasim0226 closed 2 years ago
It looks like your run is fine. I doubt if there is some gap between the data, you can test the CelebA-HQ data to check if there is any problem, if not, you may need to retrain a HairCLIP on the FFHQ dataset for better results.
I just checked the test results and found that for people with very short hair there may be a problem with visual hairstyle changes due to non-uniform coloring. For other lengths of hair there is no such problem.
However, another version of the training does not have this problem, you can retrain HairCLIP and maybe the problem will disappear.
Got it. Thank you for your reply.
If I only want color change, would the follwing setting help when I train new model? --hairstyle_manipulation_prob=0 \ --color_manipulation_prob=1 \ --both_manipulation_prob=0 \ --color_text_manipulation_prob=1 \
Yes, and it probably converges very fast. Probably tens of thousands of iterations later it's already trained.
Thank you for your suguesstions.
Have fun with it!
I want to change hair color on FFHQ data, however hairstyle of some of results are change. Did I do wrong? The following is my command
python scripts/inference.py \ --exp_dir=./experiment \ --checkpoint_path=../pretrained_models/hairclip.pt \ --latents_test_path=./latents.pt \ --editing_type=color \ --input_type=text \ --color_description=red