restoration_test.py: error: unrecognized arguments: --arcface_path ./pre-train/arcface_model.pt

[cavalry7131]

@server:/video_dir/dataSet/_workspace/code/VSPBFR-main$ python restoration_test.py
--ckpt ./pre-train/restoration_net.pt
--lq_data_list /video_dir/dataSet/_workspace/face512
--hq_data_list None
--data_name_list dark_face
--psp_checkpoint_path ./pre-train/style_encoder_decoder.pt
--arcface_path ./pre-train/arcface_model.pt
--ddpm_ckpt ./pre-train/code_diffuser.pt
--size 512
--batch 1

from .fused_act import FusedLeakyReLU, fused_leaky_relu
from .upfirdn2d import upfirdn2d

usage: restoration_test.py [-h] [--batch BATCH] [--size SIZE] [--mixing MIXING] [--channel_multiplier CHANNEL_MULTIPLIER] [--ckpt CKPT] [--ddpm_ckpt DDPM_CKPT]
[--psp_checkpoint_path PSP_CHECKPOINT_PATH] [--eval_dir EVAL_DIR] [--lq_data_list LQ_DATA_LIST] [--hq_data_list HQ_DATA_LIST] [--data_name_list DATA_NAME_LIST]
[--ckpt_root CKPT_ROOT]

restoration_test.py: error: unrecognized arguments: --arcface_path ./pre-train/arcface_model.pt

[LonglongaaaGo]

Hi @cavalry7131 ,

Thank you so much for your attention!
Could you try to remove this, and try to run again:
--arcface_path ./pre-train/arcface_model.pt

[cavalry7131]

Yes, it works now. However, the face restoration results inferred from our own face dataset have low fidelity.

[LonglongaaaGo]

Hey @cavalry7131 , which data are you using? could you also share the input data and corresponding results?

[LonglongaaaGo]

We may have uploaded a wrong checkpoint. We are currently investigating the issue and will provide you with an update soon. Thank you for your patience!

[LonglongaaaGo]

Hi @cavalry7131, after investigation by @YasuoFly, we found that it is because the input hyperparameter in our testing script is too confusing. So we finally simplified it.
Could you try it again for the testing using the well-trained model or a pre-trained model provided in our HuggingFace?

Let me know if that works for you
Thanks!

[cavalry7131]

OK, let me try.

[cavalry7131]

I try agin. It still have low fidelity.It generates a lot of twisted and horrific details, and turns Chinese people into foreigners.Due to confidentiality reasons, the results of image inference cannot be made public on GitHub. Can you leave an email address and I will send a face and its inference results to your email address?

[LonglongaaaGo]

@cavalry7131 It sounds like you are retraining the model? Could you tell me the details about how you produce the results (how you train, how you test etc.)?
My email is lwlxhl[at]gmail.com

[cavalry7131]

I did not retrain, I just used your updated test code to re-infer, and the inference result is still the same. I can provide a test picture of mine and send it to your email.

[LonglongaaaGo]

Hi @cavalry7131,

Thanks for sharing your testing data. The current performance issues likely stem from the following two factors:
1. Misalignment of input images: The images are not properly aligned before inference. Proper alignment is essential for accurate results. Please refer to this script to align your images before running the model.
2. Domain gap in noise distribution: The noise characteristics in your data may differ from those seen during pre-training. For best performance on custom data, it’s recommended to retrain the model on your specific domain.

Let me know if you need help with either step!