seanseattle / styleswap Goto Github PK

I tried to run test.py and I got the following error:

Namespace(source_img_path='./examples/images/2.png', target_img_path='./examples/images/12012.png', output_path='results/test.png', size=512, align_source=False, align_target=False)
W1107 11:21:27.417249 22688 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 8.6, Driver API Version: 11.6, Runtime API Version: 11.2
W1107 11:21:27.535109 22688 device_context.cc:465] device: 0, cuDNN Version: 8.1.
.../anaconda3/envs/paddlepaddle/lib/python3.9/site-packages/paddle/tensor/creation.py:130: DeprecationWarning: `np.object` is a deprecated alias for the builtin `object`. To silence this warning, use `object` by itself. Doing this will not modify any behavior and is safe. 
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  if data.dtype == np.object:
Traceback (most recent call last):
  File ".../StyleSwap/test.py", line 76, in <module>
    generator.set_dict(paddle.load('./checkpoints/styleswap512.pdparams'))
  File ".../anaconda3/envs/paddlepaddle/lib/python3.9/site-packages/paddle/framework/io.py", line 985, in load
    load_result = _legacy_load(path, **configs)
  File ".../anaconda3/envs/paddlepaddle/lib/python3.9/site-packages/paddle/framework/io.py", line 1003, in _legacy_load
    model_path, config = _build_load_path_and_config(path, config)
  File ".../anaconda3/envs/paddlepaddle/lib/python3.9/site-packages/paddle/framework/io.py", line 161, in _build_load_path_and_config
    raise ValueError(error_msg % path)
ValueError: The ``path`` (./checkpoints/styleswap512.pdparams) to load model not exists.

Can we download the missing checkpoints ? If so, how can we download them ?

Please remove this repo as you have no intention to deliver on your promise. Thanks.

-The owner said they aren't releasing the model, no need to check up on the repo.

Oh my god! I love this

Just looking at the preview, this looks good! It is better than simswap, and is way more friendlier than deepfacelab! I am still waiting for this but I'm honestly mind blown
I would love to see more research like this and also see if it's possible to manipulate the lips of the face! Technoligy has come so far!!

Very cool! Do you know how we can preserve the source's expression?

Hello, thank you for your efforts. I tested the same type of code and found that the biggest problem with stylegan is that the face reconstruction is not accurate. Rather than reconstruction, it is better to say that stylegan regenerates a picture similar to the source face. Even with some optimizations, the generated image is more similar to the source image. If I want to use stylegan2 to generate the exact same face as the source image, what do I need to do? Looking forward to your reply

After all this time, this would be a logical step to publish pretrained models after posting some simple test code. Or publishing the training code. Please give at least some feedback.

You should reference your code on https://paperswithcode.com
Currently only your paper is referenced but not the github repository https://paperswithcode.com/paper/styleswap-style-based-generator-empowers

Thanks for your wonderful work! Recently I'm working on another faceswap network, but when I tried to reproduce the expression error in the paper, I met with difficulties. I use the paper A Compact Embedding for Facial Expression Similarity refered in your paper to extract expression embeddings. The code I used is AmirSh15/FECNet which is the only pytorch implementation I could find. But when I compute the L2 distance of the 16-dim embeddings of target face and swap face, I found that the result is always around 0.5(I tested DeepFakes, FaceShifter.etc), which differ greatly with the result you put on your paper. So could you please give me some details about how you compute the expression error? I would appreciate it if you could show your expression error code to your convenience.

Hi thank you for the incredible work.
I'm confused with the mismatch in paper and code.
In the code, you follow the GPEN, which is your reference paper. Features from the encoder are concatenated "not before" the upsample, but after 2 convolutions in building block. Is there any reason for that?
It confused me. Because the paper's architecture is new to me, but not the implementation.

Hi, thanks for sharing this wonderful work. I wonder when will you release the code.