Disappointing result for the demo with 3 images.

bamos / dcgan-completion.tensorflow

Image Completion with Deep Learning in TensorFlow

Other

1.31k stars 388 forks source link

Disappointing result for the demo with 3 images. #8

Closed benstaf closed 7 years ago

benstaf commented 7 years ago

I ran the demo from your blog post with 3 images, but the result is disppointing:

completion_problem

My input images are: face3 face2 obama

I did not re-trained the model, I directly ran the code:

./openface/util/align-dlib.py data/dcgan-completion.tensorflow/data/your-dataset/raw align innerEyesAndBottomLip data/dcgan-completion.tensorflow/data/your-dataset/aligned --size 64

./complete.py ./data/your-test-data/aligned/* --outDir outputImages

cd outputImages convert -delay 10 -loop 0 completed/*.png completion.gif

bamos commented 7 years ago

Hi, I'm not sure what's going on with the first image but near the end the second two look reasonable. Using an adaptive optimization technique could make the optimization more stable and I just opened issue #9 if anybody's interested in implementing this, it should just be a few lines of code. It's described in Algorithm 1 of this paper and there are plenty of references of how to implement it in Python, and I like using the Torch implementation for reference.

-Brandon.

benstaf commented 7 years ago

The first image is computer-generated

The other two are more 'natural'. So this image can fool humans into seeing a face, but not the GAN. Can you explain why?