Does neural style give comparable results to fast neural style tensorfire?

[lavanyakapahi]

Hi Anish, I am interested in experimenting with artistic style transfer and found your project. Thanks! I used the great wave attached as the style and ran it through 1000 iterations. The results are attached. I then checked with TensorFire. The results are very nice! Can I achieve the same via neural style? I tried tuning the parameters but could not get good results. Content image Style image Result after 1000 iterations Result with TensorFire

Thanks so much Lavanya

[anishathalye]

I don't know, I haven't done a very careful comparison. I think it's generally easier to get nicer results with fast style transfer (which is what TensorFire uses).

[lavanyakapahi]

Thanks for suggesting the fast style transfer. I checked it and the pre trained model gave a very nice result. I need to train a new water colour style. I thought it would be a good idea to understand the tuning parameters by using the neural style before trying to train a model for the fast neural style. Not sure if that would be a right approach. So meanwhile, I continue trying out the neural style. I noticed that each time I run the script for 1000 iterations I get inconsistent results even though the parameters are the same. I ran the script with --content-weight 2 and --style-weight 1, twice. I have attached the results below. Is that expected?

[lavanyakapahi]

I used the init image option and this seems to give consistent results.

[anishathalye]

Yeah, it must be different random initializations (which is what happens if you don’t use init-image) giving different results.

On Aug 26, 2018, at 7:25 PM, lavanyakapahi notifications@github.com wrote:

Closed #125.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.