Photorealistic checkpoint and mode produces caroon-like images instead of photorealistic

[pcarre-lab]

Thanks for your work.
I have problems trying to generate photorealistic style transfer.
Using the following images:
Content

Style

And running:

python test.py --content images/content.jpg --style images/style.jpg --decoder checkpoints/decoder_iter_160000.pth.tar --SCT checkpoints/sct_iter_160000.pth.tar --testing_mode pho

Using your provided photorealistic checkpoints, I get cartoonish result instead of photorealistic:

[pcarre-lab]

I trained a custom network with a dataset of factory images and using photorealism, and the inference using photorealism using last checkpoint is still caroon-like:

Would you know what could be wrong?

[JarrentWu1031]

I trained a custom network with a dataset of factory images and using photorealism, and the inference using photorealism using last checkpoint is still caroon-like:

Would you know what could be wrong?

Hi! The main reason for the cartoon-like effect is the global style perception loss. CCPL could alleviate the local color distortion to some extent. However, since CCPL only samples a few patches for each layer during training in consideration of computation costs, the restriction is not that strict. If you want to generate high-resolution photorealistic results, I think you should replace the global style loss.