ycjing
commented
4 years ago Hi @TJCoding ,
Thank you so much for your interest in our work. It is a very interesting issue to discuss the transfer of Surrealism style. I agree that current neural style transfer algorithms are not good at handling the Surrealism style. My understanding is that neural style transfer algorithms are designed for universal style transfer, and not specifically optimized for some challenging styles like Surrealism. I believe that it is very promising to develop some algorithms that are specialized for Surrealism style transfer.
Also, it is really nice to see such a novel type of photo-surrealism transfer. The results are impressive. I have highlighted the term of "Photo-surrealism Style Transfer" in this repository, and look forward to seeing more related research.
Thank you again for such a nice introduction of this new type of Photo-surrealism Style Transfer!
Best, Yongcheng
I should firstly say that the paper ‘Neural Style Transfer: A Review’ is an excellent and comprehensive review and the following should be considered no more than a minor observation.
The paper states that surrealism is addressed in the evaluation exercise. The term ‘surrealism’ was coined by Guillaume Apollinaire in 1917 but only one of the works in Table 1 is later than 1917 and that is ‘White Zig Zags’ by Kandinsky. Many different artists are mentioned in the Wikipedia entry for surrealism but Kandinsky is not one of them nor are any of the other 9 artists listed in Table 1.
It is an interesting question as to whether surrealism can be addressed by neural style transfer, in any case, since surrealist art mostly depends on image content rather than image rendition. In a wider context, it would be a difficult challenge to construct a neural network that was capable of the creative thinking required for the production of genuine surrealist art.
On the Ostagram website [Ref 1], there are examples where people have submitted selfies as content images and Van Gogh paintings as filter images and the resultants are clearly recognisable as portraits in the style of Van Gogh. If Magritte paintings were to be used as filter images instead, then it is doubtful that it would ever be possible to create a portrait that could be mistaken for a Magritte. The essence of Magritte is that he uses a precise but inconspicuous painting style as a foil to the surreal subject matter.
The reason why I am raising the issue of surrealism is that I recently proposed the term ‘photo-surrealism’ for a particular class of imagery [Ref 2]. I proposed this term for images that look photorealistic in monochrome but which have unrealistic colouration. The following is an example.
It is interesting that the taxonomy diagram in this repository divides processing methods into photorealistic and non-photorealistic. The images above were generated by a (prototype) algorithmic method known as ‘Adaptive Recolouring’ which doesn’t fall fully into either category.
It can be noted that the original image above is not of great interest in itself. It is one of many green monotone photographs that I have taken where I have failed to capture the delicate tone variations that I was observing in the natural world. The application of colour to the green image provides some compensation by emphasising the elegant contours and texture of the subject vegetation. It seems likely that someone with better artistic sensibilities than me could achieve the same outcome with greater subtlety and finesse. Some further examples of photo-surrealism are shown below.
Ref[1] https://www.ostagram.me/lenta?locale=en [Ref 2] https://github.com/TJCoding/Enhanced-Image-Colour-Transfer/blob/master/Documents/Enhanced%20Image-Colour-Transfer.pdf