Generating images from their descriptions

I think this is paper: http://arxiv.org/abs/1511.02793

Very interesting. I'm wondering whether statistics-based photo synthesis is actually viable. The main problem is that many kinds of scenes appear to require a precise geometrical understanding, e.g. sharp shadows that get cast on detailed surfaces by a desk lamp, or that a mirror would precisely reflect another object in the scene from a different angle.

On the other hand, we don't require absolute precision, it just needs to be plausible enough to fool our eyes. If a shadow just looks roughly like the object that it casts, that should be enough. It seems to me that this still would require at least a rudimentary geometrical understanding/abstraction which will possibly require in comparison to this technique extremely sophisticated neural networks that can perform things like mental rotation, path integration, object recognition and have an understanding of object permanence.

For the life of me, I cannot figure out how to actually view the images on this page. They are all extraordinarily small thumbnails with no way to click or enlarge them?

red

hot chilli

Five days ago, I submitted an opposite project with this https://news.ycombinator.com/item?id=10518232.

Generating little stories about image.

The project is available here[1]. Author who develops this project is from University of Toronto too[2].

I thought it was same author when I read this.

[1]: https://github.com/ryankiros/neural-storyteller

[2]: http://www.cs.toronto.edu/~rkiros/