Popular app Prisma applies painting styles to photographs using neural networks, turning a snapshot into an artwork in the style of 'The Scream,' for example. But what if you could transfer photorealistic effects from one photo to another? Researchers at Cornell and Adobe have successfully demonstrated a method that will translate a variety of styles from a reference photo to another image, including things like lighting, time of day and weather.
Input image on the left, reference style image in the center, output image on the right. It's not incredibly realistic-looking, but more realistic than your average Prisma treatment.
Images via Fujun Luan
This could open up a whole new world of possibilities for 'lazy' photo editing. Say you snapped a photo of a rock formation in the middle of the day, but you'd rather it had the orange glow of golden hour. With this method, you could apply the textures and colors of a reference style image, i. e. some other rock formation at sunset, to your own image.
This photo-style-transfer method augments the neural-style approach Prisma takes by constraining the colorspace of the transformation applied to the source image. Taking a content-aware approach and classifying features like sky and water in each image helps to avoid mismatched textures and distortions.
Advanced photographers would likely be wary of making such drastic edits to their photos. However, the technology might appeal to someone who wants to apply the effects of professional lighting to a badly lit photo of an interior, for example.
What do you think? Could this technology be useful to you? Let us know in the comments.
. dpreview.com2017-3-28 10:01
