Settings

Learn More

Welcome to nostalgia box. This particular corner of the internet is a compilation of human memories. It contains an assortment of images, blended together into a soup of memories via the combined powers of machine learning and simple animation.

How it works:

This is a prototype of an extension on the paper A Neural Algorithm of Artistic Style by Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Nostalgia Box is a continually shifting soup of it's user's memories. Images, curated by the user for their nostalgic emotional impact, are dreamed over one another using a neural style algorithm. The resulting images are then overlaid into a video which never fully solidifies into any one image, but instead is always several at once. The user themselves are able to recognize what they see, but after losing track of what images they've entered, should lose their certainty in their identifications. A stranger should see only a haze of vague shapes which occasionally contains the suggestion of a face. This is machine learning, stripped of the elements of spam. You controlled what information went into it, you can understand what is being created with it, and no one is trying to sell you anything. It contains 58 images created using Deep Dream Generator. It uses a Javascript animation to fade through these images.

No image in this final version is of the original content. This was created by taking a starter image, and using it as our content image, along with a simple picture of water as our style image:

This image of water can be seen throughout the animation.

This does not, itself, result in a properly obscured image. Most images in have been run through the deep dream generator 3-4 times.

The general procedure for adding a new image is to use it as the content image, and the image preceding it as the style image. Variety in the settings used creates more visually interesting work, but, in general, keeping Enhance, Depth, and Style-Weight all the way up will result in the proper level of obfuscation in fewer iterations. Vary the style-scale. The image which is produced should then be used as the new content image, with the previous image as the style again. One or two more iterations of this process should result in images which look right to complete the fade. In general, a good fade from one image of pure content to another will require 3-5 images total, to create an appropriately smooth fade.

Graphic match is an essential part of making the fade from one image to the next turn out smoothly. Graphic match involves visual elements of one image lining up with visual elements of the next. For example, these two images have great graphic match:

(The yellow light stays in the same place between the two images, which offers a degree of consistency while the rest of the scene shifts)

While these two have horrible graphic match:

(Even though the location of the figures stays the same between the two, the shift is enough to be jarring, and there is minimal background noise to obscure it.)

Keep graphic match in mind when choosing the order of your images. You will want to be certain that all of your images are the same size, and can use the process of cropping and re-sizing images to line up major graphic details. If image aren’t looking the way you intend for the them to, adding an image from much earlier in the project to your process as a style image can help.

How it could work:

Ideally, this process wouldn’t need to be done by hand. The eventual goal of this project is to fully automate the process, such that someone unfamiliar with machine learning could still take advantage of it. The user should be able to upload a few images, and more at a later time if they so choose, and nostalgia box itself should take care of the rest.

It should maintain a database of images which the user initially uploaded as content, and images which have been generated previously. It should continually select two images at random, run neural style with one image as the style and the other as the content, return the result to the pile of images, and fade that image in, slowly, over the currently displayed image.

Other tabs in this menu explain the details of this process further.



Nostalgia Box was created by Aubrey Simonson as a Spring 2018 UROP project at the MIT Media Lab with Manuel Cebrian and the Scalable Cooperation group. For more information, contact asimonso@mit.edu.

× Home Add Content View Content Settings Learn More