A real-life demo [WARNING: SPOILERS!!!]

Folding Tile Illusions

Auxetic mechanical linkages of connected squares. Click and drag them from left and right!

Folding Tile Illusions: Videos

Here are some videos of the above illusion, in case your browser is too slow to render them nicely.

Project Details

Click the this link to learn much more about Diffusion Illusions!

Overview

The project is comprised of four demonstrations:

1. Flip Illusions: Generating images that appear as different objects when flipped upside down - For example, a giraffe could look like a penguin. The image satisfies two different textual prompts dependint on its orientation.
   
2. Folding Tile Illusions: A mechanical linkage of small squares - We splice an image up into a checkerboard-grid of small tiles, and rotate each even tile 90 degrees and each odd tile -90 degrees. This can be built as a bunch of small squares with hinges on them!
3. Rotation Overlay Illusions: Revealing multiple images through image rotation - Overlaying two images to reveal four different images when rotated. The top image, when overlayed on the bottom image and light shines through, will reveal an image such as Darth Vader. When the top image is rotated 90 degrees, it becomes a different picture such as Pikachu. Another 90-degree rotation reveals another image and a final 90-degree rotation reveals a fourth image.
4. Hidden Overlay Illusions: Revealing a hidden image through image overlay - Overlaying four images, such as Hatsune Miku, a bear in a field, Mario, and a pumpkin, to reveal a fifth image, Darth Vader, when light is shone behind it. The four images are optimized to look like the original objects when viewed individually, but when combined, they reveal the hidden image of Darth Vader.

Similar Projects

Another project with a similar goal was released in February 2023: Illusion-Diffusion, authored by Matthew Tancik. That project only achieves the first part of this demo, aka image flipping.

Our other paper, "Peekaboo: Text to Image Diffusion Models are Zero Shot Segmentors", has an extremely similar formulation to the illusions on this website - it optimizes images using a variant of score distillation loss to perform zero shot segmentation.

Inspired by our Diffuion Illusions project, Daniel Geng et al wrote Visual Anagrams. It has beautiful results and is definitely worth checking out!

How it Works: Pseudocode

To help explain how these were calculated, I've included some easy-to-read pseudocode that covers how each of these illusions were made. This code has been oversimplified for the sake of simplicity - for all the details please read our paper. The full source code can be found at github.com/RyannDaGreat/Diffusion-Illusions, or at any of the Google Colab links throughout this webpage.

Flip Illusions

Rotation Overlays

Hidden Overlays

Timelapses

Diffusion illusions are generated through an iterative process: we optimize the output images with gradient descent, starting from pure noise. Here are some timelapses showing how they evolve over time.

Flip Illusion Timelapses

For reference, here are the end results of the above timelapse video:

Giraffe / Penguin

Angelic Pretty / Victorian Dress

Hidden Overlay Timelapses

Bibtex Citation