Lighten, Screen, Addition, Darken, Multiply, Linear burn, Hard Mix, Difference, Exclusion, Subtract, Grain Extract, Grain Merge, or Luminance.
A friend of mine made a similar animated GIF type captcha a few years ago but based on multiple scrolling horizontal bars that would each reveal their portion of the underlying image including letters, and made a (friendly) bet that it should be pretty hard to solve.
Grabbing the entire set of frames and greyscaling them, doing an average over all of them and then applying a few minor fixups like thresholding and contrast adjustment worked easily enough as the letters were reveleaed in more frames than not (I don't think that would affect the difficulty much though if it were any diffierent). After that the rest of the image was pretty amenable to character recognition.
You actually don't need any image editing skill. Here is a browser-only solution:
1. Take two screenshots.
2. Open these screenshots in two separate tabs on your browser.
3. Switch between tabs very, very quickly (use CTRL-Tab)
Source: tested on Firefox
They even provide the source code for the effect:
On iPhone: screenrecord. Take screenshots every couple seconds. Overlay images with 50% transparency (I use Procreate Pocket for this part)
https://www.youtube.com/watch?v=Bg3RAI8uyVw
The effect is disrupted by introducing rendering artifacts, by watching the video in 144p or in this case by zooming out.
I'd love to know the name of this effect, so I can read more about the fMRI studies that make use of it.
What I've found so far:
Random Dot Kinematogram
Perceptual Organization from Motion (video of Flounder camouflage)
"ffs".
- The captcha would be generated like this on a headless browser, and recorded as a video, which is then served to the user.
- We can make the background also move in random directions, to prevent just detecting which pixels are changing and drawing an outline.
- I tried also having the text itself move (bounce like the DVD logo). Somehow makes it even more readable.
I definitely know nothing about how LLMs interpret video, or optics, so please let me know if this is dumb.
This is untrue in at least one sense. The patterning within the animated letters cycles. It is generated either by evaluating a periodic function or by reading from a file using a periodic offset.
Roughly you create another full size rect. On each frame add a random pixel on row 1 and shift everything down.
Make that rest a layer below the top one which has Hello cut out as transparent.
In any single frame the result is random noise.
You don't even need to maintain the approach of having the pattern within the text move downwards over time. You could redraw it every frame with random data, as if it was television static. It would still be easy to read, as long as the background stayed fixed.
This is on MacOS 15.6, Chromium (BrowserOS), captured with the OS' native screenshot utility. Since I was asked about the zoom factor, I now tried simply capturing it at 100% and it was still perfectly readable...
I guess the trick doesn't work on this browser.
The culprit had more than 10k photos of all security details for thousands of wealthy customers.
