Dall-E 2 Mini Makes Bizzare Human Faces Out of Input Texts

While there have been other versions before for achieving proximate results, why does Dall-E 2 mini is the one with a punch?

The difference between the phrases eye-pleasing and eye-grabbing is subjective – a matter of individual predisposition – particularly when the object under study is a computer-generated image. Open AI, a company known for its futuristic research in artificial intelligence has recently released Dall E2 mini. It differs from the earlier version Dall E2  for its availability to the general public and for this very reason, internet users are going ga ga over the kind of images it is generating. Yes, it is about popular personalities doing weird stuff on the internet. Whether be it Boris Johnson eating fish or Bruce Willis eating yogurt, these images are not just amusing but misleading. The images generated in the grid format just look like captchas meant for aliens. If you are buying into it, perhaps you need this primer more than anyone else.

How does Dall E 2 mini work?

Named after the popular Pixar movie, Wall-E, it is capable of gathering data on thousands of photos to train its algorithms on what a photo looks like. When provided with millions of images, it can rearrange the photos in any pattern or combination that anyone can think about. Once it receives an input, it identifies the key features of the picture such as the trumpet or the curve at the top of a teddy bear's ear. The details are then captured by the diffusion model, which is nothing but the second neural network, to generate the pixels for replicating the image in a resolution higher than we've ever seen.

Dalle-E 2, the Winslow Homer of the digital Image Generators:

While there have been other versions before for achieving proximate results, why does Dall-E 2 mini is the one with a punch? Try suggesting placing differently colored cubes over one another – apart from having dyscalculia — invariably it would fail. Dall – E 2 mini is the winner here for its ability to follow instructions to the point. Type "Charles Barkley at the Masters", and you might be able to see a bunch of compositionally accurate pictures of him hitting golf balls at Augusta. While the earlier versions are capable of generating images of one description, Dall-E 2's beauty lies in its ability to edit the picture, ie., reproducing or modifying a part, blending the details of two pictures, in painting a picture with additional details, or generating a variation of an existing picture including making modifications for a change in direction of shadows. 

So, what is the big deal about human faces?

Dall-E 2 mini clearly fails in this department. The concept is fairly depicted but if you need clarity with details, you will be disappointed. Anything which involves human faces gets wicked in the output with some small objects losing clarity in the final image. Check the images of Joe Biden dancing in the backroom or George Costanza from Seinfeld holding multiple cats, hardly one can recognize them in those unexpected sorties. The developers at OpenAI have a reasonable explanation for it. They mentioned, that they deliberately scrubbed some unwanted volume from the training data, though they didn't give the exact reason for doing so. Borin Dayma, the machine learning engineer, who developed Dall -E 2 mini algorithms says, there is much more to come, for the charm of machine learning lies in repeating.