Just like 2021 will be remembered for the rise of NFTs, it seems that 2022 will be remembered as the year of DALL-E 2 and the text-to-image generation models. And while this trend gained meteoric momentum, Israel obviously didn't just sit on the sidelines and watch: The Israeli company Lightricks – which created some of the most popular and respected photo and video editing apps on iOS and Android – has just come out with its own version of an AI-based image generator but makes it even more accessible and interesting – including the possibility to enter descriptions in Hebrew.
Lightricks has just launched 3 interesting new features: the first is an AI Creation Tool which generates images from text. It's pretty simple and very similar to DALLE-2 or Midjourney: you write what you want in a text box, and the engine behind it creates an image based on models trained on hundreds of millions of images. The nice twist about the Lightricks version is that you can be a little less concerned with the wording of the command itself (which has almost become an art in itself and includes prompt experts who will write picture descriptions for you) and work more freely. In doing so, the application actually allows you to choose the style of the image you want, which saves you the hassle of needing to be overly descriptive.
We tested the new feature ourselves, which is now in beta testing, but of course, like any other new model or tool, you will have quite a bit of trial and error. Sometimes the simple commands we gave produced a great image, and other times they gave completely unrelated images. I found that the very tool that is supposed to make the process easier and allow you to choose the style from several options slightly detracts from the final result and takes it out of context. Just look at this very amusing picture of Darth Vader suntanning on a Tel Aviv beach without a style command, and then the result with the addition of a "Photo" or "Cinema" command. Sometimes it turns out better, and sometimes it just has nothing to do with almost anything I asked for. What's more, the processing is done quickly compared to engines like DALLE-2 or Stable Diffusion, and it's not too much of a problem to click Generate Again to give the system another chance, because it just doesn't take as much time.
Some other examples:
Mister Bean as the Green Giant?
Another feature known in the world of generated images is "Fan Casting". Haven't you always dreamed of seeing Mr. Bean in the role of Hulk or Captain America? Now you can! The new Lightricks tool allows you to enter any celeb you want in whatever movie you want to "cast" them in, the character you want them to replace, and simply generate an image. You can also add a few more textual definitions to make your result more precise and cast Brad Pitt as Frodo from "Lord of the Rings". This is bound to be a hit– letting us plant ourselves on whatever we want, and you've got a recipe for a cool viral feature.
For dessert, the Lightricks’ Photoleap app will also have a Sketch to Image feature, which we have yet to try. The feature will allow you to scribble basic sketches of what you want to create, add text to describe what you have drawn, and the Lightricks engine will build a strictly AI-based image from them. This feature reminds me very much of Meta’s version (developed by Israeli researchers, no less), which also allows you to scribble a general drawing, add a prompt, and get a drawing that is more like what you had in mind. Yaron Inger, founding partner and CTO of the company, tells us that the products were built on the API of StableDiffusion which is based on Latent Diffusion models. Lightricks uses transformers (the components that translate text into tokens of numbers) and then Diffusion Models that take the numbers (these tokens) and know how to produce the image. Inger adds that Lightricks uses its huge number of users to know which prompts work better, and thus refine the results.
Free but with some limitations
The new features are available in the Photoleap (iOS only), Motionleap, and FaceTune (both iOS and Android) apps–which can be downloaded here. Currently, you can try the features for free with a few restrictions. Such limitations include the number of times the feature can be used, and the application will watermark the final image. In general, the company explains to us that these features are now being launched in a kind of open beta, with many more features coming soon on a weekly and bi-weekly basis. So, to the product managers out there, stay strong.