Google sells the idea of text-guided image editing (TGIE), and its own Imagen Editor by extension, as “a practical task that involves editing generated and photographed visuals rather than completely redoing them”. This is done by highlighting a part of an image you want edited, and then typing the nature of the edit itself. You can make small edits by highlighting a small part of an image, or completely change the image by leaving only a small part unhighlighted. No real surprise, since the basis of this tech is an image generator to begin with. On one hand, this sounds like it can be just as much fun, if not more so, than your standard text-based generated AI. But unfortunately, Google says that it is “not releasing Imagen Editor to the public”. The company cited “concerns in relation to responsible AI”. On the flip side, you can instead download EditBench, which is simply the software that the company uses to collect feedback on the quality of images edited by Imagen. Not that the general public would know what to do with it, though it’s meant to be “for the benefit of the research community” anyway. For what it’s worth, Google has published its research paper on the Imagen Editor, with plenty of examples of whacky edits done using the tool. The company claims that its own tool, “through extensive human evaluation on EditBench”, is better than DALL-E 2 and StableDiffusion. But you can give the paper a read and be the judge of that yourself. (Source: Google [1], [2])
