KakerMix posted:

Jesus I hadn't messed with Dall-E's image manipulation thing before, this is the real power. I snapped a picture looking out to my back yard.




You can then use a paintbrush to erase a portion of the image, then Dall-E will attempt to put whatever it is you ask it to via text prompt into just that area. I chose a complex area with intersecting lines of the concrete and puddle since it rained earlier. Looked like this before generation:




Results:


"a perfectly smooth chrome sphere"


"a translucent cube made of acrylic"


"a pile of sandwiches"

The sphere one it attempted to keep the puddle, sort of, but just ignored the line in the concrete. For the cube one the internal reflections don't seem to match too much, but does a pretty good job reflecting a lot of what is supposed to be around it, while deciding that the straight angled line is now a curve. The pile of sandwiches is a bit too sunny imho, but it did add a bit of shadow that's too sharp vs. the surrounding area.


I then applied the paintbrush to the entire right side of the picture, then asked it to generate "photograph of a backyard after a light rain" and these are the results:





This is essentially the content-aware function in Photoshop but on the extreme end of wholly-building entire portions of the photo with text context. It isn't perfect of course, but for a single 10~ second blot of time and it gives you multiple results is pretty impressive. Keeps the wet concrete most of the time, keeps the angle of things correct for where they photo was taken, lighting is correct.

I then erased the bottom portion of the concrete and asked it for "deep dark hole to the center of the earth"



Asked it for "a high performance sports bike"