I hadn’t paid much attention to Image Editing until recently. I’d done the occasional cut/crop or Paintbrush level “color over something in fog”, but nothing serious. Then actDave over at Dave’s Gaming pointed out to me how OpenAI had an option for AI-rendering based image cleanup. I’ve just scratched the surface of what it can do. And there’s tantalizing options there for doing things like selecting a portion of an image and re-rendering just that with a new text prompt. But for now I’ll explore some of the capabilities I already use.
Crop those Elf Ears!
One issue with AI rendering of fantasy scenes – at least with DALL-E – is that it *loves* giving people pointy ears if you give it the slightest reason to do so. One elf or half-elf character propagates to everyone else. Dwarves, Halflings, Gnomes,Half-Orcs – all often cause the same effect. Just having an Elven statue can “elf-ize” an entire scene. I used to carefully construct my multi-character renders to put any mentions of “elf, half-orc, etc” last to minimize the effect on the rest of the scene. But OpenAI Editor provides another option – crop the ears afterwards!
As an example, here is a recent extremely challenging render I did of all four characters from my Solasta series. Since Tess is half-elven, including her in a scene frequently gives some or all of her companions elf ears.
Now look at the original image and the “edited” image side by side:
For the elf ears, literally all I had to do was color-in the pointed parts of Falcon’s ear with the selection tool and then use the “Remove – Magic Erase” tool with maximum “creativity”. Boom. He’s human again.
Extraneous Weapons
Another artifact you can see in the images above is that the AI loves to add extra pointy things. Swords, bows, quivers, etc. Sometimes just adding them in when they’re not wanted at all. Other times filling the air in combat with extra weapons. Below is a particularly egregious example of this error, together with the AI-edited version. Again, the OpenAI Editor makes this very easy to remove – in exactly the same way I fixed the Elf ears.
Photo-bombing
The AI absolutely *loves* to add unwanted extras into your scenes. You can see one example in the campfire render at the top – with a fifth bearded half-elf entirely invented by the AI and added to the scene. More common is just adding bystanders in the background. This too is trivial to remove with the AI Editor. Note that its not always just humans. Many times when generating AI monsters, tattoos, heraldry, etc the AI gets over-enthusiastic. A demon with lobsters claws adds lobsters to the scene. Falcon with a falcon tattoos gets friendly avian companions. Etc. Here’s a few examples: one with both a random interloper and an unwanted crustacean, and the other with some unwanted intruders on a tender moment between Kat and Roland.
“Magic Erase” *does* have limits. Removing characters which significantly overlap other characters rarely works well – hence in the top image I cropped out the extra character rather than erasing him. He overlapped Falcon’s left side too much for the AI to repair effectively.
Facial Anomalies
The AI is generally pretty good with faces, but it does have some weaknesses. There is a “medium” range at which it tries to add more detail than the resolution supports and you get weird looking eyes and expressions. But sometimes it just glurps in small places for no good reason. The OpenAI editor has tools explicitly for fixing faces and hands when those go wrong. See the example below of Kat and Roland having a heated discussion – and look closely at Kat’s eyes and nose. Note that OpenAI *also* has tools for entirely replacing faces and changing expressions, but I have not yet played with those and so have no experience with their effectiveness/ease-of-use.
Conclusions
This is just scratching the surface of what an AI-powered image editing tool can do. And it can edit *any* image, not just AI generated ones. When I return I’ll be exploring the power of this more deeply – particularly to see what I can do with the “re-render this part of the image with a different prompt” capability. But just the above makes it *so* much easier to get usable renders – because where I used to have to just generate a ton of images until I got exactly what I wanted, now I just have to get close enough that the Editor can do the rest.
Leave a Reply to atcDaveCancel reply