DALL-E 2 technique: Close-up on primary focal point, followed by uncrop

The idea is: many images have a primary focal point. Start by generating a close-up image of that primary focal point. Then uncrop to generate the rest of the image. By improving quality of the main focal points, the overall image feels higher quality.

For example, I wanted a picture of an unhappy man standing in a hot desert. But DALL-E often has trouble with faces in full length portraits and I wanted the face (the primary focal point) to look good. So I started with this:

“Face and chest of a man standing underneath a hot summer sun in the desert. He looks very unhappy.”

I then uncropped, removed watermark, moved the face-shot to the right position in the image, and generated a second image:

“A man standing underneath a hot summer sun in the desert. Bare chest, wearing only shorts. Very unhappy. Full length portrait.” (DALLE link of resulting image)

(DALL-E didn’t flag this realistic face upload, not sure why)

Object focal point example

“Macro photo of a beautiful diamond ring in a small box.” (DALLE link)

I uncrop, remove the table, flip horizontal, and then do this:

“Hands of a man holding a diamond ring and proposing to his girlfriend.”

Multiple focal points example

You can even do this with images with multiple focal points. For example:

Focal point 1: “Closeup profile view of a man facing right and smiling lovingly. Candle light.”

Focal point 2: “Closeup profile view of a woman facing left and smiling lovingly. Candle light.”

Uncrop, erase enough of the faces to pass the face filter, and combine:

“Two lovers, a man and a woman, having a romantic candlelit dinner.”

Super-impose the two images from earlier to restore the face quality:

There are still minor artifacts from the uncrop but they can be cleaned up. The face quality ended up much better than anything DALL-E could’ve generated out-of-the-box. The hands, hair, etc. are still distorted, but it’s less noticeable since the main focal point (the faces) are so clear.

Reverse focal points example

Sometimes you have an image you already generated but you want to clean up some of the focal points. In this case you can do the opposite: crop out a smaller section of the image, remove the part you want to re-generate, adapt the prompt to produce a “close up”, and then re-combine it back into the full image afterwards. Here is an example I did:

“A family standing on a branch of an enormous tree, high up in the air.”

This was what I wanted but the faces were… blurry messes. I cropped out the middle section, erased their faces, then I did:

“Closeup photo of two girls standing in a tree.”

Still a bit messy, but when I downscale and recombine into the original image, it looks quite a lot better:

I could’ve cleaned up the other two faces in the same way, but didn’t bother in this case.