Fixing Baby Face Scans! Lets make my wife cry!

[brazen tangle]

Hey all, my wife and I are having our first child in our mid 30s. She's going to be a wonderful mom but she's very stressed out about the baby coming in about a month and a half. She keeps thinking the worst case scenarios, but we have a perfectly healthy pregnancy thus far, so I want to give her something to look forward to take away some of the stress.

Last weekend we got some of those creepy 3d scans for the baby. She loved watching the scans get made but it's not the same once we got the still pictures. I thought it would be a great pick me up/vday gift for her to get a card that I will print with a cleaned up picture. I figured GPT+ DALL-E would be good for this, but I am running into some issues where it just makes a new photo.

I tried a bunch of different scans to merge them together but I haven't really found the best custom GPT for that. I know my prompt could use some work, so as I try to read up on prompting, I wanted to post this here for anyone else to see in the future, since this isn't that common of a use case that I've seen.

Please feel free to give me any prompt tips or help me understand why it doesn't change the picture, but renders a whole new one.

[ivory mango]

Hi! First of all, congratulations to you and your wife! How exciting!

As for DALL·E recreating the scan, there are a couple things to keep in mind about DALL·E that I think will make this kind of challenging. DALL·E makes every image "from scratch" every time. If you click the image in your screenshot, then click the little "i" icon, you'll see the prompt DALL·E used to create the image, and with a txt2img model like DALL·E, it's only this text that gets used to create the image, even if in the prompt you see words that make it sound like its looking at a reference image (recreate, clean up, preserve, etc.).

It doesn't "see" reference images, or even its own images. When you provided the reference image, ChatGPT used its separate image recognition feature to create a text description of the image. This description was then passed to DALL·E, so it's a kind of img2txt2img process, which results in the detail loss you're seeing here.

This is all to say that you might just have to try a bunch of times to get a result that looks as good as you want. Maybe try starting a new GPT-4 chat, uploading your reference image, then ask for a detailed description of the image. Then, take this description and edit it so that it preserves as many descriptions of visual details as you can include about your child, then return to DALL·E and ask it to create the image based on your description.

Again, it might end up having to be a "close enough" kind of thing since DALL·E doesn't currently have the ability to edit/manipulate/see source images, but perhaps with enough written visual description, you can get close to an outcome your wife and you will enjoy!

Again, congratulations to you both -- all the best! Please let me know if you have any questions.