4

I just took a Udemy crash course in ChatGPT and Midjourney. ChatGPT has this great ability to have an iterative conversation, like "Write a product description for a condenser microphone". Then after it generates some text, I can say, "Remove the nonsense about automatic background noise removal and make it more compelling for people who want to make podcasts," and it'll come back with a slightly revised version.

I'm battling to work out how to do that with Midjourney, though. It's not easy to get everything right in one prompt. So I'll get back four candidate images, and I can either ask for more variations on one of them, or I can upscale one I like. But I can't find a way to say, "make the man's beard longer and his skin tone darker".

What is the best way to take a "nearly there" image from Midjourney and tweak it so that it's just right?

4
  • I question whether this is on-topic here (I haven't cast any votes as yet though). Seems to be better handled by MidJourney's discord.
    – Scott
    May 16, 2023 at 18:23
  • 3
    I feel like this question is just as on topic as the many "how to" questions that are asked about other image creation software. May 16, 2023 at 20:51
  • I'm not adamant about it @ZachSaucier which is why I cast no votes either way. -- It seems like a question regarding automation which is more akin to external scripting (javascript, php, etc). "how do I target this imagemajick script better" would be in the same grey area - at least to me.
    – Scott
    May 16, 2023 at 21:04
  • When I played with CLIP/VQGAN I would just use the previous image as a starting point and alter the prompt
    – user253751
    May 17, 2023 at 13:53

1 Answer 1

6

The short answer is that AI image generation just isn't quite there yet. (This probably isn't true in the near future.)

Anyone who is dabbling with image generation has this problem of how to gain control. When people post a nice and seemingly perfect generated image, we can never know if the image really is exactly what the creator had in mind or if they just adjusted their expectations after seeing the result.

Midjourney makes nice images out of the box, but perhaps lacks fine grained control. I'm not sure. I haven't used it much to be honest. I've mostly played around with Stable Diffusion. It evolves chaotically and fast. So it's less accessible, but it has more possibilities to customize.

Here's a list of things you could consider to refine your images. Some of them are only available in Stable Diffusion but some of them are general advice:

  • Keep refining your prompt (and negative prompt) to get closer to the result you want. Be creative. Sometimes you need to find a balance between positive and negative prompt words and sometimes it's not exactly logical which words to use. You have to experiment. (You might get an image of a person looking too young. Adding old to the prompt might change the whole mood of the image whereas adding zombie with a low weight might just make the person a little darker around the eyes, achieving your goal. I have no idea if it would work, but perhaps in your case trying words like wizard or hermit might provoke a long beard?)

  • Make larger quantities. You are talking about generating four images and then four more. Perhaps you need to generate hundreds of images and then search for the needle in the haystack.

  • Use inpaint to change parts of your image. This is an iterative process which might require many generations before you are satisfied with the result.

  • Use Photoshop or similar to manually edit your image. (This is probably the real trick to many of the "perfect" generations we see online.)

  • Use something like InstructPix2Pix (https://github.com/timothybrooks/instruct-pix2pix) or similar. It's months old so something better has probably emerged. It aims to make it possible to adjust an image using a prompt. So in theory you should be able to prompt make the beard longer. In my experience it doesn't work so well, but it's probably only a matter of time.

  • Use ControlNet (https://github.com/lllyasviel/ControlNet). It's not exactly what you are asking for, but it makes it possible to use different aspects of the composition of an input image to generate a new image. Lots of possibilities for creative iterative use.

(The Stable Diffusion subreddit is a good place to get updated on the latest development and get tips and tricks: https://www.reddit.com/r/StableDiffusion/)

(Most people using Stable Diffusion use Automatic1111's Stable Diffusion webui: https://github.com/AUTOMATIC1111/stable-diffusion-webui)

1
  • 1
    "Anyone who is dabbling with image generation has this problem of how to gain control. When people post a nice and seemingly perfect generated image, we can never know if the image really is exactly what the creator had in mind or if they just adjusted their expectations after seeing the result." -- I think this paragraph nails it.
    – Shaul Behr
    May 17, 2023 at 13:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.