Advanced Prompting Technique

 

Wait, you can really do all of this with prompts?

You are not a beginner anymore and you consider yourself knowledgeable in the world of AI image generation?
Well, let's bring your skills even further with this article. Today we will talk about advanced prompting techniques and turn you into a real AI master.


If you still feel like a beginner, please give this guide a read first.


Summary :

  • Checklist of a good prompt.
  • Step-by-step prompt building.
  • Tweaking techniques.
  • Common mistakes.

Checklist of a good prompt.

I'm never going to remember all of this...



To start, here's the most important question to ask: In your opinion, what separates good from bad prompts?

When I start building a prompt, I use a checklist
If and only if my prompt contains all (or most) categories specified in my checklist, then I can move forward.
Here's the checklist I use : 

  1. Character: The character you want to see in the image. What do they look like? How are they dressed?

  2. Art style: Digital art, anime, oil painting, realistic photography, etc... A common mistake is thinking this is only defined by the model you use but specifying it in the prompt is very powerful.

  3. Background: Seeing only a close-up of your character is often boring. Where are they? On top of a mountain? On the moon? 1000 years in the past?

  4. Colors & Light: Very much like photography, colors, and lighting are key factors in creating successful images. Do you think your picture would look better if your character was dressed in red and had green eyes instead? 
    Try adding keywords like bokeh, studio lighting, sharp focus, light rays, or split lightning, and see how amazing your image can become! 


     
Do you at least have a few words in each of these checklist categories?
Great! It means you have a great prompt. Now let's see the best way to put this into practice.



Step-by-step prompt building

That's a whole lot of steps...

Ok, so we have a checklist composed of character, art style, background, color, and light. 
Now how do we put that into practice? 
The answer is simply in the title: By doing it step by step.

A common mistake is writing your whole prompt in one go while in reality, iterating like this is more often than not the preferred methodology :
  • I'm writing words about my character until I am satisfied with it.
  • Then writing words about the art style until I am satisfied with it.
  • Then writing words about the background until I am satisfied with it.
  • Then writing words about colors & light until I am satisfied with them.
While this process is a bit more tedious, it will help you a lot in identifying issues with your prompt. 
Think of this like painting: If you find a mistake it's way easier to fix it if your canvas is almost empty compared to if your painting is almost done.

On Dezgo you can generate images for free without any limits,  so you can definitely take advantage of that and start to build your prompts sequentially. Trust me, it'll be beneficial in the long run!

Tweaking techniques

Hmmm yesss... finally, the PERFECT prompt!

You now have a prompt built following a checklist built sequentially, that's great but let's make some further improvements!
Here are a few advanced techniques:

  • Negative Prompt: On Dezgo, open the "More Option" panel to show the "Negative Prompt" box.
    Simply write in it what you
    DON'T want to see in your picture. For example, my scientific rat above has a pair of glasses. If I don't want to see that I could write "glasses" in my negative prompt.

  • Use prompt weights: Add a + sign after a word to imply its effect or add a - sign to decrease it.
    This advanced technique is very useful especially if you find that some words aren't taken into consideration enough to your tastes.
    You can find more information about weights over here.

  • Use proper grammar: Writing correctly is hard especially if English isn't your first language. And it's also super important when it comes to AI. If a word is not written correctly the AI might straight up just ignore it.
    I'll recommend installing a free web extension like Grammarly to fix those mistakes.
And most importantly for the last tip...
  • Have fun with it ! AI image generation isn't rocket science. Don't be scared to experiment with what works and what doesn't.
    Have fun and switch it up, in the end, everyone will build their prompt differently based on taste. Some people will experiment a lot with models or guidance switching, others will only switch their prompt.
    The whole process is supposed to be enjoyable and that's exactly why we created Dezgo ✌!

Common mistakes



Mistakes?
I never make any mistAAAAAAkes!!


Oh, you still do.
Even if prompting isn't all figured out yet, there are still some universally agreed-upon bad practices.
Let's try to figure them all out to end this article :
  • Generate multiple pictures before changing your prompt.
    Sometimes, an image can turn out badly simply because of poor luck and not because of your prompt. Since "seeds" are defined randomly, try to generate at least 3-4 pictures before giving up on your current prompt.
    Once again, that's easily done with Dezgo and our Power Mode able to batch 10 images at a time.

  • Order matters!
    Words at the start of the prompt are given more importance than words at the end. If you don't like your picture, you could consider shuffling the order of your words instead of changing the prompt entirely.


  • Don't go above the token limit.
    All AI image generators based upon Stable Diffusion have a token limit after which words are simply ignored. This limit is currently 77 tokens.
    Some words can take several tokens so I'll suggest using this online tokenizer to test if your prompt doesn't go above the limit. Simply select "Clip" in the tokenizer settings.

  • Don't copy and paste every prompt you find
    Some prompts use different syntaxes. 
    For example, if you see a prompt with "easynegative" in the negative, it's an embedding that Dezgo doesn't support. Same with brackets or multiple parentheses. Same if you see a <lora:>
    keyword.
    While we definitely encourage getting inspiration from other people, if you simply copy and paste without taking the time to learn what everything is, the quality of your final picture might be impacted.



That's it for me, hope you enjoyed the read.
Definitely consider dropping a comment and I'll look forward to answer them and chat with you.
See you around on Dezgo !


Comments

  1. Excellent article. I use Dezgo almost daily and I highly recommend the service--especially the image-to-image. However, this is more of a "beginners" article than an "advanced" article.

    ReplyDelete
  2. Hi, Clem!

    Does Dezgo have a way to lock/force association of certain descriptions? One major problem I frequently run into is trying to describe an image juxtaposing two polar-opposite elements, and the AI struggles to not mix up descriptions or smoosh them together.

    For instance, if I wanted to produce an image of a red-haired girl with blue eyes hugging a girl with blonde hair and green eyes, the AI will create an image of two identical women whose hair is both red and blonde hair, while their eyes are both blue and green. Or it'll just ignore portions of the description and give two blondes, for instance. Adjusting the accuracy just creates poorer-quality images.

    Is there an existing way to overcome this? If not, it would be nice if a way to "lock" or force the associations was added, perhaps by using brackets. So typing "[red-haired woman with blue eyes] hugging a [blonde woman with green eyes]" would ensure the AI doesn't try to merge colors or have a redhead with green eyes/blonde with blue eyes.

    ReplyDelete
    Replies
    1. try adding "diverse", at least the characters might look more different then, getting better without a lora requires a more advanced backend usually though, which most generators don't have, so a lot of it is just chance too.

      Delete

Post a Comment